the dream is materializing
as some of you know, the namesake of this newsletter is the dream machine, m. mitchell waldrop’s study of computing pioneer j.c.r. licklider. in 1960, licklider published “man-computer symbiosis,” a paper outlining a vision of computing as a tightly coupled cooperation between humans and machines, enabling humans to think and work at much higher levels of complexity than before.
by some measures, that vision came to fruition by the start of the twenty-first century. the graphical user interface, the spreadsheet, the internet, the search engine, and many other inventions of the past several decades gave humans the tools to manipulate information much more quickly and accurately, solving new kinds of problems at an unimaginable scale.
but the human-machine cooperation that licklider theorized, where humans produced ideas to execute and questions to answer and machines carried them out in real time, remained elusive, trapped in specialized contexts. why? software was expensive and inflexible. you had to anticipate particular needs and painstakingly write out instructions for a program to meet each one. and because reality has a surprising amount of detail, there were virtually infinite needs and a finite supply of people to craft programs to address them, leaving the symbiotic loop licklider described perpetually out of reach.
licklider pointed to this problem as one of the primary barriers to his vision:
The basic dissimilarity between human languages and computer languages may be the most serious obstacle to true symbiosis. It is reassuring, however, to note what great strides have already been made, through interpretive programs and particularly through assembly or compiling programs such as FORTRAN, to adapt computers to human language forms... For the purposes of real-time cooperation between men and computers, it will be necessary, however, to make use of an additional and rather different principle of communication and control. The idea may be highlighted by comparing instructions ordinarily addressed to intelligent human beings with instructions ordinarily used with computers. The latter specify precisely the individual steps to take and the sequence in which to take them. The former present or imply something about incentive or motivation, and they supply a criterion by which the human executor of the instructions will know when he has accomplished his task. In short: instructions directed to computers specify courses; instructions-directed to human beings specify goals.
last year llms crossed the threshold that fundamentally changes this equation. software—humanity’s most powerful mechanism for interacting with knowledge at scale—can now be generated at will. we’re rapidly approaching a world where software directed at any goal is virtually free and instant to create for anyone.
even if that were somehow the only use case for llms, it would change the world as we know it. but it’s not. llms are capable of much more than licklider believed necessary for man-computer symbiosis; they can make decisions, they can define goals, they can devise processes.
as ivan zhao wrote recently, we now have infinite minds. the implications are profound, and we’re only beginning to see them unfold.
i started more seriously using llms in october 2024. cursor was generating buzz among software engineers; it was, at least at the start, a much smarter autocomplete that could finish your thought with a reasonable degree of correctness. i used it on a side project; i was surprised it worked even that well.
by december, i was using llms to teach myself new subjects, a method i found remarkably useful. i’ve often struggled with structured regimens for learning—they’re not quite at the right altitude, or they make assumptions about preexisting knowledge. i had always learned best by doing, which allowed me to press against the edges of my understanding and build durable mental models. llms allowed me to zoom in and out of new information, situate it in a broader context from any angle i wished, test myself through questioning.
early 2025 was a period of rapidfire acceleration. it felt like every week there was a model release that changed what was possible to do with llms. each successive version was obviously, noticeably, smarter and more capable, with a distinct personality and set of skills. i would throw different kinds of problems at them and watch them fail miserably, often in wildly funny ways—then, later, see them handle them with ease.
in march i started paying for chatgpt pro. prior to that, i would have found it hard to imagine spending $200 a month on a single software subscription for personal use. since then, it’s been impossible to imagine not having it (i’m now primarily using claude max with chatgpt plus as a backup, though there were a few months of total abandon where i had both claude max and chatgpt pro).
it wasn’t long before i was convinced that writing code was an effectively solved problem. now that feels obvious, but at the time it wasn’t at all clear that coding agents would be able to navigate production-scale codebases well enough to take on the task of implementation entirely. but llms proved to be spectacularly resilient; often the more you delegated to them, the more useful they were able to be.
that summer at zora we changed how we build products to use llms at every level. code was quickly becoming an implementation detail. claude code, radical at the time of its launch for its choice to offer only a conversational interface and no code editor, started to feel like the logical conclusion.
by september i was using llms as part of nearly every kind of thinking. i had used them to translate books that only existed in chinese, to think through strategic decisions for work, to brainstorm puzzles for the murder mystery i was writing, to assemble research reports from hundreds of sources, to build software projects overnight, to explore and understand myself better, to speed up my editing process for my writing. voice dictation became a primary mode of input, making it easier to just talk out loud, often while i was walking somewhere, and then have the model structure the thoughts for me.
in my travels through china, i leaned heavily on llms to navigate a country unfamiliar to me, where my fluency was limited. claude and chatgpt helped me choose neighborhoods for my stay, understand subtle cultural and linguistic nuances, read restaurant menus so i could order, figure out modes of transportation, debug technical issues between my devices and chinese apps and services.
i had always been big on keeping records, but fast and free transcription and llms capable of agentic search meant that suddenly collecting as much data in my life as possible wasn’t a speculative exercise; it gave me an externalized memory, imperfect but powerful. i started recording more and more, all the time.
this past christmas i spent the week training my own model from scratch, on a dataset i had built of a hundred million tokens of books i had read, articles and papers and essays i had saved, my kindle highlights, etc., as well as my own writing. the model it produced was interesting and strange but not particularly smart, so i decided instead to fine-tune qwen, a chinese open-weights model, just training it on my macbook with mlx (apple’s ml framework for apple silicon). that was much better, albeit trickier and more sensitive than i had expected. but the core of it was: i could use an llm to train another llm from scratch, with zero prior experience. that would have been unimaginable a few years ago, and it was so easy.
all of that i say because it’s easy to grow acclimated to a new normal and lose perspective of a rapid rate of change. but that’s my experience with llms the past year or so; hundreds of millions of people around the world are having their own. we’ve entered the age of intelligent machines, and imperceptibly, then undeniably, it’s remaking the world around us.
the future is blurrier than it’s been in many decades, if not longer. few assumptions remain stable ground. but some beliefs have been solidifying for me.
agi for bits (information) isn’t here just yet, but no further advances in models need to happen to get there. the scaffolding around them (agent architecture, context management, etc.) will be enough, and we’ll probably be there before the summer.
agi for atoms (the physical world) needs to be solved for most people to feel the more apparent societal changes, like dramatically lowering costs of living. by 2030, that technology will exist on a capability level (i.e., it will be possible but expensive or unviable to produce at scale), and the next decade will be about getting it into the real world.
these milestones are necessary but insufficient for a post-scarcity future. that requires policy to socialize the benefits of these technological advances, and there’s no sign that’s on the horizon. this problem is exceedingly important and vastly underresourced compared to the first two, for obvious reasons.
coordination between humans remains valuable because humans can be qualitatively much smarter together; agents, on the other hand, are (mostly) just quantitatively smarter together. why? humans are able to reach a remarkably high level of intelligence with a spiky distribution of knowledge, which produces unique perspectives and, thus, creativity. llms need a massive amount of knowledge to reach comparable levels of intelligence, which means they struggle to escape the mean without human direction. but this reality only holds true if there aren’t significant architectural advances that change these constraints.
what happens in tech the next year offers a blueprint for what happens everywhere else. software is the first solved problem of this kind, and the effects are likely to mirror dynamics already visible in china: intensifying competition that keeps shifting the baseline for bringing anything to market, creating conditions that are excellent for everyone as a consumer, brutal for everyone as a producer.
the divide between people who use llms to expand their own capacity for learning and thinking and those who use them to outsource it will keep widening. execution is approaching a cost of zero; the work that’s left is making the right decisions, which requires absorbing a lot of information quickly and exercising good judgment based on it.
local models will become much better and more widespread—not primarily because of privacy, though that’ll be a positive externality, but because they’re free and reliable. most apps leveraging llms are bleeding costs to gain distribution right now; at some point that subsidization will stop. cloud model reliability remains notoriously bad compared to other online services because of compute needs and demand. right now, the only local model i use regularly is nvidia’s parakeet tdt model for speech recognition, but as models improve, running your own on-device will become a competitive option.
verification methods for online content (like proof of humanity, content authenticity, etc.) will become much more important. right now we’re occupying a strange liminal space where many people suspect that most internet activity is no longer performed by real people, and yet for most intents and purposes, we’re still acting like online content is by default true. “dead internet theory” has thus far been overstated because virality on the internet follows power laws and most bot-generated content is boring, but that’s changing. even now viral chatgpt posts are becoming more common on twitter and substack and reddit, and tiktok and instagram reels have started to see wholly ai-generated videos like beauty/fitness transformations take off. at some point, the sheer volume will make verification necessary for any context where humans need to know they’re talking to other humans.
within a decade, ai will produce a breakthrough cure for a major disease that changes the prognosis from terminal to treatable for millions of people. ai is already making many scientific problems newly tractable, and that work will accelerate. medicine has the clearest near-term impact; deepmind’s alphafold has already transformed the domain of structural biology in the last five years. many other institutions like arc institute are pushing the frontiers of biomedical science with research like cellular simulation to predict responses to novel drugs.
much of this change is destabilizing, anxiety-inducing. as always, technology amplifies existing problems as much as it opens new possibilities; there are no guarantees that any of this “progress” will go well. but i think it’s also hugely, wildly exciting. we’re shaping the pivotal moment in human history, the proliferation of human-made intelligences that can be designed for any purpose and scaled to any level. the cooperation that licklider envisioned sixty-five years ago, computers as intelligent partners for the human mind, is on the cusp of reality. the dream machine is coming alive.
this essay is part of a series, “what comes next.”
some news: for anyone who’d like to read from me more frequently, i’ve started a new newsletter, scratchpad, where i’ll be posting dispatches much more often and experimenting with form and substance. the dream machine remains a home for occasional fully-formed essays about technology, culture, and society; scratchpad will contain much earlier explorations of ideas and personal experiences.
it’s an experiment, and i don’t know how long i’ll keep it open to the public, so if you’d like to subscribe, now is a good time. the first series of posts covers my last few months of travel through china and japan.
as always, responses are my single favorite part about sharing to this newsletter, so if anything sparks a thought for you, i would love to hear it.