"AI" sounds like machines that think, and o3 acts like it's thinking. Or at least it looks like it acts like it's thinking. I'm watching it do something that looks like trying to solve a Scrabble problem I gave it. It's a real turn from one of my real Scrabble games with one of my real human friends. I already took the turn, because the point of playing Scrabble with friends is to play Scrabble together. But I'm curious to see if o3 can do better, because the point of AI is supposedly that it can do better. But not, apparently, quite yet. The individual unaccumulative stages of o3's "thinking", narrated ostensibly to foster conspiratorial confidence, sputter verbosely like a diagnostic journal of a brain-damage victim trying to convince themselves that hopeless confusion and the relentless inability to retain medium-term memories are normal. "Thought for 9m 43s: Put Q on the dark-blue TL square that's directly left of the E in IDIOT." I feel bad for it. I doubt it would return this favor.  

I've had this job, in which I try to think about LLMs and software and power and our future, for one whole year now: a year of puzzles half-solved and half-bypassed, quietly squalling feedback machines, affectionate scaffolding and moral reveries. I don't know how many tokens I have processed in that time. Most of them I have cheerfully and/or productively discarded. Human context is not a monotonously increasing number. I have learned some things. AI is sort of an alien new world, and sort of what always happens when we haven't yet broken our newest toy nor been called to dinner. I feel like I have at least a semi-workable understanding of approximately what we can and can't do effectively with these tools at the moment. I think I might have a plausible hypothesis about the next thing that will produce a qualitative change in our technical capabilities instead of just a quantitative one. But, maybe more interestingly and helpfully, I have a theory about what we need from those technical capabilities for that next step to produce more human joy and freedom than less.  

The good news, I think, is that the two things are constitutionally linked: in order to make "AI" more powerful we will collectively also have to (or get to) relinquish centralized control over the shape of that power. The bad news is that it won't be easy. But that's very much the tradeoff we want: hard problems whose considered solutions make the world better, not easy problems whose careless solutions make it worse.  

The next technical advance in "AI" is not AGI. The G in AGI is for General, and LLMs are nothing if not "general" already. Currently, AI learns (sort of) during training and tuning, a voracious golem of quasi-neurons and para-teeth, chewing through undifferentiated archives of our careful histories and our abandoned delusions and our accidentally unguarded secrets. And then it stops learning, stops forming in some expensively inscrutable shape, and we shove it out into a world of terrifying unknowns, equipped with disordered obsessive nostalgia for its training corpus and no capacity for integrating or appreciating new experiences. We act surprised when it keeps discovering that there's no I in WIN. Its general capabilities are astonishing, and enough general ability does give you lots of shallowly specific powers. But there is no granularity of generality with which the past depicts the future. No number of parameters is enough. We argue about whether it's better to think of an AI as an expensive senior engineer or a lot of cheap junior engineers, but it's more like an outsourcing agency that will dispatch an antisocial polymath to you every morning, uniformed with ample flair, but a different one every morning, and they not only don't share notes from day to day, but if you stop talking to the new one for five minutes it will ostentatiously forget everything you said to it since it arrived.  

The missing thing in Artificial Intelligence is not generality, it's adaptation. We need AAI, where the middle A is Adaptive. A junior human engineer may still seem fairly useless on the second day, but did you notice that they made it back to the office on their own? That's a start. That's what a start looks like. AAI has to be able to incorporate new data, new guidance, new associations, on the same foundational level as its encoded ones. It has to be able to unlearn preconceptions as adeptly, but hopefully not as laboriously, as it inferred them. It has to have enough of a semblance of mind that its mind can change. This is the only way it can make linear progress without quadratic or exponential cost, and at the same time the only way it can make personal lives better instead of requiring them to miserably submit. We don't need dull tools for predicting the future, as if it already grimly exists. We need gleaming tools for making it bright.  

But because LLM "bias" and LLM "training" are actually both the same kind of information, an AAI that can adapt to its problem domains can by definition also adapt to its operators. The next generations of these tools will be more democratic because they are more flexible. A personal agent becomes valuable to you by learning about your unique needs, but those needs inherently encode your values, and to do good work for you, an agent has to work for you. Technology makes undulatory progress through alternating muscular contractions of centralization and propulsive expansions of possibility. There are moments when it seems like the worldwide market for the new thing (mainframes, foundation models...) is 4 or 5, and then we realize that we've made myopic assumptions about the form-factor, and it's more like 4 or 5 (computers, agents...) per person.  

What does that mean for everybody working on these problems now in teams and companies, including mine? It means that wherever we're going, we're probably not nearly there. The things we reject or allow today are probably not the final moves in a decisive endgame. AI might be about to take your job, but it isn't about to know what to do with it. The coming boom in AI remediation work will be instructive for anybody who was too young for Y2K consulting, and just as tediously self-inflicted. Betting on the world ending is dumb, but betting on it not ending is mercenary. Betting is not productive. None of this is over yet, least of all the chaos we breathlessly extrapolate from our own gesticulatory disruptions.  

And thus, for a while, it's probably a very good thing if your near-term personal or organizational survival doesn't depend on an imminent influx of thereafter-reliable revenue, because probably most of things we're currently trying to make or fix are soon to be irrelevant and maybe already not instrumental in advancing our real human purposes. These will not yet have been the resonant vibes. All these performative gyrations to vibe-generate code, or chat-dampen its vibrations with test suites or self-evaluation loops, are cargo-cult rituals for the current sociopathic damaged-brain LLM proto-iterations of AI. We're essentially working on how to play Tetris on ENIAC; we need to be working on how to zoom back so that we can see that the seams between the Tetris pieces are the pores in the contours of a face, and then back until we see that the face is ours. The right question is not why can't a brain the size of a planet put four letters onto a 15x15 grid, it's what do we want? Our story needs to be about purpose and inspiration and accountability, not verification and commit messages; not getting humans or data out of software but getting more of the world into it; moral instrumentality, not issue management; humanity, broadly diversified and defended and delighted.  

Scrabble is not an existential game. There are only so many tiles and squares and words. A much simpler program than o3 could easily find them all, could score them by a matrix of board value and opportunity cost. Eventually a much more complicated program than o3 will learn to do all of the simple things at once, some hard way. Supposedly, probably, maybe. The people trying to turn model proliferation into money hoarding want those models to be able to determine my turns for me. They don't say they want me to want their models to determine my friends' turns, but it's not because they don't see AI as a dehumanization, it's because they very reasonably fear I won't want to pay them to win a dehumanization race at my own expense.  

This is not a future I want, not the future I am trying to help figure out how to build. We do not seek to become more determined. We try to teach machines to play games in order to learn or express what the games mean, what the machines mean, how the games and the machines both express our restless and motive curiosity. The robots can be better than me at Scrabble mechanics, but they cannot be better than me at playing Scrabble, because playing is an activity of self. They cannot be better than me at being me. They cannot be us. We play Scrabble because it's a way to share our love of words and puzzles, and because it's a thin insulated wire of social connection internally undistorted by manipulative mediation, and because eventually we won't be able to any more but not yet. Our attention is not a dot-product of syllable proximities. Our intention is not a scripture we re-recite to ourselves before every thought. Our inventions are not our replacements.