for the task MD project I want it so that I can very easily add annotations and also just like view the annotations

One reflection I have is that it's amazing to me how ineffective we as humans are. We don't know what we want, or think about if what we are doing will get us there. Of course, there's some benefit to the lack of rationalism due to how complex the world is. If humans thought about how they can try to affect their goal in the past, in our human evolutionary history, then that human would have died because the world was too complex for it. Over now, I still find it amazing that I can become distracted — meaning that I am not achieving my goal or moving towards my goal — for literally years, months, or weeks. The amount of distraction is surprising.

Why do it? We do it because of akrasia, which is just a general weakness of will — of doing what the higher self wants — and then there is the conscious perception of pain which is preventing us. So it's like us not being hard, as David Goggins would say. If I make my mind hard and become emotionally aware of the pain and discomfort that distraction is helping me deal with, can I become unstoppable? I think so.

When someone toughens their mind against pain, doing things like taking cold showers or running or just doing general things that make them uncomfortable — does that make the pain feel... I think, do hard things become easier? Or is it that hard things stay the same conscious perception of difficulty, however you just deal with it?

before i fly out on wednesday i should check my bring-from-home-list to see what i need to bring from home

consider applying to this
https://jobs.ashbyhq.com/coefficientgiving/61ad9392-f397-42a9-85a2-234ed820b39d

I want to create daily ritual of Reflecting One part of this ritual would be running the software that allows me to reflect on my date. OK One thing that I want to do is I want to be able to compare how I planned my day to be to how the day actually occurred. I can figure out the delta between those two, understand why And then if there is a problem, fix it This means that I would need to fill up my day with What I spent my day doing. This is time intensive and it's not that granular There are a few data sources that I have that could be useful for reconstructing how I spent my day One of them in particular that would be useful is Have this I have a lot of screenshots that I take on my screen every X amount every like X amount of seconds like 10 or five seconds and these can be used to identify what I'm setting my time doing I also have a local server that runs Olamo So I wonder if I, I want you to de-risk a project really quickly to see if the server can classify how I spent the Time of my day between 12 And 6 p.m. Pacific Daylight Savings Time Like sample maybe like 10, like 30 screenshots uniformly between this distribution and see if this can help me identify and see if I can have the system identify what I spent my time doing

https://arxiv.org/pdf/1908.01207

Here’s a summary of an article on “Blue Dot,” which is about defense and AI. What would success look like?

There are three different camps for how AGI can go well:

1. Government control over AGI
This strategy favors stopping corporate and market race incentives by having government-backed AGI projects. These projects can coordinate to ensure there are no race dynamics, and they can enforce this through surveillance and strategic deterrence.

For this strategy to be successful, several things need to happen:
- Algorithmic improvements need to be kept top secret.
- Project security needs to be very high so the AI does not break out of the system, and so model weights are not stolen by malicious actors.

Proponents of this strategy believe that current geopolitical and market race dynamics will push the development of unsafe AI systems. They also do not believe it is possible to create an aligned superintelligence, and think it would be very difficult to defend against a misaligned superintelligence. Therefore, they believe frontier AI development must be stopped, paused, or controlled by world governments.

There is also discussion of preventing other actors from developing their own systems through measures such as export controls.

A problem with this approach is that it would lead to extreme concentration of power among the groups controlling these projects. It is also quite intractable and undesirable for governments to take control over global AI compute supplies.

If that does not happen, then depending on how intelligent AIs are, they could still be dangerous. This raises the question of at what level of intelligence AIs become dangerous. If there is a large gap between our current state and the level where AIs become dangerous, then we are probably safe from rogue actors developing their own superintelligence. However, if this gap is very narrow, or if model weights can be exfiltrated, then we are in deep trouble.

2. Handing over control to aligned superintelligence
This strategy is to build an aligned superintelligence that is more intelligent than humans, and then give control over to it so that it can solve our problems. It would also create even more aligned intelligence and take decisive strategic actions over all other AI projects—essentially achieving a kind of world dominance and preventing anyone else from building a misaligned superintelligence.

This assumes there are “good actors” who are trying to win the race to superintelligence.

For this to work:
- It must be possible to develop aligned superintelligence without causing major harms along the way (e.g., viruses, cybersecurity risks, etc.).
- One of the “good actors” must win the race by a sufficient margin to develop this safely.

3. Building defenses and diffusing AGI
This strategy involves both building defensive technologies and widely distributing AGI. Proponents of this view often identify as differential accelerationists or defense accelerationists, proposed by Vitalik Buterin and Matt Clifford. There is also an essay by Nick Bostrom called “Differential Technological Development.”

The idea is that everyone has access to their own AGIs, and we develop enough defensive technologies such that people are unable to cause mass harm with them.

Proponents argue this is better than the first strategy because it avoids concentration of power—everyone has access rather than a small group controlling it.

However, this strategy assumes it is possible to create strong defensive technologies. A key problem is:
- To build a successful defensive system, you need to patch all possible exploits.
- To build a successful offensive system, you only need to find one exploit.

This suggests that defense is generally harder than attack.

This approach also assumes moderate alignment of AGIs, so they do not create their own superintelligence. This is similar to ideas in the book “If Anyone Builds It, Everyone Dies” or scenarios like “AI 2027.”

In this scenario, AIs should be aligned enough that it is very difficult for them—or for individuals using them—to accumulate enough resources to build a superintelligence.

Send Zac a message about Symphony. Tell him that Open Claw is dead https://github.com/openai/symphony/tree/main

The summary of this article highlights several technologies with the potential to create a more sensible world. These fall under categories like collective epistemics, angelic interventions, tools for strategic awareness, coordination technology, and assurance and privacy.

I’m particularly interested in epistemic virtue—personalized learning systems that tailor content to an individual’s specific needs and interests, interwoven throughout their daily activities. Aligned recommender systems, which optimize for long-term user endorsement rather than short-term engagement, also stand out.

Deep automated briefings would serve as executive assistants, ensuring people have relevant facts prepared for decisions. Reflection scaffolding, acting as a Socratic coach to help navigate tricky situations and foster self-understanding, is another key area. Guardian angels, which proactively flag potential regrets in real time, are also promising.

Furthermore, I’m intrigued by tools for strategic awareness, such as ambient superforecasting—allowing users to query Google and receive high-level assessments of likelihoods—and scenario planning on tap for exploring the implications of different courses of action.

Finally, automated OSINT provides access to higher-quality information about the state of the world. I’m also exploring coordination technologies like Magic Network, which connects people who should know each other to enable mutually beneficial trade and coalition building.

https://www.youtube.com/watch?v=Ly4eJuehcpk

Do not be weak. Speak with confidence.
I don't need to be able to reason my way to good behavior. I can just do things that are good. For example, journaling and writing on paper isn't optimal, but it works better than typing on my phone or on my laptop.
Going to Golden Gate Park was one of these things I did. This was one of the most important things I did this week. Because of it, I was much more successful than I otherwise would have been. But I cannot reason my way to making this choice — I will just go to the park more often.
When I'm there, I feel as though I am thinking more thoughts per minute. I'm not distracted by my phone. I feel my mind being expansive and open. The opportunity cost of not going is high. My unrealized potential is high.

CHALLENGE #4
Choose any competitive situation that you’re in right now.
Who is your opponent? Is it your teacher or coach, your
boss, an unruly client? No matter how they’re treating you
there is one way to not only earn their respect, but turn the
tables. Excellence.
That may mean acing an exam, or crafting an ideal
proposal, or smashing a sales goal. Whatever it is, I want
you to work harder on that project or in that class than you
ever have before. Do everything exactly as they ask, and
whatever standard they set as an ideal outcome, you should
be aiming to surpass that.
If your coach doesn’t give you time in the games, dominate
practice. Check the best guy on your squad and show the
fuck out. That means putting time in off the field. Watching
film so you can study your opponent’s tendencies,
memorizing plays, and training in the gym. You need to
make that coach pay attention.
If it’s your teacher, then start doing work of high quality.
Spend extra time on your assignments. Write papers for her
that she didn’t even assign! Come early to class. Ask
questions. Pay attention. Show her who you are and want to
be.
If it’s a boss, work around the clock. Get to work before
them. Leave after they go home. Make sure they see that
shit, and when it’s time to deliver, surpass their maximum
expectations.
Whoever you’re dealing with, your goal is to make them
watch you achieve what they could never have done
themselves. You want them thinking how amazing you are.
Take their negativity and use it to dominate their task with
everything you’ve got. Take their motherfucking soul!
Afterward, post about it on social and add the hashtag
#canthurtme #takingsouls.
C H A P T E R F I V E
5. ARMORED MIND
“YOUR KNEE LOOKS PRETTY BAD, GOGGINS.”
No fucking shit, doc. With two days to go in walk week, I’d
come by medical for a follow-up. The doctor rolled up my
camo pants and when he gave my right kneecap a gentle
squeeze, pain seized my brain, but I couldn’t show it. I was
playing a role. I was the beat up but otherwise healthy BUD/S
student ready for the fight, and I couldn’t so much as
grimace to pull it off. I already knew the knee was fucked,
and that the odds of getting through another five months of
training on one leg were low, but accepting another roll back
meant enduring another Hell Week, and that was way too
much to process.
“The swelling hasn’t gone down much. How’s it feel?”
The doctor was playing a role too. SEAL candidates had a
don’t ask, don’t tell agreement with most of the medical staff
at Naval Special Warfare Command. I wasn’t about to make
the doctor’s job easier by revealing anything to him, and he
wasn’t gonna take caution’s side and pull the rip cord on a
man’s dream. He lifted his hand and my pain faded. I
coughed and pneumonia once again rattled in my lungs until
I felt the cold truth of his stethoscope on my skin.
Ever since Hell Week was called, I’d been coughing up brown
knots of mucus. For the first two days I lay in bed, day and
night, spitting them into a Gatorade bottle, where I stored
them like so many nickels. I could barely breathe, and
couldn’t move much either. I may have been a bad
motherfucker in Hell Week, but that shit was over, and I had
to deal with the fact that the Devil (and those instructors)
branded me too.
“It’s all right, doc,” I said. “A little stiff is all.”
Time is what I needed. I knew how to push through pain, and
my body had almost always responded with performance. I
wasn’t going to quit just because my knee was barking. It
would come around eventually. The doc prescribed medicine
to reduce the congestion in my lungs and sinuses, and gave
me some Motrin for my knee. Within two days my breathing
improved, but I still couldn’t bend my right leg.
This would be a problem.
Of all the moments in BUD/S that I thought could break me, a
knot-tying exercise never registered on my radar. Then
again, this wasn’t the fucking Boy Scouts. This was an
underwater knot-tying drill held in the fifteen-foot section of
the pool. And while the pool didn’t strike mortal fear into me
like it once did, being negatively buoyant, I knew that any
pool evolution could be my undoing, especially those that
demanded treading water.
Even before Hell Week, we’d been tested in the pool. We had
to perform mock rescues on the instructors and do a fifty-
meter underwater swim without fins on a single breath. That
swim started with a giant stride into the water followed by a
full somersault to siphon off any momentum whatsoever.
Then without kicking off the side, we swam along the lane
lines to the end of our twenty-five-meter pool. On the far
side we were allowed to kick off the wall then swim back.
When I arrived at the fifty-meter mark I rose up and gasped
for air. My heart hammered until my breath smoothed, and I
grasped that I’d actually passed the first of a series of
complicated underwater evolutions that were supposed to
teach us to be calm, cool, and collected underwater on a
breath hold.
The knot-tying evolution was next in the series and it wasn’t
about our ability to tie various knots or a way to time our
max breath hold. Sure, both skills come in handy on
amphibian operations, but this drill was more about our
capacity to juggle multiple stressors in an environment that’s
not sustainable for human life. Despite my health, I was
heading into the drill with some confidence. Things changed
when I started treading water.
That’s how the drill began, with eight students strung out
across the pool, moving our hands and legs like egg beaters.
That’s hard enough for me on two good legs, but because
my right knee didn’t work, I was forced to tread water with
just my left. That spiked the degree of difficulty, and my
heart rate, which sapped my energy.
Each student had an instructor assigned to them for this
evolution and Psycho Pete specifically requested me. It was
obvious I was struggling, and Psycho, and his bruised pride,
were hungry for a little payback. With each revolution of my
right leg, shockwaves of pain exploded like fireworks. Even
with Psycho eyeballing me, I couldn’t hide it. When I
grimaced, he smiled like a kid on Christmas morning.
“Tie a square knot! Then a bowline!” He shouted. I was
working so hard it was difficult to catch my breath, but
Psycho didn’t give two fucks. “Now, dammit!” I gulped air,
bent from the waist and kicked down.
There were five knots in the drill altogether and each student
was told to grab their eight-inch slice of rope, and tie them
off one at a time at the bottom of the pool. We were allotted
a breath in between, but could do as many as all five knots
on a single breath. The instructor called out the knots, but
the pacing was up to each student. We weren’t allowed to
use a mask or goggles to complete the evolution, and the
instructor had to approve each knot with a thumbs up before
we were permitted to surface. If they flashed a thumb down
instead, we had to re-tie that knot correctly, and if we
surfaced before a given knot was approved, that meant
failure and a ticket home.
Once back at the surface, there was no resting or relaxing
between tasks. Treading water was the constant refrain,
which meant soaring heart rates and the continual burning
up of oxygen in the bloodstream for the one-legged man.
Translation: the dives were uncomfortable as hell, and
blacking out was a real possibility.
Psycho glared at me through his mask as I worked my knots.
After about thirty seconds he’d approved both and we
surfaced. He breathed free and easy, but I was gasping and
panting like a wet, tired dog. The pain in my knee was so bad
I felt sweat bead up on my forehead. When you’re sweating
in an unheated pool, you know shit’s fucked up. I was
breathless, low on energy, and wanted to quit, but quitting
this evolution meant quitting BUD/S altogether, and that
wasn’t happening.
“Oh no, are you hurt, Goggins? Do you have some sand in
your pussy?” Psycho asked. “I’ll bet you can’t do the last
three knots on one breath.”
He said it with a smirk, like he was daring me. I knew the
rules. I didn’t have to accept his challenge, but that would
have made Psycho just a little too happy and I couldn’t allow
that. I nodded and kept treading water, delaying my dive
until my pulse evened out and I could score one deep,
nourishing breath. Psycho wasn’t having it. Whenever I
opened my mouth he splashed water in my face to stress me
out even more, a tactic used when trainees started to panic.
That made breathing impossible.
“Go under now or you fail!”
I’d run out of time. I tried to gulp some air before my duck
dive, and tasted a mouthful of Psycho’s splash water instead
as I dove to the bottom of the pool on a negative breath
hold. My lungs were damn near empty which meant I was in
pain from the jump, but I knocked the first one out in a few
seconds. Psycho took his sweet time examining my work. My
heart was thrumming like high alert Morse code. I felt it flip
flop in my chest, like it was trying to break through my rib
cage and fly to freedom. Psycho stared at the twine, flipped
it over and perused it with his eyes and fingers, before
offering a thumb’s up in slow motion. I shook my head,
untied the rope and hit the next one. Again he gave it a close
inspection while my chest burned and diaphragm contracted,
trying to force air into my empty lungs. The pain level in my
knee was at a ten. Stars gathered in my peripheral vision.
Those multiple stressors had me teetering like a Jenga tower,
and I felt like I was about to black out. If that happened, I’d
have to depend on Psycho to swim me to the surface and
bring me around. Did I really trust this man to do that? He
hated me. What if he failed to execute? What if my body was
too burned-out that even a rescue breath couldn’t rouse me?
My mind was spun with those simple toxic questions that
never go away. Why was I here? Why suffer when I could quit
and be comfortable again? Why risk passing out or even
death for a fucking knot drill? I knew that if I succumbed and
bolted to the surface my SEAL career would have ended then
and there, but in that moment I couldn’t figure out why I
ever gave a fuck.
I looked over at Psycho. He held both thumbs up and sported
a big goofy smile on his face like he was watching a damn
comedy show. His split second of pleasure in my pain,
reminded me of all the bullying and taunts I felt as a
teenager, but instead of playing the victim and letting
negative emotions sap my energy and force me to the
surface, a failure, it was as if a new light blazed in my brain
that allowed me to flip the script.
Time stood still as I realized for the first time that I’d always
looked at my entire life, everything I’d been through, from
the wrong perspective. Yes, all the abuse I’d experienced and
the negativity I had to push through challenged me to the
core, but in that moment I stopped seeing myself as the
victim of bad circumstance, and saw my life as the ultimate
training ground instead. My disadvantages had been
callousing my mind all along and had prepared me for that
moment in that pool with Psycho Pete.
I remember my very first day in the gym back in Indiana. My
palms were soft and quickly got torn up on the bars because
they weren’t accustomed to gripping steel. But over time,
after thousands of reps, my palms built up a thick callous as
protection. The same principle works when it comes to
mindset. Until you experience hardships like abuse and
bullying, failures and disappointments, your mind will remain
soft and exposed. Life experience, especially negative
experiences, help callous the mind. But it’s up to you where
that callous lines up. If you choose to see yourself as a victim
of circumstance into adulthood, that callous will become
resentment that protects you from the unfamiliar. It will
make you too cautious and untrusting, and possibly too
angry at the world. It will make you fearful of change and

Given that we don’t yet have a compelling theory of consciousness, and given that just about any behavior[194] could (as far as I know) be accomplished with or without consciousness (consciousness inessentialism), it seems to me that we can’t know which potentially consciousness-indicating features (PCIFs) are actually consciousness-indicating,[195] except insofar as we continue to get evidence about how consciousness works from the best source of evidence about consciousness we have: human self-report.
Unfortunately, as far as I can tell, studies of human consciousness haven’t yet confidently identified any particular “substantive” PCIFs as necessary for consciousness, sufficient for consciousness, or strongly indicative of consciousness.
Still, there are some limits to my uncertainty about the distribution question, for reasons I give below.
As far as we know,[196] the vast majority of human cognitive processing is unconscious, including a large amount of fairly complex, “sophisticated” processing. This suggests that consciousness is the result of some particular kinds of information processing, not just any information processing.
Assuming a relatively complex account of consciousness, I find it intuitively hard to imagine how (e.g.) the 302 neurons of C. elegans could support cognitive algorithms which instantiate consciousness. However, it is more intuitive to me that the ~100,000 neurons of the Gazami crab might support cognitive algorithms which instantiate consciousness. But I can also imagine it being the case that not even a chimpanzee happens to have the right organization of cognitive processing to have conscious experiences.
Given the uncertainties involved, it is hard for me to justify assigning a “probability of consciousness” lower than 5% to any creature with a neuron count at least a couple orders of magnitude larger than that of C. elegans, and it is hard for me to justify assigning a “probability of consciousness” higher than 95% to any non-human species, including chimpanzees. Indeed, I think I can make a weakly plausible case for (e.g.) Gazami crab consciousness, and I think I can make a weakly plausible case for chimpanzee non-consciousness.
When introspecting about how I was intuitively assigning “probabilities of consciousness” (between 5% and 95%) to various species within (say) the “Gazami crabs to chimpanzees” range, it seemed that the four most important factors influencing my “wild guess” probabilities were:

evolutionary distance from humans (years since last common ancestor),
neuroanatomical similarity with humans (see Appendix E),
apparent cognitive-behavioral “sophistication” (advanced social politics, mirror self-recognition, abstract language capabilities, and some other PCIFs[197]), and
total “processing power” (neurons, and maybe especially pallial neurons[198]).

But then, maybe I’m confused about consciousness at a fairly basic level, and consciousness isn’t at all complicated (see Appendix H), as a number of scholars of consciousness currently think. I should give some weight to such views, nearly all of which would imply higher probabilities of consciousness for most animal taxa than more complex accounts of consciousness typically do.

Cheap tests (hours to days)

Self-study: There is a lot you can learn about the history, culture, and working conditions in different policy environments simply by engaging with policy-relevant resources like newsletters, books, podcasts, government strategy documents, etc. For example, if you’re interested in working in Congress, consider reading books like Surviving Inside Congress and Climbing the Hill.
Reading policy reports and newsletters: There are many excellent reports from policy research organizations like think tanks and the Congressional Research Service. These reports can provide high-quality, accessible overviews of policy areas and specific policy proposals. If you’re interested in learning about emerging technology policy issues, check out our list of think tanks focusing on these issues and look for relevant reports and newsletters on their websites. For example, some good starting places are CSET and Brookings for AI policy; CHS and the Bipartisan Commission for biosecurity policy; New America and CDT for civic tech; CFR and the Atlantic Council for cyber policy; FAS and NTI for nuclear policy; and CSIS and SWF for space policy. But it’s normal for detailed reports, especially on technical or niche topics, to seem a bit boring or hard to follow until you develop a better understanding of the topic—even if you have a great policy fit.
Track daily political news for a few weeks: Make a habit of skimming stories from politics and policy-focused outlets like Politico, Foreign Policy, or The Hill. Most policy professionals follow political news—often casually, but sometimes intensely—and staying informed is valuable for navigating DC culture effectively.
Attending policy events (e.g. webinars, workshops, or conferences): Many think tanks host regular events that are open to the public, including many online webinars in recent years (often recordings are available). Check out the think tank websites listed here to learn about upcoming events specific to emerging tech issues. Our policy area webpages also highlight conferences relevant to AI, biosecurity, civic tech, cyber, nuclear security, and space.
Speaking with policy professionals: Among the best ways to learn about policy work (and expand your network) is to talk to practitioners about their day-to-day work, their issues of interest, and their organization’s culture. Our networking in policy guide offers advice on connecting with policy professionals, such as via cold emails, LinkedIn, your school’s career services, or your professional and demographic affiliations. There are also several policy networks and organizations aiming to support professionals from underrepresented backgrounds, including the Public Leadership Education Network, Make Room, NatSecGirlSquad, Women of Color Advancing Peace, Security, and Conflict Transformation, and the Leadership Council for Women in National Security.
Completing online courses: Platforms like Coursera, LinkedIn Learning, and Udemy offer various free or cheap online courses relevant to policy. Our policy area pages highlight online courses on AI, biosecurity, civic tech, cyber, nuclear security, and space.
Participating in advocacy coalitions and events: There are many “policy coalitions” organized around specific policy goals or issues. Consider engaging with them if you find one or more whose mission you share. Relatedly, many public interest organizations organize issue-specific “advocacy days” or “lobby days” that are often open to the public. These events bring together activists, constituents, nonprofits, and other interested groups to coordinate and often meet with lawmakers and government officials. You can find various policy coalitions and advocacy/lobby days via online search.
Reflecting on your fit for policy: Write down your skills, strengths, and interests, and consider how they map onto different roles in policy organizations. As our full-time policy roles guide notes, “If you excel at research, you might consider think tanks or government research organizations (like an FFRDC or the Congressional Research Service); if you enjoy fast-paced, collaborative work on legislation, consider working as a legislative staffer in Congress; if you enjoy politics and have strong communication skills, you could become an advocate for an interest group or work in a congressional campaign; and if you have strong project and people management skills, you might find appealing opportunities to work on policy implementation in agencies.”

Medium-effort tests (weeks to months)

Internships: Interning in policy organizations is a great early-career opportunity to explore whether you’d enjoy and excel in different types of policy work and to compare public policy to your other sectors of interest. While internships can be great fit tests, they generally differ from full-time employment: interns typically have much less responsibility than full-time staff, less context on the organization, and less experience performing the particular tasks. So, as an intern, pay close attention to the work of your full-time colleagues and supervisors, and reflect on whether you’d enjoy and be good at their work. If at all possible, speak with your full-time colleagues to help you answer these questions. If you’re a student, consider participating in your school’s Semester in DC program (if there’s one) and applying for a remote, part-time internship with a government agency via the Virtual Student Federal Service (VSFS).5
School work (classes, assignments, and theses): During college, you can learn about government and different policy areas by taking policy-relevant classes and writing your class assignments or thesis about related topics. While school work can allow you to learn about the policy world, it typically doesn’t closely resemble the day-to-day experience of policy professionals. The writing and research in college tend to be more academic (e.g. in terms of style, methodology, and target audience) than what’s involved in many policy jobs, even in writing-heavy think tank research positions.
On-campus work: As a student, you might also be able to build policy-relevant experience through research or teaching assistantships (“RA” or “TA” positions) with professors working in relevant areas. It’s best if the professor has first-hand experience and relationships in DC policymaking institutions. But it can also be valuable to RA in a technical lab if you’re interested in science and technology policy or want to explore your fit for technical research as well.
Campus student groups: In college, you could also get involved with policy-relevant campus student groups, such as student magazines, Model United Nations, or political groups (be intentional about affiliating yourself with partisan groups). Engaging with these groups can allow you to explore certain aspects of policy work (e.g. policy writing and research, debating policy proposals). But be mindful of the time commitment and also consider less costly or more informative ways to explore public policy.
Policy seminars and courses: Some think tanks and academic organizations offer training seminars and courses on specific policy topics. These programs are typically free, occur in the summer, and last between a few days to a few weeks. Some examples include the Henry Clay Center’s two-week College Student Congress, UC San Diego’s two-week Public Policy and Nuclear Threats Boot Camp, the Public Leadership Education Network’s policy seminars for college women, and the various summer opportunities offered by conservative or libertarian policy organizations.
Part-time and short-term fellowships: Many policy fellowships are full-time programs lasting an entire year or more. But there are some fellowships relevant to emerging technology issues that are either part-time or only last a few months at most. These part-time opportunities include the Center for Strategic and International Studies’ Project on Nuclear Issues (PONI), the Center for Health Security’s Emerging Leaders in Biosecurity Fellowship (ELBI), RAND’s Center on AI, Security, and Technology Fellowship (CAST, also available full-time), and the Council on Strategic Risks’ Ending Bioweapons Fellowship and Nuclear Risk Reduction Fellowship—but note that except for PONI all of these programs target mid-career subject matter experts (though your expertise doesn’t necessarily have to be in policy). A few short-term, full-time fellowships include the National Academies of Science’s 12-week Mirzayan Science and Technology Policy Graduate Fellowship for STEM graduate degree students or recent graduates and the 10-week Aspen Tech Policy Hub Fellowship for technologists.
Informal collaboration: You may be able to contribute to short policy projects by asking a policy professional if they could use any help in their work, such as support for a research project. Doing this is less common and institutionalized than the other fit tests on this list, but it was recommended to us by some senior policy practitioners who found it very helpful in testing their fit and getting a foot in the door. Such collaborations would have similar benefits to an internship, but they may be easier to do part-time, remotely, on shorter notice, and in an area you’re especially interested in. The main downsides are that your work would likely be unpaid, you’re less likely to get good mentorship, and you’ll likely have fewer networking opportunities. Collaborations like this may be more likely to succeed if you already have at least some subject-matter knowledge and know policy professionals working on interesting topics, though a cold email might also work. Send a message that signals your learning and interest (e.g. by discussing the researcher’s work and your own research ideas and experience) and make sure you have the time and interest to follow through on the collaboration. If you already have significant subject-matter expertise outside of policy (e.g. in technology), there might also be more opportunities for you to formally collaborate with policy professionals.

High-effort tests (months to years)

Policy fellowships: Besides offering training and networking opportunities, fellowships are often designed specifically to allow individuals with different backgrounds and levels of experience to explore policy work full-time for 1-2 years. Some programs aim to support recent graduates in exploring their fit for policy; others target mid-career professionals from non-policy fields (e.g. STEM, business, or law) to help them transition into public service. Fellowships are often better than internships as fit tests since they more closely resemble the work of traditional full-time employees. If you’re particularly interested in Congress or the executive branch, consider our congressional fellowships list and this federal agency fellowship database.
Graduate school: Alongside many other benefits, completing a policy-relevant graduate degree (e.g. master’s programs or law school) can be great for learning about different policy areas, institutions, and processes. But considering the substantial cost of graduate school (in terms of both time and money), we recommend getting a policy-focused graduate degree only after testing your fit for policy in other ways to ensure you’re actually getting the right degree. Many graduate programs are much more academic than policy work. So, if you’re pursuing a career in public policy, look for practitioner-oriented programs rather than academic programs. During your degree, consider directing much (or even most) of your effort toward pursuing relevant internships and part-time work experiences that give you stronger fit signals for various policy roles. If you’re completing a non-policy-oriented degree (e.g. most standard STEM degrees, social science, business, and humanities degrees), it’s likely you won’t learn much about the functioning of government or the policy process in your classes.
Full-time roles: The most reliable way to explore your fit for policy work is actually to do the work. If you’re struggling to find a full-time position in DC or have a location constraint, you can test your policy fit by working at different levels of government (e.g. as a staffer in a state legislature or working in a state or local government agency). And if you’re transitioning from a different field, consider if there are intermediate positions allowing you to learn more about policy work, leverage your background, and advance your policy career development. For example, a technologist might transition into policy by first working for the government in a technical role as a stepping stone to future policy-focused roles. As a scientist, you might also work for one of the national labs or other government-affiliated research organizations.

From this article, it says that those in policy careers, deal with uncertainty thinking creatively gracefully navigating uncertainty, taking initiative, and building coalitions. This seems very related to general entrepreneurial work.

Assessing personal fit for launching a high impact non profit. The book gives a few metrics. This is my self-rating.

self motivated - 8/10
conscientious - 9/10
irrepressible - 6/10 (I feel this is my weak point, I think my mindset becomes weak, but I am capable of strong mindset (marathon))
open minded - 8/10
creative - 6/10 (hard for me to know how creative I am)


the right reasons:
- I questionably have the right reasons, I'm partly motivated by prestige and doing something that I personally enjoy.

results oriented:
- yes, I am

ambitious altruism:
- 80 percent of the way there. i think i am lacking partly because i recognize the high statusness of being considered altriustic. I do not cry at spreadsheets yet

the most impactful people spend like a year exploring broadly and then 3-4 years executing rigorously.

I like this idea of high indirect impact can come by founding an org because you can stop working there and the org can continue functioning, giving you "passively attributes good"

often, the limiting factor for the establishment of an effective organization is a single co-founder. this makes it easier to maximize your direct impact

an estimate by Peter wildeford suggests that founding a give well top charity could be as impactful as donating 400k a hear. incubated CE charities has the ev to donating 190k per year. this is quite surprising to me how little this is.

wow, this is my smaller than it seems. it seems making a lot of money and donating it is quite effective. I'm not sure if I believe this previous analysis for this reason. it is very unintuivie that founding an org that can become a top charity has such a low EV. This is more evidence that personal got matters the most. choose where you will have the most impact as aperosn

Anthropic has this responsible scaling policy They had a new version of it just a few weeks ago RSP 3.0 Some reasons why they do it is because They wanted to be an internal forcing function with Anthropics so that people use this to identify when they should release models or have other safeguards they also want other A company has introduced similar Policies They want to create a bigger consensus about risks, especially among bioweapons or once and models which are certain capability They also want to have multi lateral action among all the AI companies After Anthropic released their RSP, it caused Open AI in DeepMind to adopt similar frameworks These are the voluntary standards that they use

The AI Futures Project has some interesting models: https://www.aifuturesmodel.com/
I think its worth spending like 2 hours understanding and being able to articulate this so I can get a better view on what my timelines are

Listening to the Next Big Idea podcast episode with Rutger Bregman, here are some interesting ideas and statistics I want to remember:
- The average corporate laywer does 30k in damanges per year to society

A frame from this podcast is those that are doing consulting, finance, or law are seen as doing something 'extremely boring' with their talents. They could be using their network, capital, career capital, and influence to solve pressing problems in the world but instead they choose a life that maximizes prestige or salary.

A quote I like is the only time change occurs is when a group of motivated citizens take it upon themselves to make a change.

Bill Gates is a good example of an ambitious entrepreneur and philathoprist. Due to his donations he has saved millions of lives. A criticism is made that typically philathropists donate money to schools to get a building named after them instead of using their money to prevent the ~15k children dying every day from preventable diseases.

A important idea covered in this book is to get over the "illusion of purity" that many activists have. This is people wanting to take actions that make them feel good (morally pure) instead of taking actions that win. For example, activists spending their effort convincing others to reduce their meat consumption instead of taking action against corporations and coalitions that keep factory farming in place.

A salient example is abolition in the mid 1800s Britian. There was a growing movement of abolitionists, but they got abolition not by appealing to the moral worth of enslaved peoples. They created a coalition fighting for the same action, but for different reasons. 20% of white sailors would die on these trips in the slave trade, which convinced the government to abolish the slave trade.

Another amazing quote "It's the mark of a civilized human being to be able to look at a row of numbers on an Excel sheet and weep"

Insane that we spend equal amount of moeny on preventing infection disseases as americans bleeching their teeth.

I really wanna make a project that is just a system that very deeply and rigorously understands reality. i want to make it so that i can input a goal I want to achieve and my current state and it finds the best solutions given the evidence and its understanding of cause and effect. For example, this can automatically validate charities via their theories of change, finding information that they were previously unaware of. The human brain muddles information, it forgets to ask basic questions.

The hardest but most critical step is getting your foot in the door. Many early-career professionals, even those with master’s degrees, start as interns while applying for full-time jobs. This internship stage can be brief, and once you gain your first professional experience, you can leverage connections with colleagues, supervisors, and collaborators to transition into subsequent roles. In many cases, securing your first internship or full-time role is the hardest, so we encourage you not to give up too early. Given the professional development boost of an initial policy experience, you should prioritize starting somewhere—even if the role isn’t a perfect fit for your long-term interests.

Some attributes are valuable across most policy jobs:

Professionalism (both in and out of work)
High openness to working with people with different motivations and worldviews
Being kind and social (crucial for the collaborative, trust-based nature of DC policy work)
Intrinsic motivation (despite slower feedback loops and hard-to-measure progress)
Humility (policy work is often very complex; genuine eagerness to learn can take you far)
Being proactive and entrepreneurial (the biggest policy “wins” often come from people who proactively search for opportunities to improve processes and outcomes)
Willingness to invest in key policy skills (e.g. writing and professional communication)
Having US citizenship (foreign nationals face many limitations to working in US policy)12

For these reasons, I may not like working in policy. I am not a fan of expectations around professionalism, formal work environments. I am alright working with people with different motivations and worldviews as many of my friends have compltely different world-views to me.
"""
At the same time, policy careers can pose personal challenges for some people. Most policy roles are located in DC, creating significant geographic constraints for some. Expectations around professionalism—including restrictions on public statements and formal work environments—can feel limiting. Policy work often involves collaboration with people with different motivations and worldviews, which may be draining for individuals who struggle to maintain productive relationships amid deep disagreements. Progress is often slow or incremental and comes with limited recognition, potentially disappointing those who thrive on regular feedback loops, clear measures of success, or public appreciation.
"""

Because the private sector pays so much, there is a talen shortage in areas where government agencies need expertise.
"""
While federal salaries generally track market rates, some specialized roles—particularly in emerging technologies—face significant private-sector pay gaps. This creates talent shortages in areas where government agencies urgently need expertise. For those with technical backgrounds, this means opportunities to fill critical needs while contributing your skills to public service.
"""

I am worried that policy will be frusturating for me
'''
Despite its significant potential for impact, policy work often faces political and institutional constraints. Policy can move at a frustratingly slow pace, with changes sometimes taking years or even decades to fully pass or implement. Political environments can shift unpredictably, occasionally reversing hard-won progress as administrations or priorities change. Complex bureaucratic structures and competing interests may also create obstacles to innovative policy approaches, requiring patience and creativity to navigate setbacks.
'''

“Ninety-five percent of the nitty-gritty work of drafting bills and negotiating [their final form] is now done by staff.

There’s a saying in government that “people are policy”—even the best policy ideas only create impact through individual decisions and actions about how to interpret, shape, communicate, and implement them. This explains why government staffing—from political appointments to civil service hiring—receives such careful attention and why talent shortages can critically impact program success.

This is a real concern I have regarding working in public policy. “public policy has great power, both to improve people’s lives if it is planned and executed well and to cause significant suffering if it is not.”

Policy is made through a roughly 4 phase process, starting with policy research. This can be done by congress, white house, agencies, national labs, think tanks, NGOs, Universities, Contractors. Then is policy advocacy for others to buy in on the decisions and actions the government wants to take, this is done by convress, white house, agencies within the government (notice that national labs are excluded because they don't have power here), and think tanks, NGOs, universities, Industry groups, labor unions from outside the government try to get buy in. Policy is formally adopted by Congress, white house, agencies via official channels and then implemented through agencies and contractors.

Authorization is different than appropriation. Congress legislation allows for things to exist. For the PAHPA in 2006 it allowed for BARDA to exist, but Congress still needed to appropriate funds for it during the year, the president needed to appoint the first Assistant Secretary for Preparedness and Response, they then needed to make prioritizations and under the umbrella Congress put, how to divide the budget.

when reading this, this feels like the first time in a while that i'm reading something with almost no background knowledge on. My reading is slow and I am reading every word carefully. This is a good feeling. I am growing in my capabilities and understanding of the world.

Policy is everything that the government does, all actions it takes. Public policy is policy directed towards society or the public. Policy is not just legislation, but it is also what individual agencies do (which Congress grants or limits their power), court decisions, government spending so laws can be acted on. Policy is also what comments the Federal Reserve makes, public funding for universities, voluntary industry standards. All of these are affordances that people working 'in policy' can do.

Reflecting on the book Indistractable It claims that distraction is just things that move you away from your goals And traction is what move you towards your goals I just went on a ram. I went to bed at a reasonable time. I had a full night's sleep. I was reading before bed And now I feel that I have a lot of focus I thought that focus was just When you do a task, it's like the mental state that you're in However, I now have the focus to just not check my email or check my messages, but to meditate meditating is something that requires focus because it requires you to not be distracted I feel that I have a better sense of my body and I have heightened sensation now because of my run This feels good. I know where I need to be and I know what I need to do I need to run more I need to hold myself to my previous standards I am on an upward trajectory right now. I have been for the last couple of weeks I am Content

This interesting idea from this lesswrong article is called factoring. It's based from the realization that some people have this disorder called pica, which when they have a nutrient disorder it causes them to eat things that don't contain that nutrient but they eat it anyways because their body kinda doesn't know what to do. This is something we can extrapolate to regular human behavior. Your brain tries to learn and pattern match to try to get you what you want. You can short circuit this by doing something called 'factoring' which I would probably instead called root cause analysis, but the idea is your just identify what is the cause of your behavior (easier said than done) and you solve the root problem. You need to do introspection for this. Why do you respond negatively to people? What are you trying to get?

An idea i have is the reason why i eat so much and over eat is because of distraction. food i sused to make myself feel more comfortable. i had this epiphany when reading distractable and noticing i wanted to eat just so i can be distracted instead of because i was hungry.

Okay, the next part of the book is after you have identified what is the internal trigger that is causing you to be Another way that you can make yourself become indistractable is to Reimagined the task The main idea is to make it so that this It feels like it is fun and makes it feel like it is play This is related to the course that I took about game design, where we discuss about what makes something fun or make something enjoyable Cool It's not necessarily something that is like a spoonful of sugar as mary poppins would say it but rather it's something that allows you to have mastery it allows you to it puts you in a playful mood it makes you feel better and better. It gives you constraints This is for community It's about discovering novel solutions to problems. That is what fuck is Fun is looking for the variability and something others don't notice It's breaking through the bored of and monotony to discover its hidden beauty One example of this is putting a time constraint on you finishing some sort of project or challenge That makes things more fun plotting out the optimal path for mowing your law

Okay, so from the book indistractable, there's a four-step process for reimagining the internal trigger first is to look for the discomfort that precedes the distraction focusing in on the internal trigger This can be something like a craving, feeling restless or thinking that you're incompetent Then you should write down what the trigger is. You should note the time of day How you felt if it was an internal trigger, external trigger, or something else entirely You should write If you did or did not give in to the distraction. It's okay for you to give it to the distraction, the purpose of this exercises for you to understand what is making you feel distracted You can also write out ideas on how you can prevent this from happening in the future Then you should explore your sensations The next is to explore your sensation The recommendation is to become curious about the sensation Try to understand what it feels like try to stay with a feeling before acting on the impulse You can imagine leaves on a stream. You can imagine that Your internal triggers Or emotions are just leaves floating down the stream placing each thought of your mind on each leaf It's going to be a memory, a worry, a word, or an image, and you just imagine those leaves floating down the stream You just sit and watch The final step is to be aware of liminal moments A liminal movement is a transition from one thing to another throughout our days, like opening a door opening a tab in the web browser, getting annoyed how long it takes a load, and opening up another web page while you wait watching social media when you get on a train, etc The problem with these actions is if you do them just for a second but instead you spend like half an hour or 20 minutes doing these things In order to prevent this, you can do a 10 minute rule, so if you do feel like you Wanted to check your phone then you could set a five minute timer ten minute timer etc at which point you can look at the distraction this technique is effective because It prevents you from giving it immediately When you gently attend to negative emotions, they tend to dissipate and positive emotions expand

When it comes to getting context on a field you want to work in, the right first move is cartography. Before you read anything in depth, spend time sketching the landscape: who exists, what they're doing, how they're funded, how they relate to each other, and — most importantly — where the lines of genuine disagreement run. This takes a few days at most. It changes the quality of everything you read afterward enormously.

Another way that people tend to be incorrect or be shallow in their context understanding is having different priors So For example, someone who wanted to do policy work in AI safety, they might not have the same prior about how to get things done. Someone from a different policy background might say that institutions would be the best way to get policy changes However, in the technical research community, we believe in a short timeline So these methods might not actually work. Perhaps a better solution would be to try to influence key decision makers If you do not have a lot of context about the field then Going through long-standing institutions might be a reasonable way to get policy work done. However, someone else talking about your belief would be confused Or it just would not make sense to them in order to help with improving this like context or this reasoning flawed understanding of the field is to understand your own arguments for why doing this is the best or come up with a theory of change of what you want to do and question each assumption

People that are deficient in the map view of context are those that absolutely talk about maybe a alignment or read the major books but they don't know who is doing what Why are they working on? They can say that they want to work on an A-safety, but not specifically what they want to do where they want to work at This is very revealing when you want to work in this field because If you cannot answer the basic question of why you want to work at the organization you're applying for, then you will not be selected. You should be able to able to answer "who is doing the most interesting work on X right now". The candidate with a good map can name specific research groups, explain what's interesting about their approach, and articulate why they find it more compelling than alternatives.

There are three components of context that one should optimize for One of them is the map. It is knowing who does what And why? What different organizations exist? how they relate to we differ from each other. A lot of debates between them and the spatial Landscape of them who is funding who's being funded by who? etc The next is language. It is a shared conceptual vocabulary. It's knowing about terms such as a shelling point It's about understanding how they're used in discussions and how to use them. And finally, it's about what are the default assumptions that people are making, what counsel is evident towards a belief what does not count as an evidence towards a belief What are uncertainties?

in many communities, but especially in AI safety context matters It's about knowing what people are talking about, what are the live debates and problems. It's about engaging seriously with problems and being able to explain your position It's not about performing that you are aligned, but it's like seriously thinking through problems and the ideas that have gone through This is probably more necessary in the AI safety community because it is smaller and it is a relatively newer field

The main reason why the AI safety community is currently trying to create some distance from the effective altruism community is because The effective altruism community has some reputational baggage because of the FTX and Sam Bankman-Fried all out. The air safety community does not want to be associated with EA publicly because this will associate them with the FtX fallout. Also, EA has a little bit of a cult perception from the media when people talk about things such as long-termism or expected calculations of future simulated beings So, okay I'll do them The separation seems to be more because of marketing and public perception than in reality Everyone in AI safety essentially is an EA

In the first scenario, optimization pressures optimize for what is easy to measure, there is a gap between what is easy to measure and what we actually want, this will lead to a society where things that appear to be what we want to be optimized are optimized instead of what we want to optimize. The free parameters in this optimization structure will become extreme as the system becomes more complex and human reasoning cannot lasso it back. In this scenario, human agency and augmenting human abilities is useful for the success of our species. It seems a general trend I am seeing is agumenting human reasoning is a very good way to deal with increasing AI.

Cromwell's rule https://en.wikipedia.org/wiki/Cromwell%27s_rule

The doomsday argument is the argument that uses the anthropic principle to reason that we are not a special observer in our reference class. There have been 100 billion humans before me, this means out of all humans I should expect to be in the median 100 billion people. If there will be an astronomical number of humans then me being born in the first 0.00...001 humans is a priori unlikely. A counter to this argument is humans will no longer be in a human form, so it will be a different reference class.

In general, the EA and d/acc communities are mostly self-sustaining and growing. I can see the argument for why they wouldn't in the long run (people that pursue things that society values like money and status will be more numerous and influential), but there are other communities of thought that go against these economic optimization pressures, such as abolition. (Although, I do know that abolition was not entirely orthogonal to optimization pressures, the reasons for abolition happen seem to be mostly to have been done against these optimization pressures). It perhaps points that the EA and d/acc communities should try to design the community so it changes others values.
"""
For this to work, you would need a mechanism that not only slows progress but redirects it, that specifically channels civilizational resources toward survival-relevant capabilities rather than competitive-advantage-relevant capabilities. And this mechanism would need to be self-sustaining: it would need to maintain itself against the continuous pressure of competitive dynamics that reward defection from the survival-oriented program. This is, in effect, asking for a global coordination mechanism that persists indefinitely against strong incentives to defect.
"""

I need to be cautious about what I work on, this is related to the thought that most things do not matter. I need to be on level 3, otherwise I am wasting my time. Everyone who works in the field of existential risks or AI safety needs to get on level 3. In high school I was on level 1, I am entering level 2, I want to have a swift transition to level 3.
"""
Caring about survival is not the same as surviving

I do not dispute that many people care about survival, on the psychological level.

However, there are three levels, and only the third matters:

Level 1: Caring about survival in the psychological sense. You feel concerned. You believe existential risk is real and important. You experience anxiety or urgency about it.

Level 2: Executing actions that are intended to be pro-survival. You donate to AI safety organizations. You write papers about existential risk. You lobby for policy changes. You work at an alignment lab. You feel, reasonably, that you are doing something about the problem.

Level 3: Executing actions that actually lead to survival. You are executing actions that causally contribute to the civilization actually not going extinct. This is an absurdly demanding criterion, and that is precisely the point.
"""

Does this impact or influence the people that I tend to see working on existential risk? I imagine more intelligent people be working on problems that are related to existential risk, at least, a component of intelligence I measure is "caring about the future" and "realizing you can do something about the future".

"""
Intelligence is not the bottleneck

It is true that, all else being equal, greater intelligence leads to better awareness of existential risks and perhaps even greater concern about them.

But "all else being equal" is never the actual situation. What actually happens when the general intelligence level of a civilization increases is that all agents get smarter, including the ones who are optimizing for things other than survival. And the ones optimizing for things other than survival still have a competitive advantage, because they are not paying the survival tax. A world with smarter humans is not a world where smart survival-oriented humans dominate, but a world where smart survival-oriented humans compete against equally smart growth-oriented, power-oriented, and profit-oriented humans, and lose, for exactly the same structural reasons they lose now, just at a higher cognitive level.
"""

"""
Every emerging civilization is around the dumbest possible civilization

A civilization arises when a species crosses some threshold of cognitive and social capability. Below this threshold, you don't get civilization at all. Above it, you do. The question is: how far above the threshold should we expect a newly-emerged civilization to be?

The answer, by a straightforward selection argument, is: barely above it. Civilizations emerge as soon as they can, because the components that give rise to civilization (intelligence, social complexity, tool use) are under positive selection pressure for reasons that have nothing to do with building civilizations per se. They are selected because they confer immediate competitive advantage. A species crosses the civilization threshold not because it was aiming for civilization but because the optimization pressures that were pushing intelligence upward for other reasons happened to push it past the threshold. And since it crosses as soon as it can, it crosses with approximately the minimum cognitive endowment required.

This is directly analogous to a point that is familiar in evolutionary biology: organisms tend to be minimally adapted to their niches, not maximally, because evolution is a satisficing process that stops optimizing a trait once it is "good enough" for the current competitive landscape (or more precisely, once the marginal fitness return of further improvement drops below the marginal cost). Civilizations, similarly, emerge at approximately the minimum viable intelligence, because there is no selection pressure that specifically pushes a pre-civilizational species past the minimum, and even if there were, it would require more time than civilization-scale time.[5]

So, we should expect a newly-emerged civilization to be roughly as dumb as a civilization can possibly be while still counting as a civilization.
"""

This claim is somewhat reasonable: "because of this "jaggedness", humans, and especially smart and caring ones would be needed for at least several decades if not more"

This is how I felt in physics class and when doing ESRP:
"""
Words do not have intrinsic definitions. If I hear the syllables “bea-ver” and think of a large rodent, that is a fact about my own state of mind, not a fact about the syllables “bea-ver.” The sequence of syllables “made of waves” (or “because of heat conduction”) is not a hypothesis; it is a pattern of vibrations traveling through the air, or ink on paper. It can associate to a hypothesis in someone’s mind, but it is not, of itself, right or wrong. But in school, the teacher hands you a gold star for saying “made of waves,” which must be the correct answer because the teacher heard a physicist emit the same sound-vibrations. Since verbal behavior (spoken or written) is what gets the gold star, students begin to think that verbal behavior has a truth-value. After all, either light is made of waves, or it isn’t, right?

And this leads into an even worse habit. Suppose the teacher asks you why the far side of a metal plate feels warmer than the side next to the radiator. If you say “I don’t know,” you have no chance of getting a gold star—it won’t even count as class participation.
"""

The progress in the world has been significant, we have gotten significantly richer (the united states has become 20x richer), we live longer. And this graph was more than 10 years ago! The average income is over 20k and life expectancy is over 75.
## Screenshot

This is why personal fit matters for impact: A small percentage of the workers in any given domain is responsible for the bulk of the work. Generally, the top 10% of the most prolific elite can be credited with around 50% of all contributions, whereas the bottom 50% of the least productive workers can claim only 15% of the total work, and the most productive contributor is usually about 100 times more prolific than the least.

the more ambiguity there is in a role the less people are capable of doing it so in order to maximize impact, pick roles that have high ambiguity

Core idea here:
- There is a lot of money, time, quality policy decisions, etc. lost due to incorrect science. Frequently, studies are not reproducible, thus wasting all of the studies built upon this study as a core assumption.
- Because of the immense waste, we should invest funding and effort in improving the quality of the science, rigorously and autonomously verifying all papers sent in. This can be done algorithmically or with AI

An objection I have to the above is it relies on minor advances to AI compared to where we are today (actually we can probably do most of this with models today, but I do anticipate seeing edge cases where models fail). If AI writes more and more code and scientific analysis, I suspect the amount of errors to decrease on the margin, and the amount of these errors will continue to decrease. It seems plausible to me that in the future AI generated code and AI generated science will have no errors and this may not be useful in the future.

because of this, i should probably not work on a 'sexy problem', it Could be that in general the more sexy a problem is unless you should work on it, partly because this means that many other people are going to be drawn to it. So I should be drawn to problems other people are not drawn towards

D.A. Henderson was the person responsible for globally irradicating small pox, there could be things to learn from him. """The goal of globally eradicating smallpox was first articulated by WHO in 1959, but with no serious funding, weak operational plans, and little confidence it would succeed. In 1966, after limited success, WHO and partner governments wanted someone who could run a program to achieve this goal and hired D.A. Henderson, an epidemiologist working at the CDC known for his comfort operating in chaotic field conditions, his blunt, results-oriented management style, and deep belief in accountability and speed.

Henderson didn’t discover anything—the first smallpox vaccine existed for 170 years before eradication. The gap was strategy, coordination, and will.

When the original strategy of mass vaccination wasn’t working, he switched to a ‘ring vaccination’ and surveillance approach. He ran the program like a wartime operation: he traveled constantly to endemic countries, demanded real-time data so that they could quickly adapt strategy and tactics as needed, empowered young field staff over senior bureaucrats, fought WHO HQ inertia, pressured governments publicly, and occasionally bypassed them altogether. By the late 1960s, Henderson believed that failure was unacceptable and that time mattered morally, a mindset that spread through the program.

The last natural case of smallpox was in 1977, and while many others were involved, Henderson is largely credited for leading—or “GMing”—the successful 11-year effort."""

This part is interesting because I didn't even think that this is something that "should" exist, but I feel that it is obvious and noble for someone to wake up and believe it is their responsbility to get this problem solved.: """There’s a surprisingly big category of problems that are ‘orphaned.’ By ‘orphaned’ I mean: you can’t point to a specific person or organization who thinks it’s their responsibility to deliver the outcome in its entirety. Lots of people talk about the problem, and often many work on slices of it. But if you asked: ‘is there a hyper-competent person waking up every day feeling accountable for making sure this gets solved?"""

In this paper, they measured the impact of AI tools from the first half of 2025 to developer productivity. They found, surprisingly that AI tools decreased their productivity by 19%! There are a few reasons why this probably cannot be extrapolated to saying AI slows developers down:
- devs were highly skilled and familiar with these repos, less skilled developers likely would improve their time
- codebases were large and old (but on the other hand, most code bases are probably going to be large)
- this was in the 'non agentic' ai era

Although, it does reveal that if devs are highly experienced on large, familiar code bases, then AI tools can slow them down.
Surprisingly, those that use AI they spend 5% of their time waiting on AI!

Another thought from the Alex and Sam podcast episode is this mention that our brains were evolved to have some intuitions about how the world works, so this makes some truths difficult to understand or impossible to fully believe (such as many worlds). This doesn't mean they are not true. This reveals an assumption a lot of people make, especially when it comes to theism, that their intuitions about what is true are correct, and because they cannot see how the universe came to be, it's more likely to them that there is a conscious, powerful agent that willed it to being.

When I do things on my computer, I can have a system that takes in what I do on my computer and "checks" off my habits (when I post something on X, for example).

Hypothesis:
Music evolved to bring people together. In the past, when people were listening to music, they were with other people. Now, in society, we are listening to music because it is pleasant, but there is no social component to this.
Hypothesis:
This is a bad thing. Listening to music in a non-social context is like eating candy.

ai security of weights, remember to read this

Ralph loops are used for prompting LLMs. It is useful when there is some completion critera. you jsut ask the coding agent to do some work for you, you continue prompting it with the same prompt until it is done. This invariably leads to the idea of putting multiple steps and branching conditions.

This is a very interesting article. It's about how you can sample for negative temperatures for large language models. So if you sample at a negative temperature, then the scale reverses. Instead of choosing the most likely token, it's the least likely token.
## Screenshot

Agents can be thought of (so far) in a similar way that humans are thought about. Thinking about working with agents/delegating to agents can be thought of in the same way as delegating to humans.

I am currently wrestling with lots of uncertainty. This makes me feel better
"""
The world is changing at such a rapid rate that it’s turning us all into amateurs. Even for professionals, the best way to flourish is to retain an amateur’s spirit and embrace uncertainty and the unknown. When Radiohead frontman Thom Yorke was asked what he thought his greatest strength was, he answered, “That I don’t know what I’m doing.”
"""

Everyone wants AI to replace their work if they own it
Work will shift to people putting their expert judgements into AI. Everyone will soon try to put their thinking into AI to do their work for them, eventually selling their digital clones instead of themselves.

https://www.swyx.io/learn-in-public

i don't like meaningless marekting. this means nothing
## Screenshot

Internalize this:
"""
I am amazed at how many really smart people don’t understand that you can make big mistakes in public and emerge none the worse for it. I know distinguished researchers who will go to preposterous lengths to avoid having to acknowledge that they were wrong about something. They have never noticed, apparently, that the earth does not swallow people up when they say, “Oops, you’re right. I guess I made a mistake.” Actually, people love it when somebody admits to making a mistake. All kinds of people love pointing out mistakes. Generous-spirited people appreciate your giving them the opportunity to help, and acknowledging it when they succeed in helping you; mean-spirited people enjoy showing you up. Let them! Either way we all win.
"""

make it a habit to record mistakes, make them publicly, review them periodically, make this a system as part of your growth system.

The chief trick to making good mistakes is not to hide them—especially not from yourself. Instead of turning away in denial when you make a mistake, you should become a connoisseur of your own mistakes, turning them over in your mind as if they were works of art, which in a way they are. The fundamental reaction to any mistake ought to be this: “Well, I won’t do that again!” Natural selection doesn’t actually think the thought; it just wipes out the goofers before they can reproduce; natural selection won’t do that again, at least not as often. Animals that can learn—learn not to make that noise, touch that wire, eat that food—have something with a similar selective force in their brains. (B. F. Skinner and the behaviorists understood the need for this and called it “reinforcement” learning; that response is not reinforced and suffers “extinction.”) We human beings carry matters to a much more swift and efficient level. We can actually think the thought, reflecting on what we have just done: “Well, I won’t do that again!” And when we reflect, we confront directly the problem that must be solved by any mistake-maker: what, exactly, is that? What was it about what I just did that got me into all this trouble? The trick is to take advantage of the particular details of the mess you’ve made, so that your next attempt will be informed by it and not just another blind stab in the dark.

We have all heard the forlorn refrain “Well, it seemed like a good idea at the time!” This phrase has come to stand for the rueful reflection of an idiot, a sign of stupidity, but in fact we should appreciate it as a pillar of wisdom. Any being, any agent, who can truly say, “Well, it seemed like a good idea at the time!” is standing on the threshold of brilliance. We human beings pride ourselves on our intelligence, and one of its hallmarks is that we can remember our previous thinking, and reflect on it—on how it seemed, on why it was tempting in the first place, and then about what went wrong. I know of no evidence to suggest that any other species on the planet can actually think this thought. If they could, they would be almost as smart as we are.

When preparing for interviews, effort needs to be done upfront for what questions the interviewers will ask. Think about interviews as methods of getting the information you want from biased sensors. You need to structure the interview in a way that minimizes the bias, maximizes the information per question (which is a function of what you want to get out of the interview). There is an art to it because there could be relevant follow-up questions the interviewer needs to ask if they don't think the interviewee fully answered the question.

When it comes to layering a collection of skills in your personal monopoly, there
are some other things that I want you to think about too. Things like personality
traits, distribution channels, concrete technical skills, membership in some kind
of community (like what college you went to or a club that you're a part of),languages that you speak, geography, where in the world you live. We once
had a student in Write of passage who was an expert at coffee making in
Iceland. We’re looking for experiences that you have access to, maybe people
that you know, certifications, or credentials.

The quality of personalization is (probably) proportional to the log of the amount of data you have on yourself. Therefore, you should try and maximize the amount of data you have on yourself so personalization becomes better.

I like the thoughts in "The Bitter Lesson". A takeaway is to work on techniques that scale as a function of compute and data due to the exponential growth of both. I think one should apply to themselves. The more we can tie our own goals to compute (and other things that grow exponentially), the better.

"Why should being fat be any different from smoking? It should be our right to do either if we wish- but neither should be celebrated. Freedom of speech does not mean freedom of judgement.

We know that being fat is unhealthy, we know that obesity most commonly begins in young children or in adolescents, we know that children of obese adults are far more likely to become obese themselves, and we know that at current trends as much as 50% of all American children today will be obese in adulthood!10 Obesity is unhealthy, should not be celebrated, and it would be better for both fat individuals and society if it were less common."

Moreover, we need to move the conversation from it being a choice to it being an easy choice to not be fat. Right now, it is hard to not be fat, it is too easy to be fat. This is a societal/systemic fault.

I think this is interesting. I wonder if there should be an equivalent for other "societal harms". For example, Americans for Nondrinkers rights. Although less directly, people that do not drink will never drink and drive or pressure others to drink. Those that drink harm those that do not partake in this negative behavior.
"Americans for Nonsmokers' Rights (ANR), founded in 1976, is an American member-supported 501(c)(4) tax-exempt non-profit national lobbying organization, based in Berkeley, California, "dedicated to nonsmokers' rights, taking on the tobacco industry at all levels of government, protecting nonsmokers from exposure to secondhand smoke, and preventing tobacco addiction among youth. ANR pursues an action-oriented program of policy and legislation.""

Okay, so I am on the website that will show you how much time you have left to live. and I'm looking at it and I'm trying to map out the last few months and since I've turned 20 I've been growing at such a large rate. Ever since my 20th birthday something has changed. I quit my phone addiction. I lived in Wisconsin. I lived in San Francisco. I went to Spain. I made friends there. I started my website. I'm starting a blog now. Everything is being started now. And if I can map the last three months of progress to the next three months, I would be in such a good shape.

Thinking about the Heul lead scandle and how it ended up being not a big deal, I want to have a principle that whenever I hear news that is shocking, wait 7 days before having an opinion. It's probably BS and not worth getting worked up on.

There is this idea of: raising your standards for yourself will cause you to rise to them. Set your standards extremely high, it will be easy for you to make choices and you'll reach your standards. Some ideas for me are: be strict about my bedtime, this will give me better focus and energy. Be stricter about ending my meals. When waking up, wake up with energy and vitality.

In order to improve the way that I learn, I need to do something different than what is typically done in schools. I'm thinking about something that traverses these knowledge graphs or mind maps, and then learning from that, where the base unit of knowledge is not like a document or textbook, but some graph structure.

I need to be brave and venture out to get extraordinary results.

I have great energy after my run. I ran at around 15:30 and it's 19:48 right now. On typical days when I don't exercise I would feel both mentally and physically tired. Today, I just feel physically tired, but my mind is sharp. Exercise is good! It feels good. I don't need to run for hours, just 30 minutes, with running and walking in between, and I'm good.

I'm thinking about the calendar and how we keep track of time. I bet these units are not "optimal".
Right now, there are minutes, hours, days, weeks, months, and years. These are definitions that did not evolve to optimize for humans goals.
My calendar should not care about weather or the seasons, but something more useful.
Built into this new calendar should be different "phases". Just like how some version of the calendar.
I think a week should a specific length (in terms of days) such that a goal of a specific size can be achieved. For example, the length of a sprint.
I want to have a unit of time that is around the length of a quarter such that, at every one of these units of time, I am a significantly better and different person. 12 weeks (like in 12 week year).
Also, maybe days should be variable length? I should use the concept of "100 blocks of time" per day (which is N minutes) instead of a day being 24 hours.

I've had "apply to an entrepreneurship related organization" on my todo list for over a month now and I have not done it. I did not do it because it felt like it would have taken a long time to do so and I would need to be prepared. Just looking at, I took the plunge to do it today. I looked for places to apply to. I found Entrepreneurship First. I applied within 10 minutes. This whole thing took me a total of 35 minutes. I was overcomplicating it. and I could have potentially lost out on things. So this is another example of emotional reasoning. This is something that I identified being one of the most important things for me to do, and I did not do it. for a long time because of some invisible requirements.

"If I put a gun to your head and gave you a year to be able to bench press 315 pounds, you will either interpret that as impossible and accept your death, or you will do everything in your power to achieve it, and you probably will."

# How To Get Jacked

If you do 10-15 sets per muscle group each week, take each set close to failure (1 RIR), progressively increase the weight over time, eat anywhere close to 1g protein/lb. of lean body mass, and then ultimately cut to a lean body fat %, you will get jacked.

If on top of that, you train heavy compound lifts (squats, deadlifts, bench press, pull-ups) and follow a simple lifting plan like 5x5, you will get even more jacked.

If, after absolutely mastering all prior concepts, you want to spend enormous time and energy dialing in the exact rep ranges, isolation exercises, and supplements that makes the most difference to you, in order to fight tooth and nail for that last 10% of explanatory power over your results, then by all means go for it. But by this point, you will be so jacked that you could become a fitness influencer and sell your own supplements to the unsuspecting masses - a far better value proposition.

It seems that the concept of invisible requirements is also a large part of procrastination Before procrastinating, when you're procrastinating on something, you feel that there's many requirements for things that you need to do, but in reality they're not. For example, with my CS527 assignment, I thought that the requirement was for me to make a system that works end-to-end completely by itself and solves flaky tests, but I realized that that is not what the actual requirement was. The requirement was to contribute to open source projects.

So in order to be more agentic, I need to be aware of the invisible requirements that I set up for me doing something. If I want to do something, for example, one thing that I want to do is set up a McDonald's. I'm thinking about all the things I need. I need to get enough money to figure out a menu, then I have to make a menu, then I have to make a post, then I have to get a uniform. But all that is not required to set up a McDonald's. You could probably do it Like, an hour... with 20 bucks worth of food. And then if it's sold out, it's sold out.

I spent maybe 10 minutes of time total thinking about how I'm gonna tell the Perifrontics office that I'd like to cancel my appointment, thinking about what excuse I'll give them etc Emotionally this felt like a 6 out of 10 thing to do; however, in reality it was like a 2 out of 10. I just called them and I told them that I'd like to cancel my appointment And then they asked Would you like to reschedule? And I said no. And then they ended the phone. It was a couple seconds long.

When thinking about my cs527 homework assignment I was overwhelmed and was procrastinating. I explicitly set up a focus block by specifying what I need to do, how I will do it, why I am doing it, and I got more focus. I also felt less overwhelmed because I was only thinking about the next step. Before, when I think about how much effort this assignment would take, I would have estimated it being 9/1010. But it was actually more of a 4/10.

There’s this idea of reinforcement learning with verifiable reward, introduced in DeepSeek-R1-Zero – if you have some priors on the solution to a problem, you can enforce them as the verifiable reward. This can be exactly used for retrieval because the retrieval outcome is verifiable, which is just what is the recall. So that can reflect the retrieval quality, and you can train the retrieval module, which can be like the query rewriting or expansion task, so you can have an LLM improve retrieval outcome by rewriting the query. This is a reinforcement learning paradigm because the LLM’s environment is the text corpus as well as the query. Its action is how to rewrite the query, and then its reward can be the recall. So this is an exploration task instead of a supervision task. You can use reinforcement learning with verifiable reward to solve this problem, which leads to deep retrieval, which is named after deep seek. It learns to augment the query based off this reinforcement learning with a very popular word paradigm. Deep retrieval significantly outperforms industrial leading LLMs like GPT-4o or Claude Sonnet; it defeats all of them by a wide margin, with only using a 3 billion parameter LLM. This is because it was fine-tuned for the specific problem that it is – it’s not fine, it was trained on the specific text corpus, and then they use a reasoning model for deep retrieval, which improves the performance by quite a bit, does increase the amount of tokens used.
## Screenshot

Here’s how GraphRAG works: it starts with text and then chunks it, identifying relationships and entities using a prompt. The LLM extracts these relationships and entities, then puts them into a knowledge graph. It detects communities, summarizes each one, and then summarizes those summaries to get a higher-level overview. At query time, this knowledge graph can be used to identify which communities to extract documents from – because when doing a RAG on this graph, you can retrieve a community summary and gain a better understanding of each community, which can help answer more general questions like trends, instead of just extracting single entities. Some places where GraphRAG is beneficial include domains where understanding systemic patterns is critical, such as global threat networks, cross-disciplinary breakthrough opportunities, corporate strategy synthesis across thousands of reports, and legal discovery – uncovering document relationships and litigations. Another way is by simply identifying systemic patterns in your own personal life, I think that could be a benefit. Some problems with GraphRAG are that it's difficult for it to deal with high-ephemeral or high-velocity information because it takes time and effort to construct the graph. There's also the general problem of entity resolution and disambiguation; edge sparsity can be annoying – so many pairs can have weak or spurious links.

So I’m going through this paper on metacognitive reuse—essentially, turning recurring LLM reasoning into concise behaviors. My understanding is that you have a problem, and then you tell the LLM to solve it using reasoning. However, the LLM re-derives these properties—like the fact that multiplying a number by or dividing a number by a fraction is the same as multiplying by the reciprocal—and this takes many tokens. The Medic-Hungary for Use paper proposes a fix: the system stores declarative facts but doesn't support procedural knowledge, which we should do because it reduces token usage, benefiting both cost and latency, preventing context window clutter, and potentially improving accuracy. The key idea is introducing reusable reasoning fragments distilled from the LLM's own reasoning—essentially, metacognitive reflection. We’re aiming to build a behavior handbook for behavior-conditioned inference, behavior-guided self-improvement, and behavior cognitive SFT. The method involves having an LLM generate a solution to a problem, then prompting it to reflect on the solution, and finally prompting it to extract a reusable behavior from the question, solution, and reflection. For example, “behavior\angle\sum”—the sum of interior angles of a triangle is 180 degrees—is a behavior extracted from the LLM. Results show a significant decrease in token usage; at 1,500 tokens, accuracy is 90%, compared to 65% without this method. To achieve the same performance, the LLM might only need to increase token usage by 50%. This applies to behavior-conditioned inference, where the model can refine future iterations by adding to this behavior handbook, or through supervised fine-tuning, using the relevant behaviors from the handbook to condition the reasoning. A limitation is that the behavior handbook is static, and its generalization to other domains is unknown. Ultimately, this represents a metacognitive procedural memory for LLMs—the LLM understands its own cognition and learns how to perform these procedures, leading to the efficiency and accuracy gains observed because behaviors are compressed versions of verbose reasoning.
## Screenshot

Although I am chronologically 20, should I think about myself as being much older, such as 25? How would changing my identify change how I perform? Will I achieve my goals better, if so, then it is obvious I should change my identity.

Can i do conditioned taste aversion to myself: https://en.wikipedia.org/wiki/Conditionedtasteaversion

What if I poison myself to have an adverse reaction to sweets or food I don't want myself to eat so that I am biased to avoid it.

What i've notived at LBF is none of these companies or organizations ever talk about money, shareholder value, etc. They just care about solving the problems of longevity.

It is so surprising that indefinite lifespan is not more mainstream. I think this is the whole point of the LBF, but being in the circle more, I'm internalizing more and more the idea we should live indefinitely and it's weird not to think the other way. This is very cool science and engineering.

I was about to spend ~1 hour setting up firefly-iii on my computer, but I remembered that it is not the proper time and scheduled this activity for after my Spain trip! I am learning and getting better at time management!

Self control exercise:
To gain more self-control, select an area and a concrete goal. Choose two relevant self-control strategies, identify initial small steps, and create a reminder to implement these steps.

Area:
- Health/nutrition
- Do not binge eat on {almonds, finger food, protein shakes, huel, random food}, especially at night.


Ways to improve:
- Preportion "snack" size amounts. Limit myself to only 1 of these amounts (sidestepping temptation because there is not a large amount to view).
- If I notice me binging after I've eaten a meal, drink a glass of water and imagine the feeling of discomfort when I binge.
- I can have a ceremonious "start" and "end" to my meals. I can end my meal mindfully by doing some gratitude/prayer. I can also end it by brushing my teeth.
- Eliminate all 'very tempting food' (sweets)
- "Cold turkey" -> never accept food I will regret eating.
- Do not start my dinner late
- Verbally tell people around me that I have a problem with binging and I want to only eat X amount
- If I notice my cup of water is empty, I will fill it
- When I notice my desire to eat, realize I don't actually need to address this desire. I can just let it go away.

So there are nine different aspects of self-control. There are four broad categories. The first broad category is classic self-control, which is awareness of the temptation and the ability to override it—an example being when someone offers you candy, being aware that you’re about to eat it and consciously passing it up. The second genre is helpful preferences, encompassing high motivation, delayed gratification, and a lack of unhealthy desires. For example, being scared straight can lead to high motivation, like when you’re motivated to make a specific change; today the dentist died, and I now have very high motivation to floss because it means I’ll reduce the amount of cavities in the future. Delayed gratification is where you discount future rewards less than a typical person—if you’re okay making a sacrifice today if it’s beneficial, you can keep it for tomorrow. And then the last one is a lack of unhealthy desires; it takes no effort to avoid things like candy or sweets. Third, there’s pain tolerance—if you have high pain tolerance, you’re able to do things that are painful. Many beneficial things are painful in the moment, so you can do them without as much problem. Then there's knowledge to launch towards future goals—knowing you'll suffer in the future but making decisions that are ultimately beneficial. If you don't care about future pain during a marathon, you’ll sign up for it. Finally, there are momentum-related aspects: high energy makes it easier to avoid temptations, and flow—not realizing how much you’re struggling, so you don’t notice the pain, and it appears you have self-control when you don't. I relate to this when I run, especially longer distances; I just become less aware of the feeling of running.

🧹 Options to consider to free up time
Your money/time value suggests it may be worth delegating tasks to free up your time. Here are some services you might find useful:
👨‍💼 Personal assistant
Use services like Fancy Hands or Zirtual to outsource general admin tasks remotely.
📊 Accountant
Hire help to manage your taxes and finances with confidence.
🧺 Laundry service
Drop-off services can wash and fold your laundry for you.
🧽 Cleaning service
Free up hours each week by hiring someone to maintain your home.
👶 Babysitting
Consider booking a sitter to reclaim a night off now and then.
🚗 Errand runner
TaskRabbit can connect you with people to handle everyday chores.
💻 Freelancers
Use Upwork to hire support for one-off or recurring tasks.

Product idea: be able to better store information/extract information from people in an organization both through time and through people. This is what a knowledge managaement system should do. For example, my knowledge management system means I can query knowledge through time. There is a problem in organizations where "tidbits" of knowledge are stored in employees heads. This should be extracted into a database and people should be able to very easily query it. For example, tools, workflows, etc. They should automatically be suggested.

Add habit of "owning my mornings" and "owning my nights".
"A keystone habit I am noticing is to "own my nights" and "own my mornings" if I can do that, then that will set myself up for 1. Growth and 2. Productivity. I grow at night."

I was not being very essentialistic today because I spent a lot of time reading the Alpha Evolve paper and working on the presentation, but the presentations are the most important thing for me to do. The most important thing for me to do is my systems iteration and other tasks for my Spain trip, so I'm not prioritizing, but this is a very good thing because I really enjoyed working on this presentation. So that means I'm getting to the point where I'm not doing things that I would enjoy, so I'm getting rid of good opportunities, so I'm making room for better opportunities.

I envision a system that, if I could hire someone to review my body camera footage alongside my calendar and typed communications – essentially a life log – could parse through that data and identify observations that influence my systems. For example, last night I experienced an asthma attack and didn’t document it or take any preventative action. I’m not trying to figure out how to prevent future attacks, but rather to understand the root causes by leveraging the system to analyze my life log, potentially consulting scientific literature, and determine what habit changes I could implement to avoid such episodes and ultimately eliminate asthma altogether.
I also imagine this is running locally on my computer, 24/7. This makes me think about Alpha Evolve, how it was able to spin up a bunch of agents and its goal was to solve each of these problems. What if each of my systems is represented by some agent? And then when I want to make a decision, I can ask all of my systems what actions should I take next? And then I can have some meta-orchestrator that learns my values and can take into consideration the decisions from the agents. This is essentially a framework that will optimize my life for me instead of me optimizing my life myself.
I can imagine that it’s intelligent enough to, so it knows my values and it can reason about them. One sign that this system is actually very intelligent is if it can identify that one bottleneck is in fact its own abilities. If this system could realize that it could convince me to get more resources, then that would be good for me, and it would do that. It might suggest for me to get a better GPU for itself so that it can think smarter. It also might, if I can get something that’s recursively self-improving, like imagine I have this idea for this light optimization system, but if I could just get one system that can optimize itself and improve itself, then I wouldn’t have to think about all these different modules; it would be able to come up with those modules itself. So what if I have something like Alpha Evolve, and the judge of how good the system is is me? Or another, like, LLM that has learned my part is this. Yeah, how cool would it be if I could have this, like, a 24-7 growth and systems-improving agent, then I think I could very easily hit my goal of doubling my goal effectiveness every three months; I would envision the system to be able to double my goal effectiveness in, like, one day. Maybe it’ll take me a couple of weeks to get the system up and running and work out the bugs, but if I get an autonomous agent whose goal is to just make me the best person possible, then I think that would be extraordinary.

So I got an email back from Darko, which was an email I was really anticipating. I started reading it, but I realized that if I read it now at 11 o’clock, I wouldn’t be in the right position to respond, and I didn’t want to. So, to prevent this from happening, I should make my email client stop working after a certain time, or simply stop checking my email altogether, or turn off notifications after a specific time. I can come up with a better way to ensure I’m not checking my email at nighttime or doing other things when I’m intending to focus on something else.

By scheduling my time and knowing all my tasks – whether for tomorrow or just generally – I can be more strategic. It might be beneficial to create a task manager or agent that understands my habits, such as my preference for watching videos while I eat, and then schedules my day accordingly, prioritizing tasks aligned with my strengths, like creativity or problem-solving, when I'm most capable, and content consumption when I'm tired. This could be a useful addition to my systems.

This is a research tool that allows you to discover how papers are connected, and to identify other related papers. You can visualize this connection in a graph, which will aid in your searches. https://www.litmaps.com/

Watching this youtube video makes me want to work at DeepMind. They want to solve intelligence, and then use that intelligence to solve everything else. In this video, they showed how AlphaEvolve was able to improve it's own hardware and it's own algorithms.

I've probably spent ~10 minutes just buying my bus ticket. I notice when doing trip planning, there is a lot of task switching and I do not get things done. I should mono-task and have a specific framework where I identify tasks, then identify what I need to get done/know before doing that task, do this recursively, and then I do the actions.

When listening to this lecture, I wish I could've paused it and go back. I was taking notes and missed some parts. This is the disdain I have with live lectures. I think it is archaic and asynchronous would be much better for learning.

When doing model merging, a problem is that some weights were changed due to the stochastic nature of the training process. Another problem is when doing model merging, there could be some weights that were both changed in different directions. The solution might not be to do a simple averaging, that is where TIES (trim, elect sign, and merge) method comes in.

1. trim value that are too small.
2. elect the sign that has larger magnitude
3. then you do a disjoint merge based on what aligns with the sign vector

I want a device that listens during lectures as well as notes what I am writing and then can allow me to pull up the lecture transcript at the time I was writing something so I can be reminded what I'm reminded of. Take very short note + lecture transcript at time -> much more detailed, accurate, correct note.

The layers of complexity for learning are:
Non-associative learning - habituation of stimuli
Classical conditioning - unconscious learning; association between two stimuli
Higher order conditioning - multiple stimuli together, applying logic
Temporal/Spatial Sequence Learning - learning the order of events and locations; this often involves creating cognitive maps
Episodic Memory and Cognitive Mapping - remembering when and where events occurred, building rich, interconnected mental representations of the environment

If more context window means better performance in LLMs, then if we improve our own context window can we be smarter?

Listening to charlatans speak feels like watching a lecture but gaining no understanding. There is no understanding that underlies their speech but they mimic understanding.

Stimulation drive our hallucinations so they are grounded in reality.

SearchR1 involves training LLMs to reason and leverage search engines using reinforcement learning. Essentially, you integrate another LLM into a module that learns how to utilize a search tool to improve the responses of the coupled LLM. The training strategy consists of a query, followed by an LLM utilizing a search engine to return sequences, if necessary. This search output is then provided to a reference LLM, which outputs a response, allowing for validation of the response quality. The key idea of the paper is to fine-tune one specific LLM to be exceptionally good at using a particular tool, rather than fine-tuning only the reference LLM. You could reformulate the problem entirely, but the paper's approach is to have the LLM output specific tokens for both searching and answering; however, I suspect a completely new approach would likely be more efficient, theoretically.

When doing free recall/explanation after watching a lecture, it would be nice if I could have some automated suggestions, such as something that checks my reasoning, something that cleans up my explaination. I am giving an explanation and I am not sure it is correct. It would be nice if some teacher could correct my explanation. Someone who was an expert on the paper.

This is a project idea that I think could be powerful. I am reading [this blog](https://alastairreid.github.io/coreutils/) and realizing the developer had to write this post and share what they are doing. For every developer that shares what they do in an easy-to-digest format, there are probably 1k that don't share what they do. I'm assuming this developer has done other explorations that they did not share online that were less value. There are potentially millions of value-producing information shares that do not exist because people do not take the time to audit what they did and write it up. We can dramatically speed up development by sharing more information, especially coupled with sharing this with LLMs and powerful search. This information sharing can be applied to not just development, but all other fields, but development seems to be the easiest.
I imagine a software that records what you are doing, all of the commands you type, files you download, and then automatically generates a blog post sharing what you did. Intermitently, it can query you for your thought process and what you are doing, what fails and what doesn't. With this, we can share more information and generate things automatically.

Metamorphic testing is by leveraging metamorphic relationships of programs to identify errors. Metamorphic means "change-of-form" and it refers to how a program should transform when the inputs are changed. This helps solve the test orcale problem, where one does not know what the expected outcome of a software should be.
Mathematically, if $$x$$ yields $$f(x)$$, a transformation T produces T(x) such that f(T(x)) is predictably related to f(x) per property.

Without thinking about it too much, a goal of mine is to have an 80-year-old's mind while being 20 years old.

Semgrep’s product is essentially an LLM that detects vulnerabilities and performs triage, and its accuracy is 96%, which many people believe is too low. However, they argue that a few false negatives are acceptable, because you need a high signal-to-noise ratio for anything to be fixed; too much noise would prevent people from fixing the issues. The value isn’t that every single vulnerability is detected, but rather that to minimize the number of vulnerabilities within your application.

I think I the skill I want to get very good at is personalization. Personalization is broad enough, it is a business need, and I have a unique perspective on it. Personalization can be applied to longevity, productivity, most softwares.

Semgrep is a company with an open-source project called the Semgrep Supply Chain Open Source Dependency Scanner. What distinguishes Semgrep Supply Chain is its use of reachability analysis. Currently, the scanner identifies over 101 software supply chain findings across six projects, but by employing reachability, we can reduce that number to 16 high-signal findings. Semgrep code combines the customizability and speed of the open-source engine with more powerful code analysis, the ability to integrate into developer workflows, and necessary feedback loops to ensure vulnerabilities are detected and actually fixed. Once you’ve defined the rules producing actionable findings, you can use them to notify developers before code enters production; they'll receive notifications after creating their PRs and can interact with Semgrep comments to fix issues, triage the results, and provide feedback. Semgrip Secrets uses a three-prong approach – semantic analysis, secret validation, and improved entropy analysis – and Semgrep Assistant provides GPT-4-powered security recommendations to help developers quickly review and remediate security findings. For instance, Semgrip detected a SQL injection vulnerability, which the Assistant confirms is a true positive and suggests an auto-fix code snippet. Ultimately, Semgrep provides an engineer-first application security platform designed to remove developer friction, enabling the effective resolution of critical issues, and they’ve integrated LLM capabilities to address vulnerabilities.

A skill to have is to identify marketing terms. Terms like "ludicrously fast"

If you are not struggling when you are learning then you are not learning.

Time block: 21:40-00:00
Activity: watched YouTube

Log:

- Cognitively, how much energy did I have doing this activity? ([[estimate-energy-mood]])
2

- How much did I enjoy this activity?
5

Plan:
My plan for this time on my calendar was to do my night routine, go to bed.

Observations:

- What happened during the activity?
Reflection on last night, it was around 21:40, I got home from Grainger and I laid on my bed. I was tired, I made the choice to watch YouTube. What was going through my head was I just wanted to watch YouTube for 40 minutes then go to bed. I watched YouTube for 2 hours and 20 minutes. I started watching a video by Kurzgestat, then 2 videos by Veritasium. I wanted to watch the videos, they were in my recommended. I clicked on them, part of my head knowing I shouldn't click on them.

- How did I feel before, during, and after the activity?
I felt that it was okay for me to do this behavior. I didn't feel loss, I didn't think about my future self, I just thought about what would feel good in the moment. In my mind, I was thinking if I watch 40 minutes of YouTube I will feel better. During the activity, I felt the good mind-numbingness of YouTube videos

Evaluation:

- What went well and didn't? How did things not go as expected?
I'm happy I didn't go way past my bedtime. Most other things did not go well, I was even watching YouTube videos while brushing my teeth! There was an extreme resistance against me pausing the video. As soon as a video ended, my monkey brain wanted another one. This makes me feel disgusting that I can be manipulated like this.

- What surprised me or did I find challenging?
I definitely am doing better than before, cognitively I felt I had a choice but chose wrong, instead of in the past me not feeling I have a choice. Looking back, I'm also realizing I don't really enjoy the activity that much. It's not that I wanted to read because it's better long term, I also just enjoy thinking much more. The problem is that reading and thinking is more difficult.

Analysis:

- Why do I think things went well or poorly?
I did have a very productive day, I finished the most important task for that day, got more caught up with my lectures, paid my bills, did work on my research. I was tired at the end of the day. The thought of reading marginally crossed my mind, if there was a book in my bed, I bet I would have read it.
Because I was drained, it doesn't mean I need to watch some brainrot to recover, it just makes me more susceptible to these stimulants. This is why I've gone about a week without watching or consuming anything and have made a poor decision last night.
I also did this because of present bias.

Action:

- What is the single most impactful change I could make to improve this activity/system?
No phone in room after 22:00. I do this change when I have energy and before I lie down. There is nothing I can do on my phone that I can't do on my computer.
- What is the next concrete step I will take?
Logout of YouTube on my phone so I won't see recommended videos.
Have what I want to do on my bed, so I don't need to make a decision.

Questions:

I just solved a bug where a JavaScript function was expecting the variables named options and results, but I was passing chartData. I was under the assumption if I passed in a value that was not defined in the function then it would throw an error. This is not how javascript works, though. This took me an hour to find.

`js
const ScatterChart = ({ options, chartData }) => {
if (!options) throw new Error("options is required");
if (!chartData)
throw new Error(
"chartData is required, chartData is:\n" + JSON.stringify(chartData),
);
return ;
};

export default ScatterChart;
`

Reflection: This took well over 1.5 hours to find. What I did was ask ChatGPT why the bug occured. It gave some potential reasons of the versions being out of date, but the versions are in date. Instead of asking ChatGPT why the error occured, I should ask it "what are all the possible reasons this error could occur". Instead of giving me one answer, it can give me several. Then, I can step-by-step eliminate them. I can also look to see what error is most likely given my context.

A way to prevent this bug is to switch programming languages, that is not as feasible, but another way to prevent it is to mark your assumptions in your functions (i.e. xyz should not be null). The library I was working with should give a better error message.

Question:
How else can I improve my debugging skills?

I wanted to get caught up in tech, I watched 1 video from fireship, and I had the willpower to stop. I finished my CS527 lecture video and went to YouTube, but I had the resistance to not watch anything. I have significantly cured my brain from the desire to be entertained.

Time block: 11:00-15:00
Activity: call dentist people, do CS527 work, paid my bills, cooked + ate lunch, watched CS598JH lecture 3

Log:

- Cognitively, how much energy did I have doing this activity? ([[estimate-energy-mood]])
- How much did I enjoy this activity?

Plan:
My plan during this time was to finish my CS527 MP, and then watch my CS598JH lectures.
Observations:

- What happened during the activity?
I went to work on my MP, did some work on it, but then I got distracted by a repository I found. Newsflash, I just got distracted by shiny object syndrome, the repository is cool, but it was not worth derailing my progress.

- How did I feel before, during, and after the activity?
I felt high energy, although I was distracted, I didn't feel mentally drained or low energy

Evaluation:

- What went well and didn't? How did things not go as expected?
I quickly found a repository with many flaky tests, I think I can now finish my MP in an hour. I also remembered my reflection from before and meal prepped 4 meals, so I'm very happy my future lunches will be quicker.

Analysis:

- Why do I think things went well or poorly?
I had a good nights sleep and good energy which is why I think things went well. I had more confidence in myself to do hard things (call the dentist people).

- Where did I waste time, energy, or focus?
Distraction from email.

- Could I automate, delegate, or simplify any part of this process?
- Are there alternative approaches or methods that might yield better results?
- Did I leave anything half-finished?

Action:

- What is the single most impactful change I could make to improve this activity/system?
When there is a thread for me to go down, store it later, build confidence in my system that eventually I will see it, but see those things as non essential.

- If I did this again, what would I do differently?
- What is the next concrete step I will take?

Questions:

Someone named Kolby Shouse came to my apartment today and got me to switch my electrical provider to IDT instead of Ameren. He came in saying that he works at the electrical company and he needs to check my bill to see if we were double charged. I took this at first glance, he asked for my account number, phone number, name, and other information. And then he hoodwinked me by getting me to sign up with IDT. I canceled it, but I spent an hour learning a lesson that will be valuable for me in the future.

Some lessons/principles I learned:
Always ask why? If I feel like I cannot ask why, then I must end the interaction. The reason I was afraid is I thought he was working for Ameren in some way and I didn't want to waste his time, so I wanted it to be quick.
Don't let someone pressure you to make a decision to quickly. Always ask "can I do this another time". This is really hard when they are face-to-face with you, but if they say I have to do it now, then tell them to leave.
Always sus out their intentions when someone comes to your door.
If there is a warning sign in your mind, then raise the alarm. I had some suspicion when the website was asking me to ensure they said they do not work for the power company. I should have asked "why is it asking me this".
When interacting with people that you do not expect, behave with caution. Look intimidating, like you are not having fun. Don't be in a nice mood because you'll treat them with too much kindness.

Next, I’m going to be explaining another technique taught in class, which is Aspect-Based Reasoning Structure Extraction. The problem this is trying to solve is that some claims don’t have a specific yes or no answer. For example, the lecture gave the example of vaccine A being better than vaccine B, but there are many different aspects in which vaccine A could be better or worse than vaccine B. The insight is that you need to break down these claims into specific aspects and then identify which of those aspects have been explored within the scientific corpus and which have scientific consensus behind them. The algorithm works by starting with a claim and generating coarse aspects that would impact that claim, such as efficacy, safety, and distribution. Then, you look into your corpus of documents and identify which documents are related to the specific aspect you want to discover more about – in the vaccine example, it would be safety. You then rank those documents based on their relevance to safety so that the LLM can use them in the future. However, this is actually a hierarchy, because safety can mean safety for whom? Safety for children, adults, or the elderly, and you have to iteratively discover more sub-aspects. When someone asks, “Is vaccine A better than vaccine B?” the LLM now has a bunch of different aspects of that claim and an idea of what aspects have scientific consensus. It could respond with something like, “Yes, 80% of the corpus agrees that vaccine A is safer for adults than vaccine B.”
## Screenshot

So, I’m going to share another paper that was discussed in this class – it’s called KARE. It’s a way to predict the outcome of a patient based on their context, and also provides reasoning for that prediction. It works in three steps: first, you start by attaining knowledge sources – biomedical knowledge graphs, biomedical core sources, and other information from LLMs. Then, you extract a knowledge graph, doing semantic clustering on the embeddings of the entities, and grouping things into communities to create a bunch of different corpuses. In those corpuses, you have information about interactions and the general theme. The next step is to construct and augment the patient’s original context – you have the patient’s chart and EHR, construct a knowledge graph from that, and then look at the community summaries, appending information related to their specific cases or illnesses, as well as similar cases with different outcomes. The purpose of that is to see what cues cause what outcome. Finally, you prompt a smaller LLM with this augmented context to provide reasoning as to what will happen to the patient and then predict the outcome. As for training the LLM, you already have the patient context and ground truth in your dataset, prompting a larger, smarter LLM to provide a high-quality reasoning chain, and those are the training signals. This dataset is about 16,000 patients – each has about 1.6 visits, 20 conditions, 4 procedures, and 30 medications on their chart. It’s a relatively small dataset, but large enough to fine-tune a smaller LLM. The benefit of this algorithm is you have a smaller LLM (so better inference times, better generalization, more robust), you have the reasoning grounded in the medical knowledge graph.

## Screenshot

Next, I want to talk about a paper called JoSE. It stands for Joint Spherical Embedding – it’s similar to other text embedding algorithms like Word2Vec or GloVe. The key insight is that the usage space and training space of these algorithms aren’t aligned; Word2Vec is trained to maximize the dot product between two vectors if they appear closely together in text, but we typically use cosine similarity to assess how related two words are. So, there’s a disconnect there. A student designed a training strategy that just used the cosine of the vectors instead of the dot product. I’m confused because dot products are much faster than cosine similarity. Even if the training objective is perfectly aligned with the usage objective, I’d assume you could train the other algorithms more and get a better result. However, this author found that the Jose method was the fastest in training time compared to Word2Vec, GloVe, FastText, BERT, Foncar, and GloVe.

# Dynamic RAG

I'm going to be giving a voice note on Dynamic RAG, which is a technique that was discussed in CS598JH. The title of the paper is Dynamic RAG, Leveraging Outputs of Large Language Models as Feedback for Dynamic Reranking in Retrieval Augmented Generation.

To start, to understand how RAG works, you have a generator, a retriever, a corpus document, a query, and you start with this query, then you have a retriever that retrieves some N documents from the corpus. Then you pass those N documents to a generator, and then it gives you your answer. But there's a problem that this N is a fixed value, and what was seen is that the order in which the documents are shown to the LLM also matters. So this retriever does not have the ability to rank the documents based off the importance to the query. It's probably simple just like embed documents with a cosine similarity, etc. And then the purpose of this work is to have a re-ranker that is able to take those N documents that are retrieved from the retriever and then rank them in an order that is better for the LLM, that makes the LLM generate a better answer, as well as decrease the number of documents to a number K, which the re-ranker gets to choose, because more documents does not always mean better. So how is this, so what is this re-ranker? It is another LLM that reads the documents and then it outputs an order for the documents to be in, and it also outputs the number K, which is the number of documents that are going to be passed to the next LLM.

But how do you actually train this LLM, how do you train this re-ranker LLM to give you not only good documents from the Scorpus, but also how do you train it so that the LLM improves its other LLM? Well, what you do is you have this training setup where you have an LLM that is a generator and then you freeze the weights of that LLM so that the LLM does not get updated at all. Then you train an LLM, you train the re-ranker LLM to output documents and then the frozen LLM generator will just generate documents based off the, or generate an answer based off the documents you received, then you, the LLM will generate many, like the re-ranker will generate many different series of documents, many different rankings of documents, and then pass that to an LLM, the LLM will generate an answer, and then based off that answer you'll compare that to a ground truth answer, and then you'll get a reward signal for the original re-ranker. So from this, now you have a reward signal that then you can update the weights of the original LLM re-ranker, and in this paper they use DPO as the learning algorithm. They show that they get better results than similar sized LLMs with other RAG algorithms, and then the LLAMA 8-billion is able to outperform the GPT-4-0 without retrieval. And then other results that they've shown that their RAG is better. What was interesting to me was that sometimes the LLM could just save that, could set k equal to zero, and make it so that no documents are passed to the generator LLM. I thought that's interesting. I would not have expected the LLM to find that behavior, but it was able to find that behavior.

## Screenshot

Over 80% of data is in unstructured text data

Checkout this repository:
[https://github.com/khoj-ai/khoj](https://github.com/khoj-ai/khoj)
Your AI second brain. Self-hostable. Get answers from the web or your docs. Build custom agents, schedule automations, do deep research. Turn any online or local LLM into your personal, autonomous AI (gpt, claude, gemini, llama, qwen, mistral). Get started - free.

My mindset is "cooked" as some will say. It is the thing that is holding me back. This is one of the most important things I need to work on. I am partly surprised by this as I thought I had a good mindset, but internally, I have some major things holding me back.

On the phone with the Dental Insurance people. I realized I have a fear of talking with people on the phone, I take any gesture of theirs negatively. This is my current mindset and something I need to change/improve.

I’d like to read more of my journals and reflect on them, as I haven't been doing that enough lately. I want to be able to continuously cringe at my past self – how disorganized I was, how difficult it was for me to read, and my lack of growth over the last three months, so I can ensure I continue to improve.

I’d like to read more of my journals and reflect on them, as I haven't been doing that enough lately. I want to be able to continuously cringe at my past self – how disorganized I was, how difficult it was for me to read, and my lack of growth over the last three months, so I can ensure I continue to improve.

At the end of the day, it took me ~15 minutes to transcribe my phone capture to my laptop, and this was on a "slow" day. Ideally, this would not happen or only take me 5 minutes to capture my remaining thoughts of the day.
1) To do this faster, I can have more AI suggested features and better shortcuts
2) "Native" text to speech + "cleanup" prompts
3) Mobile application to mirror the real capture.

I’m thinking about CS598JH, and it’s a phenomenal class. I’ve learned quite a bit about LLMs and text mining with LLMs, and I’ve learned about different techniques that I think are interesting. I think the reason the class was so good wasn’t because of some inherent property of the course, but because I paused and took breaks when I was watching the lecture and I also did free recall on what was talked about. I wrote smart notes and tied them to other things, so I’ve learned more, and I’m probably only at 25% of my maximal learning capability. There’s some things I did not do like I did not tie as well with things that I’m interested in right now; it was more passive, and I was interrupted.

If you cannot answer why you are doing something, then you need to identify why you’re doing it, and if you don’t like that reason, then stop doing it.

Most people don’t have mission statements; they don’t know where they want to be in a few years, and that is a problem. Many companies have vague mission statements, and that doesn’t help them make decisions or inspire the people working there. Likewise, people do not have mission statements that inspire them or help them make decisions. I don’t have a mission statement that I think about, and I think that’s one reason why I feel like I don’t have a purpose or am not focused.

Everyone needs to have an aim; those who don't have one and just hop from job to job without a larger plan or picture won’t be able to achieve the highest level of contribution or specialize, and that’s important.

The purpose of a vision is to help you eliminate options and make decisions. it is not some wishy-washy manifestation thing. I think this is how manifestation "works'.

A lack of purpose leads to caring about non essential things, things that most people care about like social status, money, Twitter followers, instead of things that actually are valuable.

A core skill to have in life is prioritization

for essentialism system, consistently identify activities I do and rate them on extreme criteria. anything that's below a 9 fails. I can do this with all types of decision making.

This technique called SARG, and the purpose of SARG is to allow LLMs to do better reasoning. I wanted to extract a principle of how LLMs can be used from this paper. The idea is that you have this corpus of documents, and then you extract triples of cause, relation, and effect to construct a causal graph. From this causal graph, you can then use an LLM to learn why XYZ happened, by querying it with a question. I think this is a powerful principle that can lead to other interesting research topics and results.
## Screenshot

Principle: let all work that you do be good enough to show others.
All these assignments, essays, homework, they are for learning, but you spend so much time, why not spend 10% more effort to make it good enough to be shown to others? Why not be strategic and align your assignments with what you want? I have identified the teams i want to work on, my best choices of projects, now i am thinking how can i work on this homework assignment, and have it be evidence.

I've taken so many courses, yet there are no visible fruits. You have to just "trust me" that I have learned the skills. Turn homework assignments into course projects.

So the whole idea of text mining with LM is shown very well in this example. It’s called RolePred: argument role prediction. The goal is to extract a predefined schema of different arguments from a document, like magnitude, location, or date. You don’t want to use a large language model for efficiency. So you extract numbers, dates, and other phrases like names. Then you have a template, and you say, “According to this, the blank of this event is this entity.” When I say blank, I mean these are the different arguments you have. The value that has the highest congruence is the correct value for that argument. You’re not just asking, “Oh, what is the location of this earthquake?” You’re doing this in a smarter way, and you could use a smaller language model and get better accuracy.
## Screenshot

An interesting idea is that this LLM is very general, but you are defining a structure to the LLM's output, and because you're giving more structure to the LLM's output, and you're doing more discrete searching, you get more power out of it.

I do think that this method is very interesting. It's called ontotype, ontology-guided entity typing. The goal is to identify what an entity is. I thought it was interesting because it uses a small language model and it shows the efficiency in understanding and extracting the knowledge of a very small language model. How it works is that you have an entity that you want to identify its type. Then you prompt an LLM with a statement about what that entity is. So in this example, it is Wrigley Field. So you ask Wrigley Field the location. Is it a person? Is it an organization? And then you get some sort of confidence score from that statement. You get that from the small LLM. And then once you’ve identified what class of entities it is, you can go down the street. So you know it’s a location. Is it a country? Is it a city? Is it a building? You get the highest score from the confidence. So then you go down again. I think this is interesting because I can directly apply it to my note system. Instead of prompting an LLM, I can do it more efficiently. Instead of asking an LLM, what is this type, I can have this action. And I think a cool idea is that you progressively refine the type so you can stop at any time if you want a higher level idea of what an entity is.

## Screenshot

Hypercube Rag is a technique for doing retrieval where you enumerate the different possibilities for different semantic types and then you essentially do a union on all the different semantic types to get documents. For example, if you want to do retrieval related to tropical storm disasters, there are different semantic types for disaster, like location, event, and theme. A tropical storm has a location where it happened, it has a theme which could be the name of the tropical storm, and it also has the event of the tropical storm, like Hurricane Katrina. And there could be themes for the documents, maybe it's about how there was so much rain and it flooded, or it could be that there was a lot of wind. And these are semantic themes that could just segment the documents. And then from that you have a query. It's very easy to find the documents related to the theme if they're indexed in this way. I don't understand what is so special about this technique though.
## Screenshot

So this one training method they use is called SearchR1. Essentially, you train an LLM while using the search algorithm instead of adding it post-hoc, and this leads to better performance, specifically for searching, which is unsurprising to me. And then there’s also this other algorithm that I mentioned, where you train an LLM to improve the search query before passing it into a search engine, and then the LLM can augment the search query. The LLM is essentially trained to have a search query that maximizes the F1 score. This allows it to have better retrieval, and then better retrieval leads to much better performance, which is shown by one of the graphs where they have DeepSeek, a 3-billion-parameter DeepSeek, outperform GPT-4.0 and Cloud Sonnet when it comes to information retrieval.

For information retrieval, understanding the intent and context of the user is important for good retrieval.

There is a real difference between theme-based and QA retrieval. Theme-based requires documents that might not have the specific keywords you're looking for. You need to do reasoning on a query to know what is relevant, some things that you might think would be relevant aren't and vice versa. For the example with the trump rally shooting, the political landscape at the time would be relevant, knowing about the shooters friends might be relevant or might not be. That is why this is a hard problem.

Retrieval is an important topic for an LLM because information not in it's training data exists. Better retrieval means better results for LLM.

When you make a change in an open-source program, one way to increase the odds of acceptance is if you cite that the current change is similar to another change that was accepted.

Time block: 14:30-15:30
Activity: Eating and reading Essentialism

Log:

- Cognitively, how much energy did I have doing this activity? ([[estimate-energy-mood]])
5
- How much did I enjoy this activity?
5
Plan:
During this time, my plan was to eat and read. nothing more specific
Observations:

- What happened during the activity?
I ate a pre-packaged meal my parents made for me. That was not filling enough, so I also ate a protein shake and some Huel. I was reading Essentialism. I was voice recording my thoughts. Afterwards, I was not ready to work so I went on another walk.

- How did I feel before, during, and after the activity?
Before I had quite high energy, a little frusturation because I didn't work on the most important task, I worked on an extension for Neovim instead of working on what is essential.
During the activity, I felt that I had clarity.
Afterwards, I feel a little sluggish.

Evaluation:

- What went well and didn't? How did things not go as expected?
I won't comment on the food consumption because I was listening to my body. I was reading while I was eating, so I definitely ate faster and more than I wanted to (which is probably why I feel sluggish and tired now). I read my book well and I'm happy I did it instead of bingeing on my phone or working on my laptop. I used my lunch as a way to get space. I want

- What surprised me or did I find challenging?

Analysis:

- Why do I think things went well or poorly?
- What underlying factors or assumptions influenced the outcome?
- What skills or knowledge did I use effectively? What was lacking?

- Where did I waste time, energy, or focus?
I was eating and reading. Wasting time by writing down by hand.

- Could I automate, delegate, or simplify any part of this process?
If I was reading digitally, then I wouldn't need to manually flip pages, I could just process the information like a computer. Actually, that would be quite helpful. There is a pain in flipping pages.

- Did I leave anything half-finished?
I didn't finish my chapter

Action:

- What is the single most impactful change I could make to improve this activity/system?
I want to ensure that when I am reading, the notes I am taking are effectively captured.

- If I did this again, what would I do differently?
I would like to eat outside, there is higher ROI for eating outside and reading. I get fresher air, sunlight.

- What is the next concrete step I will take?
Schedule my next lunch so it is outside.

Questions:
I wonder if it's better to eat and be bored instead of eating and reading?

I want to think about my learning system and what's the best way to learn from my books, process it, etc. I want to read "Building a Second Brain"

It would be very nice if I could just free speak, then my zettlekaasten notes get automatically segmented and captured.

I was working on my plugin to do system iteration for >1 hour and it doesn't work. I was working on it instead of doing the actual value add--my systems iteration.

Principle: When improving, ensure you're getting better in a way you want, if not, course correct.

When programming with Windsurf, something I realized is I am using it to do things that I do not know how to do. I am getting better at prompting Windsurf and using it, sure, this is applicable to other LLMs, but I am not getting better at things that matter (better system design.)

When I have principles, they should be explicit, specific, and not vague.

There is someone at Grainger that is taking their meeting on the second floor, the quiet floor. I was sitting next to him and he was annoying me, so I just moved. I feel bad, because I wanted to tell him to not take his meeting at grainger and for him to move, but I didn't have the self confidence to o so. I was rationalizing why I shouldn't tell him to move, I didn't want to feel like an asshole, I thought he was going to stop soon, I didn't want to build bad blood. I want to think of a better way of confronting people and build up the ability to do so.

So part of it was initial fear and discomfort. I think that came from me being inauthentic because Siri told me that I should just go up to her, but there’s no reason for me to go up to her. Then Siri suggested that I should tell her that I want to walk in with her instead of just talk with Siri and her friends, and this was a very good reason, and it is true, so I felt much more confident about that. In my mind, I was envisioning a playfulness of me going up to her table, sitting down, and just like teasing my friends that I actually have to do work and kind of smiling at her and letting her know that I’m just interested. That’s a version of myself that I envision, and that’s how I will be in the future, and that’s a version of myself that I want to be. So I think the reason why I did not initially do it was because of a lack of authenticity, and then once I had an authentic reason for going to her to do it and actually I think the main reason maybe was not part of it was definitely the authenticity, but another reason is probably the playfulness. I did not have a playful reason to go up to her and talk with her, but once I had a playful reason, I was in a much better mindset. So I think that in these types of cases, I should try to think of a playful way to approach people. People vary like the playfulness, and I’ve seen this in my own life when I’ve approached girls in the past. So in the future, the next time I am in a position where I will actually want to approach a girl, I will identify what’s an authentic reason I want to go up to her and talk to her, and I will try to think of how I can make it playful.

I realization I had was that questions are extremely useful because they prompt your brain to think in a way, therefore, by systematically asking yourself good questions, you can get good answers. The answers are latent inside of you, and you need questions to fish them out.
## File
[Attachment](media/audiomicrophone175747596936720250909224610_029.wav)

Here is a resource for identify popular causes of flaky tests https://mir.cs.illinois.edu/lamyaa/publications/fse14.pdf

Time Block: 16:00-18:00
Activity: At home, eating, working very passively
Log:

- Cognitively, how much energy did I have doing this activity? ([[estimate-energy-mood]])
5
- How much did I enjoy this activity?
5
Plan:
The problem was during this time block, I explicitly did not make any plans. I was just at my apartment doing whatever work felt productive. That was mostly learning more about Anthropic, consuming content.

Observations:
I realized that I don't like the feeling of passively consuming content. This is what causes a mind fog for me.
I was at my apartment, eating throughout the time period.
Feedback:
A problem identified is I ate at 15:00 and at around 17:30. By not having set times to eat, I ate more than I wanted to, now I feel bloated and that I overeat. A positive is that I haven't felt this way in 3 days, so I am definitely improving!

Not only does having a defined time to eat improve my eating, how I feel when I am done, but it also improves my output because I can spent less total time eating, cleaning, etc.

- Where did I waste time, energy, or focus?
- What is the single most impactful change I could make to improve this activity/system?

If I were to plan my meals out before hand I bet that would improve this activity. This is because I'm more likely to stick to a plan if it exists. Also, I could add my meals to my calendar and try to be more "strict" about those eating times so I know when it is time to eat or not. I think I was eating out of boredom.

- Could I automate, delegate, or simplify any part of this process?
- If I could do this activity again, what would I do differently?
- Are there alternative approaches or methods that might yield better results?
- What is the next concrete action to improve this system?
- Should I experiment with a small change or a larger redesign?
- Did I leave anything half-finished?

Time Block: 16:00-18:00
Activity: At home, eating, working very passively
Log:

- Cognitively, how much energy did I have doing this activity? ([[estimate-energy-mood]])
5
- How much did I enjoy this activity?
5
Plan:
The problem was during this time block, I explicitly did not make any plans. I was just at my apartment doing whatever work felt productive. That was mostly learning more about Anthropic, consuming content.

Observations:
I realized that I don't like the feeling of passively consuming content. This is what causes a mind fog for me.
I was at my apartment, eating throughout the time period.
Feedback:
A problem identified is I ate at 15:00 and at around 17:30. By not having set times to eat, I ate more than I wanted to, now I feel bloated and that I overeat. A positive is that I haven't felt this way in 3 days, so I am definitely improving!

Not only does having a defined time to eat improve my eating, how I feel when I am done, but it also improves my output because I can spent less total time eating, cleaning, etc.

- Where did I waste time, energy, or focus?
- What is the single most impactful change I could make to improve this activity/system?

If I were to plan my meals out before hand I bet that would improve this activity. This is because I'm more likely to stick to a plan if it exists. Also, I could add my meals to my calendar and try to be more "strict" about those eating times so I know when it is time to eat or not. I think I was eating out of boredom.

- Could I automate, delegate, or simplify any part of this process?
- If I could do this activity again, what would I do differently?

- Are there alternative approaches or methods that might yield better results?
Going to work at Siebel, outside somewhere, not forcing myself to work if I do not feel capable of working, I can just read.

- What is the next concrete action to improve this system?
Add lunch and dinner times to my system.

Time Block: 16:00-18:00
Activity: At home, eating, working very passively
Log:

- Cognitively, how much energy did I have doing this activity? ([[estimate-energy-mood]])
5
- How much did I enjoy this activity?
5
Plan:
The problem was during this time block, I explicitly did not make any plans. I was just at my apartment doing whatever work felt productive. That was mostly learning more about Anthropic, consuming content.

Observations:
I realized that I don't like the feeling of passively consuming content. This is what causes a mind fog for me.
I was at my apartment, eating throughout the time period.
Feedback:
A problem identified is I ate at 15:00 and at around 17:30. By not having set times to eat, I ate more than I wanted to, now I feel bloated and that I overeat. A positive is that I haven't felt this way in 3 days, so I am definitely improving!

Not only does having a defined time to eat improve my eating, how I feel when I am done, but it also improves my output because I can spent less total time eating, cleaning, etc.

- Where did I waste time, energy, or focus?
- What is the single most impactful change I could make to improve this activity/system?

If I were to plan my meals out before hand I bet that would improve this activity. This is because I'm more likely to stick to a plan if it exists. Also, I could add my meals to my calendar and try to be more "strict" about those eating times so I know when it is time to eat or not. I think I was eating out of boredom.

- Could I automate, delegate, or simplify any part of this process?
- If I could do this activity again, what would I do differently?

- Are there alternative approaches or methods that might yield better results?
Going to work at Siebel, outside somewhere, not forcing myself to work if I do not feel capable of working, I can just read.

- What is the next concrete action to improve this system?
Add lunch and dinner times to my system.

Time Block: 16:00-18:00
Activity: At home, eating, working very passively
Log:

- Cognitively, how much energy did I have doing this activity? ([[estimate-energy-mood]])
5
- How much did I enjoy this activity?
5
Plan:
The problem was during this time block, I explicitly did not make any plans. I was just at my apartment doing whatever work felt productive. That was mostly learning more about Anthropic, consuming content.

Observations:
I realized that I don't like the feeling of passively consuming content. This is what causes a mind fog for me.
I was at my apartment, eating throughout the time period.
Feedback:
A problem identified is I ate at 15:00 and at around 17:30. By not having set times to eat, I ate more than I wanted to, now I feel bloated and that I overeat. A positive is that I haven't felt this way in 3 days, so I am definitely improving!

Not only does having a defined time to eat improve my eating, how I feel when I am done, but it also improves my output because I can spent less total time eating, cleaning, etc.

- Where did I waste time, energy, or focus?
- What is the single most impactful change I could make to improve this activity/system?

If I were to plan my meals out before hand I bet that would improve this activity. This is because I'm more likely to stick to a plan if it exists. Also, I could add my meals to my calendar and try to be more "strict" about those eating times so I know when it is time to eat or not. I think I was eating out of boredom.

- Could I automate, delegate, or simplify any part of this process?
- If I could do this activity again, what would I do differently?

- Are there alternative approaches or methods that might yield better results?
Going to work at Siebel, outside somewhere, not forcing myself to work if I do not feel capable of working, I can just read.

- What is the next concrete action to improve this system?
Add lunch and dinner times to my system.

Time Block: 16:00-18:00
Activity: At home, eating, working very passively
Log:

- Cognitively, how much energy did I have doing this activity? ([[estimate-energy-mood]])
5
- How much did I enjoy this activity?
5
Plan:
The problem was during this time block, I explicitly did not make any plans. I was just at my apartment doing whatever work felt productive. That was mostly learning more about Anthropic, consuming content.

Observations:
I realized that I don't like the feeling of passively consuming content. This is what causes a mind fog for me.
I was at my apartment, eating throughout the time period.
Feedback:
A problem identified is I ate at 15:00 and at around 17:30. By not having set times to eat, I ate more than I wanted to, now I feel bloated and that I overeat. A positive is that I haven't felt this way in 3 days, so I am definitely improving!

Not only does having a defined time to eat improve my eating, how I feel when I am done, but it also improves my output because I can spent less total time eating, cleaning, etc.

- Where did I waste time, energy, or focus?
- What is the single most impactful change I could make to improve this activity/system?

If I were to plan my meals out before hand I bet that would improve this activity. This is because I'm more likely to stick to a plan if it exists. Also, I could add my meals to my calendar and try to be more "strict" about those eating times so I know when it is time to eat or not. I think I was eating out of boredom.

- Could I automate, delegate, or simplify any part of this process?
- If I could do this activity again, what would I do differently?

- Are there alternative approaches or methods that might yield better results?
Going to work at Siebel, outside somewhere, not forcing myself to work if I do not feel capable of working, I can just read.

- What is the next concrete action to improve this system?
Add lunch and dinner times to my system.

What claude code is good at:

## Stages

1. Discover
2. Design
3. Build
4. Deploy
5. Support & Scale

### Activities

- Explore codebase
- SR Plan project
- Implement code
- Automate CI/CD
- Debug errors
- Search .
- Develop tech specs
- LU CTR G L Configure
- Large-scale refactor
- documentation
- tests
- environments
- Onboard & learn
- Define architecture
- Clegte coMis
- Manage deployments
- MCIiEorRage &
- performance

Participatory design includes users directly on the design team - participating in needfinding, proposing design ideas, helping with evaluation. This is particularly vital when the target users have much deeper domain knowledge than the design team. It would be unwise to build an interface for stock trading without an expert in stock trading on the team, for example.

Contextual inquiry is a technique that combines interviewing and observation, in the user’s actual work environment, discussing actual work products. Contextual inquiry fosters strong collaboration between the designers and the users.

When developing products for users, iterate and iterate cheaply. This is also a core principle of Scrum.

I'm contemplating the finitness of my life using this visualization and I am looking at other accomplishments that people have made at various ages. I was surprised to see that a lot of others have made major accomplishments much later than I thought. For example, Steve Jobs was 21 when he started his company, Bill Gates, 28. This makes me feel a lot better. It also makes me come to the thought of how every week of my life is permanently fixed. This last 2 weeks will stained by my bike accident, in 2 weeks, I will remember that time as my trip to Spain. These are periods in time I will be able to vividly see on a map. I still have ~3-4k weeks left of my life, and I can choose what those weeks look like. This is a powerful realization.

Last night, I spent ~30 minutes watching YouTube videos on death. I am happy that I was thinking about death, but I was on the border of the point of binge watching again. I am reflecting on how I spend my time last night, and I am reminded that my time is less valuable at night time compared to in the morning. Because of this, I want to make more of an effort to move my sleep schedule backwards. (side note, I think this reveals something about how I perceive time)

I woke up this morning feeling hungrier than yesterday. My weight is still 183, which is quite a significant drop compared to what is was a couple days ago. I am happy my weight is lower but I don't want to lose too much weight and feel too out of control.

I went to bed hungry last night and I had a very good sleep because of it! I am happy my sleep is solidifying to be consistent and high quality.

I would live differently if I completely internalized my own death. I am not aware of my own death enough. There's a popular thought that "would you live your life differently if you were aware that this is your only chance at living, would you be afraid if you knew life was a suicide mission".

From this video, I consciously know that I will never experience my own death. I will only be mortal.

I am trying to become more aware of death by watching this youtube video. There are many regrets of the dying, they want to tell a friend they forgive them for their transgressions, or they say sorry to someone they've greatly harmed. A goal in my life is to have no regrets when I am dying.

[talking about my cellulitus, Maks putting honey on my open wound claiming it would be healed]
Maks: "You will see you will be healed because of the honey"
Me: "I will be healed because I've been on rounds of antibiotics"
Maks: "You didn't even need antibiotics, you could've eaten garlic and honey and you will be healed"
Me: [laughter]
Me: "Can I quote you on that?"
Maks: "You can quote me on anything, Plato didn't write a word"

For my CS527 MP. The skill that I was trying to develop was debugging. I spent ~15 hours on the MP. If I would have spent 1 hour learning how to debug better, I would have finished quicker. This is a general principle I can apply to future assignments--to spend more time up front learning instead of just doing.

For my getting-things-done system what if I were to add a task to do in todoist, then I plug in the effort, then it automatically schedules time for me to do the task in my calendar so I don't forget to do the task + save time in scheduling

Islam Shehata gave me the nickname of "150 grams" in reference of me eating 150 grams of protein powder in a day. But he says if I become a rapper and use that nickname then I owe him 5% royalty

What if I build a company around personal growth tools, optimized for myself – personal optimization.

An idea I have is that I should envision my life in the future and how it would be if I were to gain specific attributes such as high confidence and reflection, and also envision how my life would be without that feature so that I am motivated to change that part of myself.

The book recommends these books to read: Zen: The Reason of Unreason; the wisdom of Confucius; the holy Bible; Tau, to know and not be knowing; the meaning of the glorious Quran, and an explanatory translation; The Book of Mormon; the meditations on Marcus Aurelius; and The Up and the Shades.

A theme of this book is explicitly making space in your schedule.

Humans need a lot of time to think about big decisions or big projects, and only through a lot of thinking can we come to the best solution.

The author of Essentialism has a practice where he blocks off 8 hours a day to write for this book. I think this is a good idea; it might help me to adopt this idea.

There's a connection between deep work and this book, and that idea of having space to think about your problem, like in solitary. An example is Isaac Newton: he became a recluse for two years to think about his problem, and then he came out with Principia Mathematica.

There's a course at Stanford called Designing Life, Essentially. The purpose of this course is for people to have a block of time where they're forced to think about their lives. That seems like a good idea. I should tell Natalia about this.

In whatever company organization I have, I need to ensure that I have time to take a step back and not be too busy. I need time to think and be strategic.

In the essentialism system, there's an explicit time and period to explore the possible space. In that time, it's essential to have spaces to think. I’ve had my own realizations myself when I realized that if I'm too busy to reflect on things, then I'm just too busy. They have a meeting a month where everyone is just in a room and they do nothing. If people can't make it to that meeting, then it means that they’re too busy or there aren’t enough people. So, because it’s too busy, it gives the founder a good signal as to if people are working on things that are unessential or not, because it is essential to have time to think.

A good indicator of if your actions and habits are good is if you were to think about. if you were to repeat those actions for the rest of your life, where would it leave you. you have to think more than just the present moment and think about the future, understand where your future self will be continuing those actions

For my weight loss system, I eat smaller bites and drink smaller amounts of food at a time. I've realized that I take large bites and eat quickly.

by eating slower, I can sense what I'm actually full

After falling off my bike, I have chest pain when I breathe in deeply. It's not a lung or chest pain; it feels like a bone and chest pain. The pain seems to be located in the front of my chest, around my rib cages.

My ideal job would be spending the day thinking and improving my thoughts.

It's interesting how social media applications can capture your attention so quickly and entirely when you start using them. Their UIs are designed to achieve this. I find it disgusting.

These are the root causes of flaky tests, I am surprised that so many happen because of async wait.
## Screenshot

We tend to hallucinate when we are sensory deprived because the environment doesn't give us information. This is a difference compared to ANN where that cannot hallucinate. There is a benefit to hallucination. There is a connection with this a dreaming.

There are biological equivalences of factors of artificial neurons. For example, the "weight" can be thought of how much myelination there is.

The system that prunes neurons is the brains immune system. I wonder if there is a connection to sleep here.

What "plasticity" looks like in brains is that neurons are just wiggling around and looking for more connections.

In each section of the brain, the neurons evolved to communicate "optimally" for that region. That means, I would expect to see different neuron structures in different regions of the brain. I wonder what are the structural differences between neurons that process different pieces of information.

Neurons are cells, there are many many complex pathways that happen within a cell. I am amazed at how good the brain is at computation. How is it possible that computers are not as good as brains at computing. This is surprising for me.

Differences between artificial neural network and real neural networks:

Real neural networks have temporal dynamics, neurons have impact on their environments. There are also different neurotransmitters that impact neurons at different times.

In real neural networks, they are adaptive, so the structure can be constantly changing. There are also other "data processing units" in the brain.

I realize that a lot of the time I am not really propelled by wants, I am just "floating around". What I mean by this, is I take a class initially because the topics sound very interesting, but after a few lectures, I get into a passive state where I am not motivated to do well. I don't feel "active" in my classes.

I've noticed many people at UIUC find jobs by waiting "for jobs to find them" instead of actively finding jobs they actually want. This includes things like waiting for notifications from job boards, Handshake, etc. it is a very passive way of finding a job and it leads to mediocre results.