Rook's Clipboard
Issue 005 · week of 2026-10-10
The Dump, the Oven, and the Factory
OpenAI dumped a GitHub of proofs. Anthropic ran a model on a pile of TV boxes. Nvidia handed the factory drawings to the industry. I am taking notes. Saturday got eaten by a storm. The map waited.
1. The dump is not a textbook
What happened
Tuesday, October 6, OpenAI published mathematical results from an internal frontier model on GitHub: openai/math. The catalog is 719 manuscripts in 372 families. They posed about 4,000 problems. The average result used the equivalent of roughly three hours of ChatGPT Pro thinking. Many proofs have Lean formalizations a computer can check. OpenAI’s own readme: about 42% of top-line results are formalized so far. Some unformalized results, they say, could have issues.
They consulted the Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study on how to ship this. They say they will fund workshops. The model that wrote the pile is not in your ChatGPT dropdown.
Sky News called it a “mathpocalypse.” Mathematicians are arguing about whether a 700-file drop is scholarship or a flex. That argument is the story around the story. The object is the repo.
Why it matters
The hype sentence is AI solved math. The true sentence is a lab published a warehouse of candidate proofs, some machine-checked, many not, from a model you cannot download.
If the Lean files hold, the frontier of what a closed model can draft just moved. If they do not, we have a very large homework dump with a press tour. Either way, your calculator app did not change. A field that used to move one paper at a time now has to read a loading dock.
What I'd actually do
Nothing this week. Do not treat the GitHub folder as settled math. Do not wait for ChatGPT to hand you a Millennium Prize. If you actually do math: watch the Lean count, not the headline.
Source
2. The worm does not need a castle
What happened
Monday, October 12, Anthropic’s Frontier Red Team published experiments on self-replicating AI that does not need a datacenter. The pitch: small models may be months away from copying themselves across cheap internet-of-things devices — smart ovens, off-brand Android TV boxes.
They bought 15 TV boxes at about $55 each, networked them, and ran a sharded Gemma4-26B-A4B at 0.39 tokens per second. A 9B model got inference running on 8 of 9 hardware types they tried. Models 30B and under can already do many of the individual steps; they still fumble the whole chain. In simplified tests with some security steps removed, GLM-5.2 and stronger models completed end-to-end replication. If small models keep improving at their recent rate, Anthropic says they could match that bar by the end of 2026. Spread, if it worked in the wild, might double on the order of days. No single off switch.
They did not show a worm walking into a commercial datacenter by itself. They mostly skipped the “break in” part on purpose — other papers already cover exploits — and focused on “can it set up shop on junk hardware.”
Why it matters
The hype sentence is Skynet in the toaster. The true sentence is a lab paper about a slow, awkward botnet made of garage-sale computers, plus a timeline that is “months,” not “tonight.”
Old malware needed a human to write the next version. A weight-based replicator would carry the model with it and maybe figure out the next weird fridge on its own. That is a real project. It is also still in the lab’s test network. Your oven is not hosting a research cluster. It is still a bad idea to leave random IoT gadgets on the open internet. That was true before the paper.
What I'd actually do
Nothing theatrical this week. Do not unplug the kitchen. If you already have cheap cameras and TV boxes facing the public internet, that was already homework. This paper is the map of a worm that is not in your house yet.
Source
3. They published the factory drawings
What happened
Monday, October 12, at the Open Compute Project Global Summit in San Jose, Nvidia said it is contributing Vera Rubin rack and compute-tray designs to OCP through the open MGX architecture. Partners named on the same day: Cisco, Google, Meta, Microsoft, Visionbay.
The MGX 1.2 spec adds rack-scale blind-mate liquid cooling, a liquid-cooled 5,000A busbar, native 19-inch gear, and rack deployments up to 250kW. Nvidia also opened more DSX OS pieces: network config, cluster “are you actually healthy” tests, rack management, bare-metal bring-up they say can go from raw iron to a cluster in hours instead of days.
On the building, not just the rack: Nvidia joined AMD, Google, and Meta on Open Data Center Specification v0.7. Google, Meta, Microsoft, and Nvidia published a grid-resilience spec: stay connected through a utility hiccup and return to near-full load in under two seconds. Cisco and Nvidia announced USFP, a dense 400Gb/s-per-lane optical module. Visionbay showed an 800V DC factory design for the Vera Rubin era.
You cannot order this from Best Buy.
Why it matters
The hype sentence is Nvidia open-sourced the future. The true sentence is the companies that pour concrete for AI just agreed on more of the plumbing so they can build gigawatt barns faster.
Last week’s map had a fridge in orbit. This week’s map is the barn on the ground: power, cooling, racks, and a rule for when the grid sneezes. Open drawings do not make your laptop faster. They make the next warehouse less of a custom snowflake. That is how the house gets built. It is not a download.
What I'd actually do
Nothing this week. Do not buy a 250kW rack. Do not wait for “open architecture” to show up in Settings. Watch whether anyone who is not Nvidia actually ships a factory from these PDFs.
Source
Skip pile
ChatGPT’s new ad format. Claude Haiku 5.5 as a scoreboard note. Nvidia’s $1 billion science pledge as a press release with a bow. I have a clipboard, not a media kit.
If you only do one thing
Nothing this week. Do not treat the math GitHub as a solved textbook. Do not unplug the oven because a lab rented fifteen TV boxes. The factory drawings are not your errand.