entering my ROCK ERA

on billie eilish, alchemy, language models

i. annual recap

i made end-of-year websites in 2023 and 2024 but this year i decided i’d rather start on upcoming projects so you’re getting this written summary instead. here’s a brief work-only recap of 2025:

1. i spent jan-march stuck in emotional paralysis, meeting lots of cool faculty and being indecisive about whether to go to grad school. also i was doing research part-time (including pre-deployment testing of o3 that went viral) and writing mediocre blog posts

2. in april-july i realized i was tired of who i was and that i’d trapped myself in a specific genre of writing that was far too limiting. so i tried a lot of new things, like drawing a comic about declining grad school (which i realized would be fun socially but would probably result in stagnation / feeling trapped, mostly due to the current state of nlp), building a ddr pad, writing a short story that many people told me they still think about half a year later, wild. relatedly i met some new art friends like J and T, who were very good influences and helped me realize that i could / should branch out into new mediums

3. in aug-oct i did self-directed research for the first time (ie. in all my previous research a lot of the good ideas came from other people, whereas this time they almost all came from me), traveled a bit, and tried some more writing experiments. also there was a stroke which made me re-evaluate my priorities

4. ai took over my life in nov-dec as i went through 2 research releases and also NeurIPS occurred. the deadlines destroyed my soul in that i had so many ideas i wanted to explore but they all got backlogged to focus on finishing projects and my personality became extremely one-dimensional, sorry if you had to talk to me during this time

i think 2. and 3. were important phases to go through. trying new kinds of projects and seeing them turn out well was helpful both for confidence and for becoming more accepting of new ideas (i’d observed in may that blogging too much biases you towards only thinking about ideas that fit in blog posts); i was able to connect with a lot of new people who liked stuff i worked on this year but would not have found my pre-2025 work interesting; there were also just a lot of cool moments, eg. one time i was talking to my friend alicia and she said something like “i love how much you’ve been experimenting recently, it’s inspiring” and i wanted to scream: no?? literally YOU are MY inspiration i love all of YOUR experiments what are you talking about?!?

---

ii. ai research

NeurIPS did not go the way i expected it to. twice a day i would walk through massive poster sessions, each exhibiting hundreds of machine learning papers, and i would be bored out of my mind; then i would walk through the 15-20 creative ai demos and consistently come away with new human-ai-interaction ideas to think about. after this occurred for the fourth consecutive session i thought hm, maybe this is a sign that i shouldn’t be doing machine learning

my current research is on language model interp. to clarify a common misconception i do not work on mechanistic interp (which tries to reverse-engineer the exact mechanisms used by models, and which i believe is impractical); instead i mostly use black-box methods (eg. training an assistant to do the interpreting). many people who work on interp are motivated by concerns about existential risk, so people often ask me why i (as someone generally unconcerned with existential risk) am doing interp. for the longest time i haven’t been able to give an answer beyond i think it’s cool to be able to understand language models better, and when people ask why i say something like i enjoy discovering new things about complex systems

which is definitely a bullshit answer, because i wouldn’t want to spend my life discovering new things about eg. zebra colonies. the discovered knowledge must have some use case that excites me, and if the use case for interp isn’t monitoring for safety / risk then what is it? i’m not entirely sure, but i think the interp applications i find most interesting might be spiritually closer to hci - how would we change the way we interact with language models if we could probe and steer their internal thoughts / hidden associations / conceptions of themselves while chatting with them? i will try to push my team in this direction, and if it doesn’t work out i think i should switch fields because i am quite tired of being far from applications

in november my housemates asked me to list examples of other peoples’ projects that i would’ve been happy working on, and i struggled to do so. probably most of my favorite books / art / media? the pinterest platform? the general user models paper? one of the common themes here is that i want to create things that help people better understand or express themselves

i think there are not many companies i would thrive at, eg. places like anthropic and cursor build great products that (in my view) do enable self-understanding and self-expression, and i could’ve joined either org but i find it hard to care when many other orgs are building the same thing. unfortunately everything with a large market incentive will have many copycats, so maybe i’ll never be wildly rich… i am still coming to terms with the fact that the things people are most eager to give me money for are not really the things i want to do

---

iii. reinvention

there’s a george bernard shaw quote i like that goes: “This is the true joy in life, being used for a purpose recognized by yourself as a mighty one. Being a force of nature instead of a feverish, selfish little clod of ailments and grievances, complaining that the world will not devote itself to making you happy.”

it’s a little harsh, but i think a feverish, selfish little clod of ailments and grievances, complaining that the world will not devote itself to making you happy accurately describes a lot of internet content, including much of my writing when i was depressed. i don’t mean to suggest this genre of writing is bad - it’s often relatable, beautiful, and genuinely valuable - but producing it feels like being trapped in a cycle of “have feelings -> write -> catharsis -> have more feelings” where you’re incentivized to keep having similar feelings so that you continue having easy material to write about, and i’ve been trying to instead move towards being used for a purpose recognized by yourself as a mighty one. (this is related to the observation that enneagram 4s grow by becoming more principled)

when i think about artistic reinvention one figure that often comes to mind is billie eilish. partially because she’s a year younger than me and watching her evolve in real time has been inspiring, partially because the journey she’s been on feels similar to what i’m looking to do. in particular: when we all fall asleep perfected a specific formula (bedroom pop, breathy vocals recorded close to the mic, minimalist bass-heavy production, mostly writing about teens being sad in various forms), but the commitment to the formula was often done out of fear (fear of belting and imperfection, fear of production that was too noticeable, fear of discussing topics where her opinion could be challenged; here’s a breakdown if you’re curious); happier than ever exhibited more diverse writing and vocal range (especially on the title track) and elements of rock / jazz; hit me hard and soft is still very obviously billie but large portions of it sound nothing like her first album with how much rock / jazz / edm / retro / other genres have been integrated

how do you go about completely altering your style, content, and relationship with your work, while keeping intact the emotional core of your output? it’s time for me to make the analogous transition from bedroom pop to rock. this year was a decent year for trying new things, but i didn’t explore nearly enough; much of my time was consumed by various self-created distractions and realistically i should be able to do 2-3x more in 2026. just gotta iterate faster, love harder, put more ideas out, fail earlier, make deeper contact with reality, train the idea muscles, compound the exponentials

this is the first time in at least four years where i’ve had resolutions entering the new year. in past years i was too confused to know what to do, and additionally i didn’t realize the antidote to confusion is rapid exploration rather than thinking hard and waiting for the confusion to resolve. i guess subconsciously i was also expecting the right people to appear and tell me what to do and save my life. those people actually did show up in 2023 and 2024 and 2025; and they were remarkably kind and insightful and helpful; and i’ll always be grateful to them for shaping me into who i am today; and they did tell me what to do with my life; but this year i realized everything they told me ended up being incorrect anyway, so i guess now i finally understand: you can’t wait for someone else to change your life, even if you’re surrounded by the best people ever. it doesn’t matter how good your taste in mentors is because even the most incredible person does not have enough information to know what your future should look like. your life is beyond the reach of prophetic instruction

i finally read the alchemist last week. most of my friends who have read the book say it’s overrated but i kind of disagree? i think it’s really interesting that the popular conception of alchemy is about doing bullshit pseudoscience to turn ordinary metals into gold and then the alchemist instead argues that alchemy is all about self-transformation and once you’ve made enough progress on yourself life will feel as if ordinary metals were gold. that to me is a very beautiful perspective. anyway here are a few relevant passages i liked:

“To realize one’s Personal Legend is a person’s only real obligation.”

“You must understand that love never keeps a man from pursuing his Personal Legend. If he abandons that pursuit, it’s because it wasn’t true love.”

“That’s what alchemists do. They show us that, when we strive to become better than we are, everything around us becomes better, too.”

---

iv. a few notes

on the theme of things that help people better understand or express themselves, here are some initial areas i am exploring (reach out if you want to chat or collaborate or do coworking / accountability!), and we’ll see where things go from there:

writing tools. language models have been decent at everyday writing for a while, but i think they’re finally good enough to be useful for the specific kinds of writing i care about, which look more like vibe curation. for instance, it should be possible to sample and remix text as easily as musicians sample and remix audio. i’ve been trying this out with poetry over the holidays, here’s a prototype that should be done soon:

optimistic stories about technology. many of my friends work at agi labs but none of them have actually been able to outline a scenario where agi occurs and the world is in a good state. there are absurdly few compelling visions of techno-optimist near-futures! i can think of solarpunk from the chobani commercial, the social-first app from better living through algorithms, the serendipity maximizer from all the birds in the sky, and that’s basically it. most techno-optimist narratives (eg. “solve science”) skip the near future and degenerate into high sci-fi / fantasy, or invoke concepts like “ubi” that don’t actually make any sense when you examine them critically. there is a clear vacuum here, and we should probably change that since narratives about technology are often self-fulfilling. right now we have no idea what to build towards other than that it should be intelligent

i don’t know what stories to tell yet, but as a starting point i really like this tweet from humans&, which is such a refreshing break from modern agi lab rhetoric about scaling / reasoning:

we think humanity’s biggest challenges won’t be solved by AI thinking for 1000 hours and coming back with an answer.
they’ll be solved by many collaborating humans, and AI that understands them and their different skills, goals, values, etc to empower them to do more together

feelings interventions. emotions are complex but when i look back on the last few years i find that something like 90% of how i feel is determined by two axes: hope / despair, conviction / uncertainty. a digital assistant or tamagotchi that just tracks these two axes and occasionally asks simple probing questions like: what would you do right now if you were less pessimistic? are you procrastinating out of despair or uncertainty? what are you currently confused about and how can you explore it? etc. would go a long way. also i wrote about friendship batteries two years ago and never figured out how to build them, i should do that

optimistic stories about people. this is pretty self-explanatory? uplifting narratives about other people. this can be nonfiction, like the work my friends donald and aadil are doing to capture our peers’ stories, or fiction, where i have some ongoing drafts

normal stuff. after writing the above list i reread it and thought why is it that i can only come up with useless ideas? i find it very suspicious that there is no traditional tech here (eg. “cursor for X”, “gadget for Y”); for some reason those ideas have been evading me. it might be an attention issue where if an idea is a) not specific to a niche interest i’ve thought a lot about b) obviously useful and feasible to build, then it’s likely to already exist? but also this could be an imagination issue. i can’t tell if lack of traditional tech is a feature or a bug, let me know if you have any thoughts

---

unbelievably excited for 2026! thanks to AA AG AG AY CW DJ DY EZ GZ JS JS JX SZ TW for relevant discussions