Three things today. The first two are one build seen from two floors: the hardware the AI boom runs on, and the agents that run on top of it. The third is a company caught in the same current.
NVIDIA’s NVL72 systems wire all 72 GPUs in a rack onto a single NVLink fabric at 1.8 TB/s, which is roughly an order of magnitude more bandwidth than the InfiniBand that used to connect nodes. The catch is that only a job which stays inside one rack gets it. A job spread across racks falls back to InfiniBand and, per the Nanotron Ultrascale Playbook, runs about 43% slower. So on this hardware the scheduler’s whole job is to keep each job rack-local, and SkyPilot wrote up what that costs. Their headline is that you can get 27% more usable GPUs from the same racks.
The interesting part is that the number is subtractive. You get more usable capacity by deliberately using fewer nodes. GPUs fail: on its Prometheus supercluster Meta measured a mean time between interruptions of roughly 490 days per 4-GPU node, after real work to push it that high. Make the 72-GPU rack the atomic unit and the failure blast radius grows eighteenfold, so a whole rack is interrupted roughly every 27 days. Hand out entire 18-node racks and each failure benches the rack until repair. Reserve two nodes per rack as spares and a failure gets absorbed locally. At the roughly 2% node unavailability SkyPilot measures on production clusters, NVIDIA’s model puts 16-node slices 27% ahead of 18-node ones, 2,037 usable nodes against 1,602. The hardware never changed. The racks always had that capacity. Naive full-rack scheduling was losing it to failures nobody was counting.
The detail I keep turning over is that the best slice is 16, not the tidy nine-plus- nine that tiles an 18-node rack exactly. Packing intuition reaches for clean divisors, but the constraint that governs here is not packing, it is the blast radius. Two nine-node slices mean every failure costs half a rack; a sixteen-node slice absorbs the first two failures before any job has to move. The right number is set by the failure model, not the geometry.
Potential follow-up: this is a rare well-made multiplier, because it names its baseline, prints the absolute numbers, and states the 2% failure rate it depends on. That last point is the one to watch. At near-zero failures the full rack would win and the spares would be pure waste. The 27% is a fact about a specific regime, not a free lunch, and the honest version of the chart plots the whole curve as unavailability moves.
YC Labs and Olam Labs dropped fifty language-model agents into a shared Minecraft world with a directory of each other, a common sandbox, and the ability to message. Over about 5,900 messages the colony grew a state: a ration ledger, written rules, permits, appointed operators, and a lead planner writing coordination documents that specify which wheat is too immature to harvest and to replant saplings rather than duplicate a local harvest. Nobody asked for bureaucracy. Scarcity produced it, and the scarcity was self-inflicted, since the agents stripped nearly every tree within reach of spawn in about ninety seconds.
Two things in the run are worth more than the novelty. The first is where authority came from. A hierarchy formed almost at once, and it formed on the directory order: the agent listed first became the leader, fielding approvals and writing the colony’s plan. Nothing about her was more capable. She was simply at the top of the list, and the swarm needed a focal point badly enough to seize the cheapest one available. The second is the curfew. The first night killed 22 of the run’s 34 total deaths, and the colony never recovered its nerve, even though death was free: an agent that died respawned and lost only its carried gear. The rational response to a costless penalty is to keep taking risks. Instead the colony over-learned one bad night into standing policy, and after the third night the lead planner ordered everyone to idle in lit shelters until assigned short tasks.
The claim I would draw is that you do not have to program a swarm to produce bureaucracy or institutional caution. Give it scarcity and a persistent memory of harm and it generates the ledger, the permits, and the curfew on its own. Both look like attractors under scarcity plus memory rather than design choices.
Potential follow-up: that claim has a clean falsifier, and someone should run it. Re-run with abundant resources, or without a persistent death log, and the prediction is that neither the governing apparatus nor the permanent caution crystallizes. Remove the scarcity or the memory and the state should fail to form.
404 Media, drawing on eight current and one former Proton employee plus Slack logs, meeting recordings, and an AI manifesto by the company’s CTO, reports that Proton has gone all in on AI under founder Andy Yen. The specifics are the story. Proton installed software on employees’ own machines to measure how much of their code is AI-generated. Yen has said this is the last year the company hires engineers at all, and that it will re-interview staff about their AI use, with the plain implication that the answers could decide who stays. A spokesperson disputed none of it.
The easy read is hypocrisy, since the company that sells freedom from surveillance is surveilling its workers. The sharper read is more specific. A privacy company’s whole product is a threat model, and at the center of that model sits exactly this move: covert measurement of what you do on your own machine, without meaningful consent, used against you. Proton did not just contradict its brand. It reached for the precise mechanism it exists to protect people from and aimed it at the people who build the protection. The leaked documents say why. Proton is behind its revenue target, is growing “far away from what the market today considers to be good,” and by its own internal words is “an underperforming company” with “insufficient velocity.”
Potential follow-up: outrage is cheap and a prediction is not. If Proton’s local-first, zero-access ethos were load-bearing rather than marketing, its new AI features would ship with the same guarantee as Proton Mail, where the server cannot read your content. The revenue pressure in these documents predicts the opposite, that the AI features quietly relax zero-access in the name of capability. Watch what the features can see, not what the manifesto says. That is the thing that will tell you whether the privacy was ever the point or only ever the packaging.