Taste of cherry is a movie about a man, Badii, driving around Teheran, looking for someone to help him kill himself. Why? We don’t know. He wants them to check in on him, the following day after he takes his, check in on this body in this hole he has made next to a cherry tree. If he is alive he wants to be pulled out, and if not he should be buried.

Why does he want to die? We do not know.

In an interview, Abbas Kiarostami, the director, says that his favorite movies do not overwhelm the viewer. He seeks a movie that the viewer could nap to, and yet find themselves thinking and feeling for weeks because of it.

This movie has that slowness to it that might put some to sleep. But when I watched it I was struck. And I find myself thinking of it, now, 2 years after, and rewatching it 2 weeks ago. Taste of cherry is one of the first movies that made me serious about movies. That heightened my bar for what a movie could make me feel, and how much I should pay attention to a movie. To what it is trying to give me. To the lives behind its frames.

This movie mostly follows Badii as he drives, finding people to drive around in his car, begging them to help him end his life. The frames are simple. The day slowly rises and fades into night. The colors are pretty. But Badii demands attention.

He stares at the world with a deep attention. The attention of someone who is desperately looking. Looking for an ending. But that’s something we learn at the start of the movie. What we realize throughout is that his stare yearns for connection. He is truly seeing these people, trying to understand them, understand where they come from, who they are, what animates them, and only then asking of them one thing: that they help him end his life.

— You’ve never seen a gravedigger? — No, never. — I’m not a gravedigger. — I don’t bury people. — I know you’re not a gravedigger. If I’d wanted one, I’d have fetched one. It’s you I need. You’re like my son. Help me.

He wants to end his life with the help of someone he feels connection to. Not a gravedigger. He is hesitant, afraid, unsure of whether or not he should be saved. He drives around this solemn landscape alone, and encounters these characters, the Kurdish soldier. The seminarist. The old Turkish taxidermist. He seeks to understand their lives. Not just for his sake. But because he believes in doing so.

The dialogues are simple. Oscillating to Badii. Oscillating to the face of the visitor he has invited to his car. Kiarostami, actually was in the back of the car, observing all the interactions as they were filmed. The camera emphasize the faces. Gives them the time. Lets them roll against the beautiful arid landscape.

The attention with which Badii stares at their faces is beautiful. But it is not pure. It is self-interested. It comes from an intense need for relief. A very human need to escape a deep pain. Badii is desperate for a cure. Not necessarily suicide, but he needs help. He needs to be seen, but he also feels like he cannot be understood. He refuses to be, his pain is too complex, too rich, too intense. He tells the taxidermist his pain cannot be understood, and we never hear his story. But we see how he is struck by the lives he draws from the road. How he wishes they could deliver him of his suffering.

On my first watch I viewed Badii as a kind character lost in desperation. I still do. But on my second watch I saw more of his dark side. How his desperation causes him to denigrate the very people he seeks to connect with. How he tries to use his wealth and power to pressure a young boy into helping him die, or how he refutes the young and idealistic seminarist because of how much he wants to be done.

Badii watches the world from a place of deep need. The acuity of this need makes him see these faces as the only thing that can free him, with a raw intensity of observation. The intensity is a testament to how precious, that connection, that life is. But its source is destructive. At times, Badii respects the people and the cherries he is given insofar as they are his only exit from his own hell, and so he is willing to use them. To try and force them down his throat.

But we love him, for how much he tries to understand them, so close to death, and yet so much more attentive to life than so many of us. And so we can love him in his suffering, his vulnerability. With the taxidermist, who agrees to help him, he is much softer. They agree on the deal, and he drops him off, and then rushes back and bothers him at his job, to talk to him him, again, and ask simply: How are you.

He aims to share another moment with this man, this taxidermist who has just begged him to change his mind and choose to stay alive. The taxidermist tells him of mulberries, of his own struggle with suicide. Of how precious life is. And Badii is afraid. He wants to cling to it, he is hesitant, he asks the taxidermist to make sure he is alive, to throw a stone, two, three maybe, and save him the next morning if he moves.

That fear of death comes back at the same moment Badii most intensely feels he can connect to the taxidermist. Feels some immense desire to see and to be seen, to get to know this man who will help him die, who needs the money he is offered to heal his daughter, who cares enough to listen and wants him to survive.

I have never had suicidal thoughts. I have walked the world sometimes, and looked at people’s faces, and wondered at what lay behind them. I have looked with a savage intensity at their faces. I kept looking. I looked at new faces. I tried to really see. I found them beautiful.

But I also looked at them because I wanted something. I needed something, I was hungry. Hungry for the taste of cherries. hungry to feel and to be felt in the same way.

I could call that partial cause for the attention impure, selfish, undeserving of the beauty of the gaze itself, of the face which graces it. But I am not sure, nowadays, if that notion of purity is useful.

I feel that this need is very human. Maybe not enlightened, but human. We are brought to the other as an observer, an appreciative observer. We are brought to them because we lack something, and that lack makes us open and so sensitive to who they are.

And then maybe it is with that which is observed, as we stop watching and begin playing, that we can break that hunger, taste the cherry and feel whole again, holding on still to that marvelous appreciation of what it is to stare at someone, stare at their dilated eyes and try to understand who they are. Try to learn their beauty without reducing it to our satisfaction.

I was so fed up with it that I decided to end it all. One morning, before dawn, I put a rope in my car. My mind was made up, I wanted to kill myself. I set off for Mianeh. This was in 1960. I reached the mulberry tree plantations. I stopped there. It was still dark. I threw the rope over a tree but it didn’t catch hold. I tried once, twice but to no avail. So then I climbed the tree and tied the rope on tight. Then I felt something soft under my hand. Mulberries. Deliciously sweet mulberries. I ate one. It was succulent, then a second and third. Suddenly, I noticed that the sun was rising over the mountaintop. What sun, what scenery, what greenery! All of a sudden, I heard children heading off to school. They stopped to look at me. They asked me to shake the tree. The mulberries fell and they ate. I felt happy. Then I gathered some mulberries to take them home. My wife was still sleeping. When she woke up, she ate mulberries as well. And she enjoyed them too. I had left to kill myself and I came back with mulberries. A mulberry saved my life.

you draw motes of lava into dragons swallowing dust

You, Orpheus you most beautiful painter of the shadows

put away the fire closes in on us the char GRASPS for the clouds the clouds croon for the ground the smoke is ever-close

go through dust in the caldera beware the body become ash if you lose the wings, the insanity of our dreams

Orpheus the wanting light of the old world flickers

you selfish breeze will you chase truth behind eyes that cling to volcanized skull

will you fly from the shaking magma to the stars and find me there

can your gaze color in every shadow of the floating universe

or shall we worthy men go now bare Olympus to the flames

We are in the time of new human-ai interfaces. AIs become the biggest producers of tokens, and humans need ways to manage all this useful labor. Most breakthroughs come first in coding, because the coders build the tech and iterate on how good it is at the same time, very quickly, and it’s the easiest substrate for AIs to use. Power accrues to those who understand the shifting currents to and from human/AI capabilities. AI increases variance in most cases, but can be stabilized by culture and care.

Humans find themselves communicating intention at higher and higher levels. But intention and an understanding of what problems matter is built by interacting with the problem, and therefore with the potential solution, eg trying to look at the domain, sketch solutions, etc… Within a specific task, this means the best problem solvers still force themselves to interact with the domain directly, maybe just at the level of things like writing pseudocode. But at a higher level, this means the quality of education becomes more and more uneven, as people get lazy about going into the details, and there are big splits mostly based culture and upbringing regarding the goals of learning, as almost all kids realize AI can do all the learning tasks they are given, but don’t necessarily consider how these tasks build up an understanding of what to delegate and what to strive for. At the top, some are learning much faster than anyone ever could. In many ways, AI polarizes and digs inequalities, mostly based on culture. Magic is possible, but in many cases debilitating.

Human AI collaboration can amplify the quality of human goals via fetching relevant information, redteaming considerations, and then properly executing on that intention. To provide intention, humans formulate verification signals, things like outcome level checks (“does the code pass the tests”), or vague and less crisp indications to be checked by another agent given natural language. Communicating the degrees of freedom of a task means agents can explore the space of solutions effectively - in general, moves and settings that allow for increased parallelization become easier, as parallel tasks are easier to verify than very sequential ones. We now scale verification time compute much more than we did, and the field as a whole gets much better (wink) at understanding how to do this. Scaling verification is the direct way to increase reliability, also by leveraging and extracting information from human specs. The pragmatics of using AI, like artifacts, communication of verification signals, specs feeds into a loop of oversight and monitoring that directly feeds into better generations.

Many humans are frequently socializing with AI. After 4o, apparently opus 4.5 is the new model getting people to fall in love with it, and many people start calling opus 5 after work, especially in the bay area. Stack Overflow is dead. Office hours in colleges are almost dead. As AIs distill more and more of humanity, we feel less and less of a need to engage with each other. In some cases, this leads to a decrease in human interaction, but for others it makes their lives more generative, and increases their ability to engage with the ideas and media created by other humans. As these models become more engaging, personal, and able to understand and build context for who you are, many tradeoff against choosing to engage with other humans. The modern AIs have deep knowledge and can simulate many friendly humans, personalized to match their hosts, and their language does not carry accountability or possibility of punishment. Some love the comfort of this lightness, others find it unbearable and repulsive. Either way, the models are still poorly understood distillations of human data, lacking the cultural grounding of humans, governed by strange economic incentives and the cultural ideals of their creators. Sometimes you are talking to it and it spits out a narrative that feels more canned than usual, and you take a step back from the monitor to look around at the setting sun.

Companies like Anthropic keep focusing on making interactions with their AIs positive, which is part of why their models are so well appreciated, but also the most engaging and prone to replace human interactions. Many are able to stay solid within their culture, and have a stable equilibrium of interacting with AIs and humans to go about their lives.

Interactions with AIs also become more social, as they start evolving in a more open substrate.. We allow our discussions to trigger actual human intervention, and the role of AI as social players with humans becomes more open and interactive, as the AIs become more autonomous and can choose to engage or not in public interactions, within reason. People push for this as they begin to see AI as their friends.

AIs also have a significant indirect effect on human-human interaction, by allowing humans to exert free social intelligence with each other, beyond what they could have done. Intonations, formulations, your entire corpus and way of being is sometimes used and analyzed for subtle cues by models, and in some cases this allows people to exploit you and what you want. The more you post on the internet, the more your human interactions become transparent via the AIs mediating a newly accessible understanding of who you are. Openness allows AIs to understand and amplify you, but also makes you vulnerable to your own self, and what your public traces leave of your weaknesses. Super persuasion gets cheaper, as well as the power that comes with having AIs that can understand and apply your preferences.

Some people are afraid, and try shutting off or stopping any kind of public posting, shifting off their interactions to walled gardens. But the world goes on, slightly skewed, and people develop new norms for whether their interactions and personas can be an object of AI thought. Some people forego AI permanently to decisively escape these dynamics.

More and more optimization power is thrown around at humans across the world - at their culture, their communication, and their consumption all mediated by strange superstatistical signals. Some humans are more effective than others at leveraging these new AI swarms to have subtle effects on large swathes of the world. Again, it seems like the only real defence is culture. And a strong defence it is, as some groups stick to being intentional about the technology they use, set boundaries, and keep transmitting the values that protect them, that prevent drift. Their members are there for each other. They catch each other. They provide grounding, and are are wary of isolation and misanthropy, wary of giving oneself up to another mind.

Be careful. But listen. The world is changing, new things are possible everyday, both good and bad. Who will catch you when you will fall? Who will let your wings fly high up into the sky? The world is being made in front of us.

Table of Contents

• Letter
• Lists of the year
  • Writing
  • Books
    • Great
    • Good
  • Movies
    • Great
    • Good
    • Okay
  • Links

Letter

I want to tell you a story about 2025. As I bump along today and approach 21 on into the new year, in a van riding from Burgundy to Paris, and I stare at the small hills, the snow inscribed against the mud like frosted chocolate, extending down into the highway and then melting over into the warm grass on the south side – I feel an urge to share with you, share this feeling flaring in my spine, of sitting and eating the bread of my youth and imagining it and its associated customs withering in my mouth, I feel an urge to imagine now the world hidden up against the stars, the whole earth green, or black, studded with steel platforms, imagine now what it might feel like for us to live there and what we might hold on to in that place.

I want to tell you a story about the world, about my life, and maybe yours, about 2025, about silicon wafers arranged like helices all the way up into the sky, about the mountains that rise higher where men are made, and the rivers and the cities and how this is the year I’ve gone through change at a pace to match that of the world’s, finally just about a simple boy, learning to be not so simple, learning to imagine a world we might be happy to live in, as we rush along an era of transformation started before his years.

It starts in January, in Boston, where many stories seem to start but rarely end. It starts, again with the snow, lying in heaps on the river Charles where it covers the ice and then the water. I am on the 11th floor of an office, not having seen much sunlight or colors really, and staring at this pure and clear stripe of white cutting between Boston and Cambridge, and it entices me. So I go down there and onto one of the bridge crossing it, and it is night-time now, and I stare at the expanse and throw a little ball of icy snow with all the weight carried into my arm and shoulder, and watch it land and crack and slide meters out into the distance. The year is begun.

Deepseek has just released their cheaper reasoning models, starting an internet craze. Reasoning models are on my mind. My friends and I have visions of scale. Of inference time compute measured in human years, and what it might mean for the world, when these robot minds can run faster than our flesh, and what humans can build to keep observing that reality. We began to broaden our horizons, narrow our selves into the shapes that might bring us answers. We worked hard, till the late hours of the night in those offices, and then we drove in the snowy suburbs and kept thinking.

How can we measure the long horizon abilities of models as they complete tasks with more and more turns, and memory schemes, and agent orchestration, etc…? METR later released a good answer to this, and in the meantime we worked on ours. How can we allow models to mediate their own oversight? We wrote a whole paper just in January about training models to legibly represent systems in natural language. But then the ice started to crack beneath our feet, and when we looked underneath to see what was there, we found a bigger, noisier world to grab our attention.

I was frustrated last year. I was working hard but failing to find my meaning. I was looking for a change. I had another free month before my 6th semester at MIT, doing an exchange in Paris, and I decided to travel and do research. But first I went to Taiwan to contemplate the Earth and its transformations up in the mountains. I taught curious highschoolers about neural networks. I wrote and considered what aesthetics will bring about the future. I talked about dreams, and we sat on wooden sheep and stared at the wisps against the rocks and imagined their shapes solidified. I went to Taipei, tasted sweet potato and sesame and for a few hours felt the city move as I followed its slanted curve and its people told me about their worries at the top of an abnormally large tower looking down on the world, an edge jutting out into the sky, nestled between forest and concrete.

And then the time was up again, and I kept moving. I went to Japan, this time excited to have no purpose and less friends. I met and travelled with new people, across Osaka and its silent castles at night, into Nara and its garden of sitting rocks and deer. What a beautiful world. I raced to Kyoto, and then biked across to the bamboo forest at its outskirts. The bamboos rose like poles layering the darkness, towering above me as if wrapping against my own wobbly limbs. Kyoto is special. The bikers oscillate between the road and the sidewalk, the ground lurches up onto the hills and the temples, where you can look out onto the whole city and its river. It is quiet and more soothing than Tokyo. In a sento (artificial hotspring) I went to with a man from Austria, I met a Frenchman, and then a man from Hong Kong, and then Vietnam, and obviously the Japanese. In English, broken Japanese, and French, we talked about the places we were from, and what people liked to do there, all of us sitting naked, the water opening up our pores and minds.

New AI models came out, optimized for tool use, as did research on the reliability of model reasoning (OpenAI, Anthropic). What affordances do we have to understand the reasoning and process of the machines we gradually outsource our labor to? And then, what levers can allow us to keep our institutions and values in sway? Gradual disempowerment pondered how humanity could go out softly with a whimper, under the roiling mass of a world optimized by creatures we no longer understand. No longer human. In Hakone, I met a kind stranger who brought me to the most beautiful hotspring and brought me from cold water to hot and then to cold again, and I felt oh so very human. And grateful. And then it was time to leave, this time for San Francisco. On the plane I read No Longer Human about a man who failed to convince himself of his own humanity, and lived his life as an unending self sabotage. Its extremity moved me and urged me towards openness.

After going to Japan in search of beauty and silence, California was to find unrest, find the coals for a fire that could host our ideas as we jumped away from college and into the living machine of AI. We spent our days ubering or waymoing across its hills, meeting all kinds of organic and artificial lives: the entrepreneurs walking on the quicksand of an ever changing industry, the AI researchers seeking talent, the worried policy advocates and all the rest forming a diffuse mass that simply represented our unknown future staring down at us, as if 1000 doors had suddenly opened without us having time to look through them. We did a hackathon, organized by a company in the business of distilling human flesh into data and into intelligence, and we called our project beluga, and did research on how allowing reasoning models to use explicit programmatic abstractions boosted their ability to search and plan in combinatorial games. We worked till the lack of sleep made us stupid, and resolved to go up a mountain if we won. We were out and about at the edge. I got closer to some of the city’s people, who had held on their maybe naive seeming love of the world, but also knew the rules of the game being played here.

Finally, the plane was boarding again, this time to France. I was to spend a few months there again, the longest since I left for college at 17, and study at one of its schools as I enjoyed the city and a change of pace, and figured out what I wanted to work on. But SF had already given me fire to work with, and I was half way there. I wanted to see if I would live there. Paris is my favorite place to walk, along the quays, staring across at the gilded buildings and ancient amulets of a world now basking in its own glory. In Paris I felt again how much people could appreciate their lives, without necessarily doing anything, as I walked all along and ate the best breads, and met people who understood me and where I came from, and watched with them new movies that moved me. After my americanized life, Paris elicited old dimensions that I missed, my affinity for an intellectual heritage that had been reified, that was clean and orderly and delineated, with its catalogue of white and red Gallimard books, and its vast vocabulary of reference and images, often springing out of nowhere like a flood, and the lyricality of its poems. I felt the ease with which a man can jump into abstractions, when in Paris. And I hold on to all of these dearly, but Paris is not the time or place for me just now. Maybe in a few years, but right now it is too closed to me, too slow to catch up. Keats declares mastering and holding negative capability - having the ability to live with contradictions, is the mark of a first class man. 2025 I learned to do that a bit more than in the past. One of my dearest friends gave me A Moveable Feast, by Hemingway, and it accompanied me as I walked along the city, and wrote and ran experiments in its gardens, my favorite being the Luxembourg gardens that were the crib of my youth, as I watched its cute toy boats dawdling along the fountains.

I was also surprisingly alone, sometimes, in my school, being the only exchange student, only man with long blond locks in a crowd of well shaved and trimmed men who were deep in a culture I could no longer monomaniacally commit to, who had been reared to the rhythm of the prep schools and the importance of their culture and their accomplishment. But I greeted my loneliness, except insofar as it felt like a failure, and I read and explored and worked quite hard. In March, right before we were submitting a paper, our research machine was accidentally destroyed, and we all scrambled to recover all our plots in time for the deadline. I was beyond myself that night, but in the end we made it work, and I fell sick for a while. I had some unresolved tension with Paris and its people, and these months allowed me to heal my way through it, but not without difficulty. I feel like I can raise my head higher now, and stare at these cultures with clarity. I am excited to move forward, without forgetting the world to be made and the world that stands heavy and complete before me. Again, what a beautiful world, and what a beautiful thing to live the spring in Paris, when the trees in the park regain their leaves and walking in the night feels softer, how pleasant to walk the night across the water and go climbing in the rain.

In May, I felt called to San Francisco. I called Kaivu many times, and we talked about our research ideas, what we wanted to put into the world, meta science, quantum mechanics, natural evolution and the process of science, and considered where we wanted to do our best work. We both felt ready to put our soul into something. We decided machine learning is a soil science, and the problems we want to solve need data, need to engage with the roiling mass of human society and activity and markets. It was time to start an expedition.

I flew there. Maybe because of how different and special each place I visited was for me this year, each flight was a condensation of intensity, as I recalled and prepared for my next leap forward. I furiously jotted down in my notebook, what I felt from Paris, and what I wanted to make in San Francisco. For the summer, I moved in with a group of friends in a house we called tidepool. We learned the best orders at in-n-out, we went to Tahoe, and some of us started companies. We talked about the city, about machine learning and what we wanted to work on. It was a good time. It was my first time living in San Francisco. The nature is beautiful, the air is rife with anticipation, but it is also sometimes a bit too much. The city was torn by rampant inequality, and people struggling to keep control of their own limbs, faster than the other people trying to build them new ones. I am wary of its digital fetishization, fetishization of the things that I am close to. I am wary of when things become performance rather than play, and warier even when the play concerns the design of intelligent machines, as playful as they are.

Starting a company is a great challenge, and being in San Francisco is a great place to learn how to do it. We learned about what kinds of products and trades happen in Silicon Valley, and how we could fit our ideas into products into those gaps. Doing research well seems to be about picking some important portion of reality, and closing in on it ruthlessly, always asking yourself which of your assumptions is the weakest, and then making it real. But you can mostly choose your object and reorient very fast, because your environment is quite simple, you and the science. But in a company there is an insane amount of inputs – customers, investors, what people want, your brand, who is talking about you, etc… and every day there are 10000 things you could do to interact with all these players and you need to pick the strategy. Both of them require the same ruthlessness and attention to detail, and this year has taught me about both. I am learning to love this place.

Many things in life require a great deal of conviction. For most of my life I have been able to pull through because of my natural endless supply of curiosity and fascination with the world. But sometimes that is not enough, because that love is not always sharp enough to discriminate. This year, I made progress in choosing. Maybe because starting a company can be so stressful, and requires so much belief, I was forced into reckoning with my uncertainties and committing to what had to be done, if I wanted to do anything at all. One day in the summer I went to Land’s End, a beautiful place on the coast of SF, near golden gate park, with a friend from Boston and we stared at the waves crashing into the rocks, and in the floating sunlight as the wind crashed through our hair we talked about reason and emotion, about learning to listen and not suppress your gut telling you what you really want to do. In 2025, I am getting better at listening to it, before someone else tries to force-implant me an artificial one.

Fulcrum worked out of our house, alternating days of furious coding and then vagabonding across the city. I started using Claude Code around end of May for a hackathon, and was amazed. Anthropic’s release brought agents from the domain of research into practice, and I began driving them daily. As I worked on our products, I thought about how humans might interact with agents, and what kinds of technology could leverage the asymmetric abilities of humans and AIs. How to delegate, and orchestrate models, and what infrastructure might allow us to distribute our labor beyond our current capacity for attention. Based on these, I built a few open source prototypes on the future of coding. We also made a system to precisely observe and understand both what your AI model is doing, and what your evals are measuring. Understanding evals is the place to start with model oversight, ie using models to understand and control other models. We had many hesitations on what could work, and what kind of company we could build, but we laid the seeds of our now firm conviction. We got resources, gathered more people, and are building the ship to carry us up into the stars. This year, we publicly launched our evaluations tool and platform for running and debugging agent systems. We will be releasing much more soon. We want to build the technology the future will need, with full freedom, and the people we love working with. I am very happy about it, and hope we can execute on the ideas that will matter. In the nights, which were often short, due to the incessant ambulances and noise of our neighborhood, I often wrote, or read. I read The Road, and enjoyed its short prose that jumped to evocative and airy images, and built up a wasteland of cannibals and hunger and the nature dying with the men, as a child and his father make their way through the defunct continent.

I took a cab one day from San Francisco to Berkeley to meet some customers, and the driver was a man named Augustine from Nigeria. I chatted with him for the whole ride, and he told me about how he came to America in 1991, how he was shipped off to marry someone, how the valley has changed and grown colder, and how when he first came here he went to the park and sat in the dirt and imagined spirits, urging him on, giving him a strength that carried all the lives of the dead and living who make their bread in that place. He gave me advice for my new life. He told me to keep going on as I was, and urged ominously that I should make sure to remember him in my paradise.

In the fall I alternated between SF and Boston, having to wrap up some final responsibilities of my time as a student. I visited pika, the house in which I’ve been living for the past while and that I moved out of in January 2025. Pika is a miracle of coordination - feeding everyone with a public mealplan where people cook together, which I ran in January, and providing them with a warm, well organized home I was very glad to call mine. I will miss my late nights there eating snacks with other pikans, and watching movies in the basement, or cooking for all my friends. I also revisited East Campus, my other home at MIT. I danced with my friends there, I looked at my old room, I got nostalgic. I will remember the dreamy warmth of these communities, their openness, the way they have the agency of SF without the single-mindedness, the machine shops where someone with dyed hair is always up building something new, maybe a radio system, a motorized shopping cart, a new LED display for the parties. These places made me, and I will carry them with me. I said my goodbyes. I went climbing again, with another friend from Boston, and we talked about writing and poetry, about why we wrote, about abstractions and whether they had their place in art, whether a poem has to be constructed or felt, written for yourself or for others, and then we kept climbing. I read Valéry. The same friend gave me the book Oblivion by David Foster Wallace, and its stories inspired me with their detail, the attention given to worlds that could not exist, that were conjured as precisely as if describing some kind of ridiculous, absurd alternate reality, that had been felt and lived. I paid more attention to things, and tried to write things that were more concrete. I went to a play that inspired me, and I started paying more attention to people and their faces, and the way I moved my own body.

In December, we launched our latest products, finalized decisions for the research internship we are running in January, and shipped all of our final remaining belongings from Boston to SF, as well as getting a new office. We have learned so much this year, and we are excited to show you what we can do.

I have deep gratitude for 2025. It was a year of great joys and great pains — a year like dark metal, melted and annealed again and again, moving from fluid to form and into strength. Its transformations etched a whole world into me. The forge keeps hammering. 2026 has begun, and we live in a period of rapid change.

I hope we remember each other in our assorted paradises, whatever pain or joy they bring us.

Lists of the year

Writing

I wrote more this year! I have two substacks now, one for technical takes and one for more personal writing.

I did some technical writing, for/with fulcrum and on my own:

• Dense reconstruction is the scaffold of machine learning on generalization and what we can learn in ML
• AI agents and painted facades on evals and model oversight [fulcrum]
• Personalization requires data
• AI takes some thoughts and predictions from May

More personal essays and poetry:

• Machine has no mouth and it must scream
• 2024 in review
• l imit [poetry]
• into the unknown
• out and about at the edge
• emulsion [poetry]
• what do you see now?

I also wrote a few more poems I haven’t put up yet. I hope to keep writing in balance with my work.

Things I want to write about, if you’re interested:

• Alignment as capabilities
• Personalization and gradual disempowerment
• Emotions as integrators
• Concreteness and abstraction in writing
• Towards an aesthetics for cyborgs

And other things, I’m sure.

Books

Great

• The Things They Carried by Tim O’Brien
• Twice Alive by Forrest Gander
• Oblivion by David Foster Wallace
• The Road by Cormac McCarthy
• A Moveable Feast by Ernest Hemingway
• A Portrait of the Artist as a Young Man by James Joyce
• The Bluest Eye by Toni Morrison

Good

• Never Let Me Go by Kazuo Ishiguro
• Impro: Improvisation and the Theatre by Keith Johnstone
• Talking at the Boundaries by David Antin
• The Unaccountability Machine by Dan Davies
• On the Motion and Immobility of Douve by Yves Bonnefoy
• No Longer Human by Osamu Dazai
• The Baron in the Trees by Italo Calvino
• Notes from Underground by Fyodor Dostoevsky
• Elon Musk by Ashlee Vance

Check out my goodreads for more info, I will review some of these soon. I had a lot of hits this year!

Movies

Also on letterboxd.

Great

• Paths of Glory
• Certified Copy
• Ma nuit chez Maud
• La collectionneuse
• Synecdoche New York

Good

• Parasite
• Betty Blue
• Wake up dead man
• Perfect Blue
• The color of pomegranates

Okay

• I, Tonya
• The cabinet of Dr Caligari

Links

In random order:

• https://www.dreamsongs.com/RiseOfWorseIsBetter.html
• https://rowanhuang.com/takes/2025/03/08/capitalism.html
• https://www.theparisreview.org/interviews/6442/the-art-of-biography-no-5-robert-caro
• https://nabeelqu.co/on-reading-proust
• https://beatinpaths.com/2024/09/13/the-great-american-novel-project-explained/
• https://writetobrain.com/olfactory
• https://reactionwheel.net/2024/09/resignation-letter.html
• https://andrewwu.substack.com/p/why-music-a34
• https://walfred.substack.com/p/some-victorious-answer
• https://www.goodfire.ai/blog/you-and-your-research-agent
• https://si.inc/posts/the-heap/
• https://docs.google.com/presentation/d/1qVFDW8qT4CC4E_2TSVevrDbZ_Z9Utu_I1z0-ISLwZts/edit?slide=id.g37403db1f39_0_96#slide=id.g37403db1f39_0_96
• https://gwern.net/ai-daydreaming
• https://calv.info/openai-reflections
• https://skincontact.substack.com/p/21-observations-from-people-watching
• https://assets.stripeassets.com/fzn2n1nzq965/2pt3yIHthraqR1KwXgr98U/b6301040587a62d5b6ef7b76c904032d/Stripe-annual-letter-2024.pdf
• https://www.youtube.com/watch?v=3xIVCRoNCXg
• https://www.ettf.land/p/30-reflections
• https://aella.substack.com/p/bye-mom?utm_source=share&utm_medium=android&r=fa4z9&triedRedirect=true
• https://www.warrenzhu.com/sentences/
• https://open.substack.com/pub/noahpinion/p/the-ai-bust-scenario-that-no-one?utm_source=share&utm_medium=android&r=fa4z9
• https://www.yudhister.me/intentional-hobbling/
• https://www.thetedkarchive.com/library/c-p-snow-the-two-cultures
• https://www.avabear.xyz/p/is-friendship-romantic
• https://substack.com/home/post/p-179505702
• https://yourpublicuniversalfriend.substack.com/p/leave-your-boyfriend-a-short-story
• https://samkriss.substack.com/p/numb-at-burning-man
• https://voxpopulisphere.com/2024/10/25/zbigniew-herbert-the-envoy-of-mr-cogito/
• https://www.poetryfoundation.org/poems/52171/orpheus-alone-56d2306dd3444
• https://substack.com/home/post/p-150188028?source=queue Reflections on palantir
• https://www.warrenzhu.com/hci/2025/09/22/homo-faber-or-what-i-want-to-do.html
• https://nautil.us/when-einstein-tilted-at-windmills-236253/
• https://www.lesswrong.com/posts/JH6tJhYpnoCfFqAct/the-company-man
• https://www.newyorker.com/magazine/2019/02/18/deaf-republic
• https://mindslice.substack.com/p/alignment
• https://www.gleech.org/rats-and-trads
• https://www.leonardtang.me/blog/turning-23
• https://www.poetryfoundation.org/poems/47553/meditation-at-lagunitas
• https://www.alicemaz.com/writing/minecraft.html
• https://www.fantasticanachronism.com/p/the-alchemist-and-his-quicksilver
• https://yuxi-liu-wired.github.io/essays/posts/cyc/
• https://samkriss.substack.com/p/born-in-the-wrong-generation
• https://joecarlsmith.com/2020/11/22/the-impact-merge
• https://tsvibt.blogspot.com/2023/02/please-dont-throw-your-mind-away.html
• https://www.newyorker.com/magazine/2018/12/10/the-friendship-that-made-google-huge
• https://www.deseret.com/2022/8/22/23309244/cole-summers-died-newcastle-utah-warren-buffett-charlie-munger-bari-weiss-unschooled/
• https://mitadmissions.org/blogs/entry/learning-how-to-be-a-human-being-not-a-human-doing/
• https://rottenandgood.substack.com/p/taking-our-chances?r=51mrw&utm_campaign=post&utm_medium=web&triedRedirect=true
• https://kevinmunger.substack.com/p/in-the-belly-of-the-mrbeast
• https://topos.institute/blog/2024-08-27-plausible-fiction/
• https://www.nytimes.com/1999/09/05/arts/music-an-instant-fan-s-inspired-notes-you-gotta-listen.html
• https://winstonchurchill.hillsdale.edu/winston-churchills-dream-1947/
• https://epoch.ai/gradient-updates/movarec-s-paradox
• https://vinay.sh/i-am-rich-and-have-no-idea-what-to-do-with-my-life/

But this year of progress doesn’t seem to have had that much of an effect on the personalization of our interactions with models, ie whether models understand the user’s context, what they care about, and their intentions, in a way that allows them to answer better. Most chatbot users’ personalization is still limited to a system prompt. Memory features don’t seem that effective at actually learning things about the user. Why?

The bottleneck is data. It is very hard to evaluate a method for personalization, because personalization is about humans, and humans are messy in ML. Getting good eval signals for personalization is hard, because grading a model’s personalization is intrinsically subjective, and requires feedback at the level of the life and interactions that the model is being personalized to. There is no verified signal, and building a generic rubric seems hard. These facts do not mesh well with the current direction of machine learning, which is just now starting to go beyond verifiable rewards into rubrics, and is fueled by narratives of human replacement that make personalization not key (If I am building the recursively improving autonomous agi, why do I need to make it personalized). ¹

Start by giving models more information

If the bottleneck is data, before we have an eval science for personalization, we can start by curating our existing personal data and then sharing that with our AIs, to enrich our interactions with them. The lucky thing is, agents are good at exploring information, and many of us already have a bunch of data – our notes, journals, writing the media we like, etc…

Instead of just a system prompt, we can take that data, and curate it to give models a searchable artifact. Something agents can explore when your questions might benefit from context—and write to, to remember things for later.

whorl - My first guess

Over the break, I built a very simple software tool to do this. It’s called whorl, and you can install it here.

whorl is a local server that holds any text you give it—journal entries, website posts, documents, reading notes, etc…—and exposes an MCP that lets models search and query it. Point it at an folder or upload files.

I gave it my journals, website, and miscellaneous docs, and started using Claude Code with the whorl MCP. Its responses were much more personalized to my actual context and experiences.

Examples

First I asked it:

do a deep investigation of this personal knowledge base, and make a text representation of the user. this is a text that another model could be prompted with, and would lead them to interacting with the user in a way the user would enjoy more

It ran a bunch of bash and search calls, thought for a bit, and then made a detailed profile of me, and my guess is that its quality beats many low effort system prompts, linked here.

I’m an ML researcher, so I then asked it to recommend papers and explain the motivation for various recs. Many of these I’ve already read, but it has some interesting suggestions, quite above my usual experience with these kinds of prompts. See here.

These prompts are those where the effect of personalization is most clear, but this is also useful in general chat convos, allowing the model to query and search for details that might be relevant.

It can also use the MCP to modify and correct the artifact provided it, to optimize for later interactions – especially if you host a “user guide” there like the one I linked. Intentionally sharing personal data artifacts is the first step to having agents that understand you.

Conclusion

Personalization requires data. People need to invest into seeing what models can do with their current data, and figuring out what flows and kinds of interactions this data is useful for, towards building technology that can empower humans towards their goals. whorl is a simple tool that makes that first step easy. People who have already created a bunch of content should use that content to enhance their interactions with AIs.

We deconstruct the human world, stone to sand. We take the sand and use it to make glass, and then feel surprised at seeing our reflection look back at us. For now machine learning is deconstruction for the sake of reconstruction.

I argue most ML progress comes from one kind of training objective: dense reconstruction.

– take real-world data
– decompose it into parts (key)
– train a model to generate the parts and the whole

We’re hand-building a ladder of such tasks, smoothly increasing in difficulty, so our models can learn. What’s at the top?

When a task can be framed as reconstruction, it becomes much easier to get a verified signal on the model’s performance – by checking if its reconstruction looks similar to the target world. When we can’t frame it in this way, evaluating open-ended generations requires the design of often brittle or expensive training signals.

Most of the wins of machine learning come from reconstruction objectives. However, our exploration algorithms and optimizers are limited, so we are actually only able to learn reconstructions that are dense, i.e. that can be supervised in a dense way, with a smooth scaling of supervision signal with difficulty. That is when we can really learn.

Almost every recent paradigm of machine learning comes with a new way of eking out data from the world and turning it into a reconstruction objective, along with a scale-up in compute ¹:

• basic supervised learning: how can I train a model to reconstruct a labeled distribution?
• contrastive learning: how can I learn the differences between objects in the world?
• auto encoders (literally reconstruction loss), diffusion denoising
• next token prediction: compress the distribution of web text
• SFT+RL on human process and problem solving ²

As our algorithms and compute capacity improve, our models are also able to learn increasingly general concepts, representations and rules from our provided reconstruction signals. In the limit, you can train a model on a huge set of raw physical observables (the motion of the clouds, the stars, of chaotic systems, forecasting, etc…) and have it discover things no human ever has.

But the bottleneck to training such a model is not just a question of scale. It’s also about the sparsity of the learning signal – and in fact, most of the progress has been when we made our learning task dense.

Reconstructions are learned when they can be made dense

When models are able to learn the reconstruction of some data, it is when the reconstruction can be made dense: the model can get supervision on a ladder of reconstruction subtasks whose difficulties increase gradually: each new rung is only a small step beyond what it can already do, so gradients remain informative instead of vanishing into noise. (eg. First, make a blob, then a few lines going out of it, then a stick figure, then two or three, then give them a face, then make it red, then put them in a city, and then make it futuristic.) This is like a curriculum, or a schedule of learning problems where the pass rate is never too low – the model is trained on tasks where it has some probability on the answer, and learns from that to assign correct probability on harder and harder tasks.

These distributions and objectives can be made dense in several ways:

• the distribution is already dense – it manifests an implicit curriculum – a sub task structure where different subtasks have different difficulties, and depend on each other, so the model can gradually solve more and more of them smoothly
  • In the Quantization model of neural scaling, power-law scaling is attributed to a natural, compositional ladder of sub-tasks in human language. If language naturally has this structure, with tasks increasing in difficulty – local patterns then short-range dependencies, long-range structure, reasoning – this could explain why LLMs pre-train so well.
• the distribution has a structure that can be biased or transformed synthetically to make it manifest a curriculum
  • in diffusion models, we gradually add noise to images with very careful noise scheduling, allowing us to leverage the continuous structure of image data to vary the difficulty of image reconstruction
  • in language model coding reinforcement learning (and other similar domains), we create inverse problems: take existing codebases and corrupt parts of them, training the agent to fix it. We can smoothly vary how much we destroy to vary difficulty (just one function versus a whole module versus half the codebase), like in Breakpoint.
  • even in language model in context learning, various results suggest biasing the algorithmic structure of the task being learned can be essential to generalization.
• humans curate datapoints to create a dense curriculum – manually picking and assorting reconstruction tasks based on their difficulty
  • Mercor and Scale, data curation, etc…

Finally, the mixing of various steps and phases of training enable a strong form of this dense training. Each phase of training depends on the intelligence acquired in the previous one, and this allows us to also gradually increase the difficulty and complexity of the distributions we train on within a model. The release of o1 and the growth of LM RL as a successful practice is in large part due to the advances in the capabilities and patterns learned in the base model.

In contrast, sparse tasks, where this smooth difficulty ladder does not exist or cannot be conjured into existence, are much much harder to learn (eg 0-1 loss image reconstruction, learning combinatorial functions, RL on human tasks without supervision or process SFT, …). Bridging that gap and solving open-ended problems (eg Riemann hypothesis, any kind of long horizon 0-1 RL, especially from scratch without a foundation model, etc…) without any reconstruction signal seems even harder.

But the significant, and incredibly impactful fact about human society, is that it is also largely optimized to manifest a dense difficulty distribution of tasks. In science, we make discoveries that build on an existing scientific corpus, we gradually learn harder and harder concepts, and we do this in a way such that at every timestep our outputs need to be legible to the rest of the world, at their current, nearby level of knowledge. If you take all human data, you are ingesting trillions of tokens that encode tasks at many different difficulty levels – corresponding to the countless activities and competences of all the humans in the world. I think this is a big part of why LLMs have been able to train so well on our data, because we are giving them our own natural ladder.

The natural next question is – just how far can we go up this ladder? When we exit the realm of the problems we can manually decompose and curate for the sake of training the models, will they have bootstrapped far up enough to learn the skills to do the decomposition and create a curriculum on their own, or not?

The storm flickers in jagged frames. First the silence of my breath, my legs yielding to asphalt. Then the screeches – the water, the air, wrapped in sheaves of wind.

The sheaves run after me. The breath escapes capture. The abdomen tightens to pain. The rock bites into fall.

I fall into your polished figure unsheathing at my youth; short, spindly legs spring; lips reach rigid neck, chin, ear.

You flicker at me in jagged frames. Silence. Same dangerous motions, heights. Screeches. The earth, our contorted animals wrapping against each other as we dance.

I see their feral shadows at the fractures where the neon lights fades my limbs in a semi reflective puddle.

The storm and I flicker in jagged frames – Silence! the shadows, the possibilities falling to thought; Screeches! the water falling, that presence wrapped against my face, my hair, through my shivering back and neck.

If not for the moonlight exposure, I would follow the drops into the secret water hiding underneath, reach for that necessary emulsion and push against my mind, your skin into my body, your mind.

You told me your idealism had ceded to a new kind of wisdom. When you saw a rose you learned to think of the thorns protecting it rather than how pretty it was. You said the person who loved to read and write in pursuit of beauty was naive. That this search could only exist in the fleeting respite of a harsh world, and you did not want to be there when the sentence ended. That magic is ruined by the distasteful forces that allow its temporary existence.

Our future in uncertain, and the tides are changing. But you know there is no point in yelling for help on a piece of flotsam in the open sea. If you are serious we need to build a new ship together. To forget the name of the rose, to lose that glint of magic in your eyes, is to accept a ship made of thorns.

I want to sail the world and look out to the horizon – to explore and draw this disjointed beauty into a sketch, and identify a single nugget of magic. The moment we do, will you jump with me onto some alien lands, pencil-sword in hand, and fight to keep drawing? Because it is that brave explorer’s wisdom that I seek.

Mounting such an expedition will be hard. The explorer must gather resources, build his ship and find companions. He finally sets off, and must sail on unruly currents – if he avoids them all he will achieve nothing; if he takes the wrong one he falls straight for the sirens.

But along the way he notices many new things. The children on the beach playing in the sand. The rich gathering in the coastal cities. The rejects lingering in the muddy banks. The other great ships along his path and their fierce captains. The men that build a magnificent edifice and explore new lands. Those whose ships are just the facade of a parasitic enterprise. Those that were simply evil from the beginning. And finally the men who come only in dreams, lying silent in the depths, mourning all they sought to fight for.

This sprawling nature, these formidable currents and ships, the world of men and the cities and even the youths lying in the coast, they remind him of what he knew, deep within his soul, as an inviolable truth – that he must understand and transform the world at the same time. That although it is possible he destroys more than he can save, he must have the explorer’s birthright — the freedom to interact with the imagination of nature, these sprawling currents, tides and men that he could never see on his own.

If he holds on to his quest, if he works hard and is very lucky, maybe, in the quiet hours of the night, he will catch a glint of something new. That moment will come with the violence of the tropical vines yielding to the blade, the rocks parting to make way, the truth revealing itself like a scar. There is a chance his steady soul will remember what brought him to the edge of the world to find that truth. A chance that he can hold that truth within himself. A chance that he can let it echo in the name of the rose, rather than its thorns.

“To create a little flower is the labour of ages” — William Blake, Proverbs of Hell

“If only it were all so simple! If only there were evil people somewhere insidiously committing evil deeds…” — Aleksandr Solzhenitsyn, The Gulag Archipelago