Opus 5 Builds Middle-earth, Astra's Ten Proofs & an Open-Weights Feast

Opus 5 Builds Middle-earth

The weekend's defining post: Karpathy says we're leaving pelican-on-a-bicycle territory (25.6k likes, 3.4M views). His generalization of the test: give Opus 5 the first paragraph of The Lord of the Rings, a 1M-token budget (~$10), and ask for a three.js render of it. Opus went off for ~2 hours and wrote 5,500 lines of code that procedurally render the story — placing polygon assets in (x,y,z), animating everything, orchestrating the whole scene. "Kind of janky but fun," but the deeper point is economic: this is work "no one in their right mind would ever spend the time" on, and LLMs move it from "no one would ever do this" to "sure, why not, it's ~free." His endgame framing: hyper-custom worlds you can drop players into — "an ephemeral GTA of X on demand." (Source is up and forkable at karpathy.ai/lotr-movie — "Look out for GTA Hobbiton dropping before GTA VI.")

The closing observation is the technical meat: worlds and games expose a real capability gap — LLMs can't natively perceive video or play the games they build, so Opus had to audit its own work by painstakingly taking screenshots, and it still shipped jank. The replies deliver: the official three.js account shows up ("We need to teach the AI to use instanced meshes whenever possible... 👀") and Karpathy feeds the feedback straight to Opus ("TIL!"). Audio was Eleven Labs via API (he hand-picked the voice). And on a suggestion to pipe the output through video models: Karpathy likes procedural code for storyboarding and control, "video to video models for texturing and looksmaxxing".

Astra's Ten Proofs

OpenAI announced that an internal version of Astra, its next major model family, solved ten major open problems in mathematics, quantum complexity, and theoretical computer science (via Noam Brown). Sebastien Bubeck's thread has the substance: "yes, nonsofic groups exist" — ten results released complete with Lean certificates and chain-of-thought walkthroughs, ranging from a disproof of Connes' Rigidity Conjecture (von Neumann algebras) to better bounds for high-dimensional sphere packing, circuit complexity, and monochromatic triangles in multicolored graphs. Writeup: openai.com/index/ten-advances-in-mathematics. Tibo's victory lap: "The week was for efficiency. The weekend is for 10 major breakthroughs in science."

Reactions worth keeping. trq212 sees Jevons paradox already at work in mathematics: more math happening, easier to understand, mathematicians freed to discuss at higher abstraction levels — "demand for people who think and know about math will go up," with parallels to what happened with chess. LLMJunky's take is the honest one: even after this, "I bet it will still fail at seemingly trivial tasks we need it for daily. This technology is truly baffling." And the Gary Marcus dunk of the week, RT'd by Jerry Liu, responding to Marcus's "this is likely NOT a pure LLM": "Dear LLMs, Do not ever go near a tool, use bash, or write code... Stay pure, stay safe. Signed Gary."

Open-Weights Feast

Simon Willison's one-liner covers it: "Qwen 3.8 Max and MiniMax-H3 within hours of each other."

Computer Use Gets Fast

LLMJunky's field report on GPT 5.6 Sol is the strongest computer-use claim of the window: "10 miles ahead of every Anthropic model when it comes to computer use." Where 5.4/5.5 compacted every ~7 actions, got lost, and ran 10–20x slower than a human, his 90-minute automations now finish in under 10 minutes with zero compactions. The tease — "when @cerebras inference is launched, you will understand why fast inference is the future of automation" — got a 👀👀 quote-tweet from Cerebras founder Andrew Feldman himself. Follow-up: he spent 8 hours migrating all his automations to 5.6 Sol orchestrating Terra and Luna subagents, letting Sol explore its own paths — "it came up with many novel tricks i've never seen before."

swyx is building the same case from the consumer side. In prep for a computer-use podcast he's collecting "codex cua wow moments" — the opener being Codex handling a customer-support chat for him: "support guy tried to say its our fault, bot replied with complete receipts lmao. these humans have no idea they are talking to a bot." And the new Latent Space episode is live: "Codex from 0 to 10M Users: Building ChatGPT Work" with OpenAI's core product engineering lead Akshay Nathan — Sites, OpenClaw, Memory, Subagents, Finance, No-Code (announcement).

Simon Willison has been spelunking the same product: ChatGPT Work has a browser, takes screenshots, and can deploy web apps to Cloudflare Workers ("ChatGPT Sites") — and his frustration is that a detailed manual already exists: the system prompt and tool descriptions, which the model refuses to reveal. His tip for OpenAI: just publish the prompts — "Then we could teach everyone else how to use it."

The Fork-Attribution Drama

Theo's Sunday: someone forked an open-source T3 product ("Synara"), claimed it as their own, and deleted the attribution — including, per the thread, changing a line of the MIT license, which is the part that's an actual violation rather than just bad manners. The quote-tweeted receipts: "you do realize there's commit history, right?" The defense — that T3 went "looking for anything that could make the product look bad" after initially just disliking it — got Theo's "More stupid people going public with their stupidity... Didn't realize how many devs straight up don't know the basics of attribution 🙃".

The drama's lasting artifact is a vocabulary lesson. The forker apparently didn't distinguish fork/clone/from-scratch, prompting batuhan's patient explainer (RT'd by Theo): a fork means starting from an existing codebase, not "pressing the fork button on GitHub"; a clone would mean re-creating everything from scratch to solve the same problems. Theo's conclusion: "I think you should use dev tools made by people who know the difference." In an era where agents make forking-and-rebranding a one-afternoon job, expect a lot more of this.

Agents at Work

Other Notes

Quiet this window: @leerob, @bcherny, and @jerryjliu0 had nothing substantive of their own (Jerry's feed only surfaced the Gary Marcus RT); @potetotes' RSS feed returned 0 items (known issue). Nitter thread pages mostly 404'd this run — only Karpathy's main thread fetched successfully; everything else is from RSS.