2025 year in review
Value updates
I find it helpful to keep a doc describing my values. (It’s like the Model Spec, but for humans.) The primary benefits I see:
- It lets you spend less time thinking through the same tradeoffs on individual decisions. For example, I didn’t know if I wanted to spend money on DoorDash vs. investing upfront and recurring time into learning to cook. At some point I got fed up with thinking it through every time and just wrote the decision down.
- You get the chance to think through tradeoffs more deeply. I think this is helpful to avoid falling into the trap of doing something because it’s the lowest-effort pathway. There are a bunch of areas (relationships, what I care about in a career) that I wouldn’t have thought through if I hadn’t tried to write a fully self-consistent, comprehensive values doc.
I’ve been doing this for a few years, but here are some updates I made this year:
The material
- Aesthetics are key (aka, it’s not always worth being a utility maximizer). If I look at something multiple times a day, I want it to look nice instead of stressing me out.
- Consider buying a Vespa for the memes, even though it’s slightly more expensive and annoying than Lyft Bikes + Muni + Waymo/Uber.
- Get a nice apartment (though hopefully not too expensive).
- Get a nice and fuel-efficient car, not just a fuel-efficient one (I haven’t done this one yet)
- Vibes are key. Also I want to get better at dancing.
- Drinking is good (to the worried: in moderation!). My friend started dragging me to bars this year, and I admire:
- How novel bars are compared to most restaurants, for which I feel I usually know which food is best. (Not to mention how different bars are in architecture and design, whereas restaurants feel much more samey.)
- How good of an excuse it is to hang out with friends :)
- Slowly I should try to learn all the types of alcohol. Could be a fun side project.
- Updating a bit on travel being good:
- It exposes you to changes in low-level stimulus, which can break you out of personality basins.
- You get to spend time with friends and pay full attention to what’s going on.
- Regarding money: first spend money on things that compound, not single experiences. For example:
- In:
- Co-purchasing a vespa with a friend, which naturally compels me to hang out with them and visit them biweekly to drop it off
- Trips with friends
- Out for now:
- Expensive dinners (just have a normal dinner; unless it enables a fulfilling discussion with other people who would appreciate the price!)
- Business class trips (you only get the benefit once)
- In:
Relationships
(Friendship and otherwise)
- I want to play life cards-up. Essentially: keep minimal secrets, mostly because it makes life more complicated. No security by obscurity; the best strategy should win even if your opponent knows it perfectly. (E.g., I want to be open about sharing personal anecdotes even if they’re a little embarrassing.)
- I really like having fulfilling conversations with people, and a very small number of people I’ve met are clearly >10x (maybe 100x?) on this criterion.
- This is the closest concept I have to a sparkly person. I want to seek out these people, and I also want to improve at conversation so I can find the thing that sparks these conversations.
- This is probably the biggest thing I value in a romantic relationship and is also super important in friendships. (conditioning on my other romantic criteria, I think I’ve met ~4 people in the last 2 years who have satisfied this which is pretty low!!!)
- I want to create friendships of virtue with people I admire, and fewer friendships of convenience and utility.
- Is this the kind of person that, if they were old and ugly and had soiled their pants and needed your help, then you’d still say, Yes? “Yes, I would love to. I want to give this person my care so they can unfold their life with dignity.”
- Honestly, I think life is too short for the other kinds. Also, friendships of virtue follow naturally from being a long-horizon agent; unbuffeted by short term issues, you should find people whose core tendencies bend toward progress/kindness/success and support them wholeheartedly.
- Relatedly, I also want to learn how to give someone my full attention. I try (don’t look at my phone while talking). But this is clearly not enough.
- Someone who I dated briefly mentioned that they felt worried about our time spent together competing against all the other things in my life. This is what I really don’t want to happen. If I choose to spend time with someone, I would like for it to feel unconditional.
- I draw inspiration from the exact 1 person I’ve ever met who makes me feel seen when I talk to them, even though we only meet glancingly every few months.
- Potential ideas:
- More trips where I drop all obligations; don’t check slack, don’t think about work, etc.
- Drop recurring relationships that are not of virtue.
- Practice full presence at parties, 1:1 interactions, etc. Would be curious if anyone has better ideas.
- I want to be good to people. I think this means doing both little and big things for others. I draw inspiration from the small things I’ve seen people around me do, which I confess I never learned growing up:
- Spending time to draw a bookmark and organizing a “writing gratitude cards for friends“ event (I am very sorry for not attending)
- Packing into the back of an uber when there’s 3 people instead of sitting (more isolated) in shotgun
- Getting me Chipotle when picking me up from the airport, and providing a concrete day’s agenda so we avoid decision fatigue
- I think part of this involves having a little bit of slack in life. So I’ll try to do that next year.
Reader’s Digest
(In the style of kipply)
LLMs. DeepSeek R1. o3 is out. GPT-5 chart crime. How many bits are learned in posttraining? 100 parameters can recover 50% of the performance of a finetuned model. Models generalize from reward hacking to misalignment. Chains of thought are surprisingly monitorable, and it doesn’t decrease over the course of current RL training.
Real world impacts. The Stargate Project announced. Claude Code shames the rest of the industry into realizing models can code. Can models replicate AI research? What about astrophysics? 3 million PRs merged with AI agents (and this significantly undercounts Claude Code!). GDPval starts the era of optimizing AI models to produce economic value. AI for science is now strikingly real but still banal (where’s my flying car?). Epoch continues a great tradition of making cool graphs (Epoch Capabilities Index; Frontier Data Centers). Calvin on OpenAI culture. Ilya deposition. “The answer to that question will reveal itself. I think there will be lots of possible answers.”
AI timelines. The progress of AI feels both over- and underdetermined. End xAI and a dozen startups will rise from the ashes. At the same time, the number of people pushing the frontier is very small:
- Everything is fractal; even inside AI research orgs, most people are not working on the frontier
- A general property of operations research is that the number of value adding steps is a tiny fraction of the process
- Most people in the economy are “keeping the lights on” which is extremely valuable but not pushing the frontier
Learning. Reading about operations research makes me see everything as a production line (including the art of making the production line). As always I learn the most from articles written in the distill.pub format. Ben and Mindy build a house. The hardest working font in Manhattan, and the analogue for San Francisco. How to pay less when you buy a house (in the Bay Area). Very good todo list for project management.
Travel.
- Waterloo. I was there for the Socratica Symposium (cannot believe this was still 2025). Reflecting, it makes me realize that the place matters much less than the energy of the people who inhabit it. I wish I spent more time hanging out instead of rushing back to work.
- Singapore. Lovely but feels like a giant version of air-conditioned America. I admired their floor cleaning robots, driving fees, and the efficiency of the MRT.
- Vietnam.
- Always travel with friends who have slightly more energy than you and will drag you everywhere, it’s super fun :)
- Embrace last-minute chaos and be willing to go somewhere new, even for only 2 days.
- Taiwan. I was only here for 3 hours, but it is the only place I ever found a free shower in the airport, and that is some crazy technology.
- Massachusetts. I admire the skill of the EA who books a redeye and a hotel room for the previous & following night, so one can check in at 6am and crash until 10. Beyond that, it reminds me that I can be happy in most places. Also, Dunkin Donuts is legendary and I miss rainbow sprinkle munchkins.
Friends. I love conferences because you can talk to everyone you know in SF, but they’re actually free to hang out after dinner. It’s either that or make jam and challah with them. Or drag them to a motorcycle safety course for a full weekend.
Plans for 2026
- Solve logistics — Move to a new apartment (currently planned). Make my personal space joyful, minimal, and low maintenance. Solve logistics for the rest of my life.
- Pay attention — Create at least one recurring time to be fully present with someone.
- Lock in more — I think the ideal schedule is ~6 days working (weekend more chill), 1 day fully off,1 with at least half the day unbooked (to provide the slack time to reflect). Couple this with trips or vacations where I fully don’t think about anything, which is quite different from trips this year.
- Learn more — be able to understand ~every systems and ML-related discussion in the domain of language modelling. Ideally spend a few hours per week on this. Make myself at least a tiny bit better every day.
Rating my 2024 goals
- ❌ Look at my phone less — by sleeping earlier so I have less sleepy mornings.
- Not accomplished :( though setting up my phone as a managed profile, with twitter blocked via DNS, did help a lot
- ⏳ Purposefully spend more time with close friends. Create a list of people I want to stay in touch with or get to know better. Actively try to organize at least 1 recurring event with them (a book club? dinner?).
- Actually did host some events this year! But still much to do here.
- ✅ Take at least one concrete action to prepare financially & physically for AGI.
- ✅ From Charlie Munger: Make myself at least a tiny bit better every day.
- ✅ Have a weekly cleaning hour.
- Empirically this is Sunday night when I do the laundry overnight for Mon (don’t ask when I fold clothes pls)
- ✅ Go to the gym before work.
- Not before, but started going after with a friend :)
- ❌ 3D print at least one puzzle.
- No :( gave away the 3d printer last week
- ✅ Write a 2025 year in review.
- Here we are!
Vibes
(Inspired by Alexey Guzey)
- Riley Walz’s website
- The Unbearable Lightness of Being (read alongside: Details about METR’s evaluation of OpenAI GPT-5.1-Codex-Max)
- An extremely useful property of the universe may be that text is cheaper than video, meaning chatbot tutors, polymaths, and engineers arrived years before video superstimulus (tw)
- There are no secrets to success — just preferential agglomeration and doing well every single day!
- But sometimes there are a few things you should do better
- who watches the Waymos?
- Notes on Cruise’s pedestrian accident
- https://comma.ai/neurips
- Plane Auto-Lands by Itself and Saves Pilots’ Lives! (a robot announcing an emergency over ATC is chilling…)
Many thanks to Miles Wang and Melissa Du for prerelease feedback. Happy holidays!
-
I know people who don’t need this, but I need some time to question my life choices or otherwise I go crazy. ↩