Four of us and a guest, because one member asked whether anyone would mind if he brought a celebrity guest and then introduced his mother. She fell asleep beside him before the hour was out, which he found funnier than anybody.
The founder opened from his terrace, on the first evening cool enough to sit outside after weeks of Italian heat, rain finally falling and the dog at his feet. He had been reading The Infinity Machine, the biography of Demis Hassabis, and wanted first the thing the book is mostly about. Working out how a single protein folds into the shape that gives it its function was once five or ten years of a career, an entire doctorate spent on one molecule, and then it became hundreds of millions of proteins in a few months. Every molecular biologist alive had their working life redefined between one moment and the next. Whether they had just finished that decade of work or were about to begin it, the science they thought they were entering no longer existed. When people ask what a technological singularity would feel like from the inside, it already happened to those people, and it happened in public.
The passage he had marked is Hassabis answering his biographer. Language models trained on everything humans have ever written turned out to hold a representation of the world that was powerful and useful in a way he had not expected, and the biographer asks whether that means we are less original than we thought. Maybe there is just less variety, Hassabis says, and reaches for the proverb: there is nothing new under the sun. The founder smiled, because that proverb is the thing his own book was written against. Something New? carries the subtitle AIs and Us: The Coming Age When There Is Nothing New Under the Sun Is Not True Anymore, and he suspects Hassabis catches the fallacy somewhere in the pages he has not yet reached, since the man is still at work building the very thing that disproves it.
His position is that the proverb was never true. New things have always appeared in the universe, and what was missing almost every time was anybody present to recognise the phase transformation that produced them. What makes this moment unusual in the history of the universe is not the novelty; it is that the novelty has witnesses. He treats that recognition as a responsibility as much as a privilege, and the responsibility is to hold the position out loud, so that skepticism argued from the high ground of a limited experience does not stop the people who can see the change from acting on what they see.
Last week’s argument returned here with a harder edge. Anyone exploring something new meets the obstacles that come with the exploration, and then meets a second set that comes from their own community, from the majority who prefer the convenience of experience already acquired. When those people counsel you to be modest, to be meek, to stay inside the normal distribution of ambitions, they do themselves the disservice of never finding out what they might have done. The larger cost falls on everyone else, because that counsel drains the energy of the people who were ready to try, and the energy is the scarce thing.
Which is why the founder thinks the meetings are only half of it. The other half is telling people who are not in them: the member’s mother, hearing about all of this firsthand for the first time; his own mother; the colleague who threw an arithmetic puzzle at a model two years ago, watched it fail, and settled comfortably into the conclusion that there is nothing here. That person is at peace with the negative result and will not try again. Enthusiasm, said out loud by someone who is finding things, is most of what reaches them.
A member answered with an image he has been turning over for a while. He had watched a child playing with a balloon animal and started wondering why the twisted dog is more functional, more embodied, than the balloon it was made from. What the clown does is pinch the balloon, form a ring and twist, and the whole thing holds because the material is flexible enough to take the twist. Bridging to people is the same operation. You cannot be fixed in your ways; you have to see both sides and find the twist that brings them together, which is why the communicators who manage it tend to work through irony and humour, offering grace where there was none.
Then the demo promised a week ago. The machine is a DGX Spark, which Nvidia calls a desktop supercomputer: 128 gigabytes of unified memory and four terabytes of SSD. The founder believes he bought the last one available in the Bay Area, from a computer shop he had never heard of in the San Francisco financial district, the day before he flew home. The same configuration is now listed fifty per cent higher, and the expectation is that memory prices keep climbing for years on AI demand alone. Apple, which is as good at supply chains as any company alive, will currently take your order for a Mac Studio and deliver it four months later.
Unified memory is what makes a machine this size viable. In a conventional computer the operating system gets its own RAM and the graphics cards get theirs, the two do not talk to each other much, and you have to provision both correctly at the moment of purchase. When the CPU and the GPU address one space, the model and the operating system draw from the same pool. They still compete, and you stop having to guess the split years in advance.
The tour ran through Warp, which he uses because the terminal itself is agentic and he can ask it in plain language which processes are eating the machine, a question that otherwise means reading a process list and wondering what is safe to kill. Inside Warp he runs Claude Code, and that afternoon he had asked it how to demo the Spark for this meeting; it told him to install Open WebUI, so he did. The result looks like an ordinary chat window with an ordinary model picker, except that everything in the picker runs on his desk: Nvidia’s Nemotron, Google DeepMind’s Gemma, Qwen in the newest version people are still mapping the limits of, and GPT-OSS at 120 billion parameters. He also installed Falcon at seven billion, because it is supposed to be strong in Arabic, and found that it is not. The live query then refused to run, twice, having worked fine minutes earlier.
What he could show instead was an answer from earlier in the day, where the model had listed twenty myths about the Roman Empire, decided on its own to narrow them, double-checked itself and settled on seven. Before any of that reasoning appeared there was a long pause while the model loaded from disk, which is the difference between dense models, which have to be loaded whole even though only part of them answers, and mixture-of-experts architectures, which load a piece at a time. Meanwhile, while he was talking, he had asked Claude Code in the background for a diagram of the machine’s architecture, and it was waiting for him when he turned back. The Spark sits on a VPN, so he reaches it from his phone and can give other people access to it.
The workflow he is happiest with is transcription. Audio from a meeting like this one transcribes at five times real time, and the same local pipeline then separates the speakers, strips out what is not speech, repairs the punctuation and produces a structured note, with nothing leaving the machine at any point. The one he is least happy with is coding. Qwen benchmarks well on code, and a coding result depends as much on the harness as on the model, so he pointed Claude Code at the local model instead of the usual backend and gave it a real task. Four or five hours, hundreds of tests, loops of correction, and it never finished; at the end it concluded the job was not worth continuing. He measured success and failure only, not difficulty, and he has not yet tried Codex or another harness. His reading for now is that the machine is genuinely good at structured language work and not yet competitive with Claude for development, and that the rate of change matters more than either number, which is why he wants a fixed suite of evaluations to measure against. Without one, too many pieces move at once to tell whether anything improved.
The pushback was immediate and direct: local models are not worth it while subscriptions are subsidised this heavily. The member making the case runs fourteen accounts and around fifty agents through a harness of his own that balances the load across them, and fifty local models is a server farm, not a desktop. He has connected Ollama and OpenRouter, tested Kimi K2 and K3 and DeepSeek, and found that whatever he saves in tokens comes back as extra iterations and wall-clock time, and iterations are the number he is trying to reduce. He conceded the privacy case at once, having built a local email automation for exactly that reason.
The founder agreed about the subsidy and put a number on it. ccusage reports the tokens a subscription has consumed, which makes the comparison arithmetic: the same work billed through the API would have run to something like a hundred thousand dollars against five or six thousand paid. His own habit is to split the work by model, planning with Fable, writing code with Sonnet, writing tests with Opus and running the correction loops with Opus, on the grounds that planning with the largest model repays its cost and coding with it does not. The machine earns its place for a different reason again. He is building workflows for regulated environments where the question of whether the data would really be at risk never comes up, because the data is not permitted to leave the organisation at all, and on-premise inference is the only compliant answer. In some regions Microsoft will guarantee that your data stays in region at rest and give no guarantee at all about where it travels during inference, and for certain regulated uses that gap ends the conversation.
The other member then showed what a bad month had forced him to invent. His codebase has reached five million lines, his account-switching tool had been working perfectly and was comprehensively broken by the AI over several days, and being unable to use it pushed him into rethinking the whole approach. He calls the result contract coding. The code is cut into blocks, each defined by its inputs and outputs and each carrying a description; inside a block anything may happen. The AI stops reading code and reads specifications. A usage map records which block calls which and what variables pass between them, so an agent can see the dependency structure without opening a file. Input and output ranges are declared, which means a value straying out of range can throw a debug log by itself. The blocks carry their history: which agent touched which line and when, and every task, plan, feature and piece of feedback attached to the block it concerns, with enough description that a documentation agent can read the blocks and publish a wiki without reading a line of code. He estimates it removes ninety per cent of the tokens and ninety per cent of the reading.
His diagnosis of why he needed it went further than the mechanism. He already had a memory system he considers good, with full file-change tracking and per-agent task history, so the problem was never that the AI could not find things. The problem was size. With hundreds of connections between the parts, an agent that is even slightly lazy greps for a few fragments, reads them, changes logic it has not understood, fixes one bug and opens two more. Specifications small enough to be read in full are a way of making that laziness harmless. The two of them converged on planning from opposite directions. He runs plans of twenty or thirty pages, detailed to the point of tedium, generated and criticised by different agents until they agree. A good plan, he finds, is half the total work and makes the implementation straightforward, while a bad one pushes the work later, where every fix breaks something else.
A third member closed by saying he runs local models on a laptop that turns out to be more capable than he expected, that he would like a Spark of his own, and that he will have something to show next time. The founder ended on his book, which can be bought on Amazon and downloaded free from his site under a Creative Commons licence. A friend sent him a screenshot yesterday of it sitting on a pirate site, remembering that he had said ten years ago he would be proud to find it on one. He is, even though nobody needs to steal a book that is already free, and he takes it as a sign the thing has value.
Thursdays, 21:00 Rome time, on Discord. Bring what you built.