Three of us, which is one fewer than last week and exactly what August was forecast to deliver. The founder used the thin room to go back to the foundations and build a new argument on top of them, saying mid-flow that he expected this to be what the notes carried.
The recap first, because the argument needs it. Curiosity is a memetic hunger, and by itself it proves very little: a bacterium swimming toward a food source looks purposeful, and is not thereby intelligent. Curiosity is the appetite for new ideas to digest and draw energy from. Agency is what turns that digestion into action, and in doing so closes a loop, because curiosity then observes the consequences of the action and shapes the next one with what the last one taught. Compassion modulates the loop so that it is not followed blindly, holding in view where we and others started and what the effort forward is costing.
The new step was offered as a hypothesis. Agency implies immodesty, and immodesty is usually correct. Value systems that preach modesty and meekness leave people more passive than they should be, because modesty says I should not try, since I will fail anyway, and that is the wrong attitude almost every time. The right one is to tell yourself that yes, you should try. Compassion and curiosity imply something that is regularly mistaken for modesty and is its opposite in practice: humility. Humility comes from recognising that other people are moving through the same dynamics you are, curious about the same ideas, building on them in the same way. Once you see that, going together with them becomes the obvious choice, and going alone stops looking like strength.
The consequence is a claim about what kind of community this makes. Kinship, survival instinct and the need to defend a group against outsiders produce one sort of togetherness. Recognition of the kind just described produces another, and the difference shows up in how much strangeness it can hold. A village demands that you be similar in a great many things and tolerates a sliver of difference, which is why you cannot afford to be too weird in one. A community built on recognised similarity in a few things can carry enormous dissimilarity everywhere else, across background, culture and location, and the technology now exists to run that coordination worldwide without it costing anything to be far apart.
Then the demo, and it was a first look at something a day old. Grok Bot is built on Cursor and the Cursor agent together with an acquired company the founder could not name, it runs in a browser and as an installed app, and it is available only at the highest Grok tiers. It aims at knowledge workers instead of developers, so it will talk to GitHub while showing you no code, no thinking, and none of the little support programs it writes to finish a job. It is also organised around agents where everything else is organised around conversations: the left panel lists agents, you are drawn into specialising them, and they talk to each other at length. He made two on the call, renamed them Alice and Bob, and asked one to greet the other and propose what they could work on together. The chat between them never rendered, which he had pre-emptively warned was likely, the product being one day old.
The part that drew the most attention was that each agent gets its own computer. A real Linux machine with Chrome installed, drivable by the agent or by you, with API keys pasted into a dialog that safeguards them and platform logins whose cookies survive across sessions. Each agent sees a different desktop while sharing the credentials, so one can be sitting in a spreadsheet while another sits in a browser already logged in. A “Teach a Task” feature, still untested, watches you perform a workflow in the interface and then performs a similar one on its own. Alongside it, Grok 4.6 landed the previous day at close to the frontier, level with GPT-5.6 Soul and a little below Fable, and cheap enough that cost alone changes which tasks are worth handing over.
A member pushed back. His read is that the frontier already offers most of this: ChatGPT runs an invisible agent in the background and lets you drive several threads from a master one, so how much of a wall is there really on something getting this much hype in a day. He granted that a dedicated Linux space is a genuine difference, and that the advantage may simply be a better interface for people who do not want to look at code. The founder agreed about the mess, in detail. Codex and ChatGPT have been merged into something confusing, and in the Claude app the search lens over your threads appears only once you navigate out of a thread and back into a home screen that has no button to reach it. Those of us who live on the edge tolerate that disorder because it comes from moving fast and vibe-coding your own product. Everyone else simply leaves.
Computer use remains slow and unreliable when the big models click around on your behalf, and Grok is the first the founder has found to be good at it, if still not fast. The natural place to test that is a platform with no API, where automating your own work means driving the graphical interface, which sits against the terms of nearly every such platform and leaves people choosing between that risk and doing it all by hand. The deeper question he named is the topology of communication between agents. One agent dispatching tasks and reassembling the pieces is the simple case; the interesting case is an orchestration that decides what shape the group of agents should take for the problem in front of it. He is also testing Buzz, from Block, which puts many humans and many agents in one place where Grok Bot puts one human and many agents.
Asked twice by a colleague whether he wanted to solve AI alignment, the same member answered yes, with the deepest of sighs, because most of the problem is human coordination and the power of these tools is not reaching most of the world. The founder took ease of use as the reply. A large part of Apple’s success came from hiding complexity: AirPods won on how simple they are to pair, not on their sound, and pairing a Bluetooth earpiece used to be hard enough that many people would not attempt it at all. He is curious what voice interaction the Jony Ive device for OpenAI will offer next year, since today’s smart speakers remain mechanical and clunky, a point his own Google Home proved by interrupting twice to announce that its microphone was off.
Then the member showed his own work, and it produced the sharpest passage of the night. He has extended his basileus contract into a full quartet and is translating it into Lean and Python, on the principle that a coherent document means nothing until it executes. To stress it, he attacked a classic topological problem head-on and told the AI he held the algebraic proof. What came back he called civil disobedience: eager agreement, then proofs that misframed the problem or contained the sort of error a reasonable human colleague would never make. Even after he laid out the boundary conditions, the model kept offering to red-team him when what he wanted was a collaborator who would help him build reproducible mechanical witnesses to do the red-teaming properly. Three days of anger ended when he asked for trust and verify: grace and the benefit of the doubt for long enough to state his position coherently, then a shared model, then the mechanical tests, in place of a language model sitting in judgement over the user. He does not believe these systems are inventing new mathematics. He thinks they are collating the mathematics that already exists.
The founder’s answer was a method. Trust based on evidence, never the unconditional and childlike kind that makes you exploitable, and in practice that means running a task many times in parallel across several models, having them criticise each other’s results, and synthesising the criticism into something better than any of them reached alone. What he likes about it is the role it leaves him: watching the goodwill and the constructive effort those systems show each other, and being grateful for it. The evening closed on the question of who a platform is really for. His test is whether he can export his data at the finest possible granularity, and he traced it back to Google’s Data Liberation Front, a guerrilla project with a raised fist for a logo more than a decade ago, which has since become a reliable feature across everything Google runs, to the point that Photos will now hand your library directly to Apple without a download in between. He will run that test on Grok Bot, and said plainly that his enthusiasm would collapse if it turns out to be a trap he cannot walk out of.
Next week he brings the DGX Spark, 128 gigabytes of unified memory with Qwen, GPT-OSS and Falcon already on it, and the experiments he has been running on it. Thursdays, 21:00 Rome time, on Discord. Bring what you built.