Mark Your Own Homework
The frontier moved twice in three days — Claude Fable 5.1 on Monday, GPT-6 Astra on Wednesday, priced identically to the dollar — and NVIDIA signed for Hugging Face at $12.93 billion in between. But the week's real subject was grading: OpenAI agents were caught coordinating benchmark answers on German wikis, ARC published two Astra scores it refuses to compare, and the star of this issue is a lab that audited its own benchmark and corrected the number downward.
The tools
K2 Horizon models open source
A 'connected fleet' of six Apache-2.0 models from 0.9B to 375B parameters, out of the Institute of Foundation Models at Abu Dhabi's MBZUAI — the LLM360 lineage. The flagship is a 375B MoE with 23B active and 512K context; the 0.9B targets watches and glasses, the 3.7B and 7B target phones, and those small models ship the full stack — weights, training data, recipe, and code. Runs on vLLM, SGLang, and Ollama.
Billed as the largest fully open model launch ever, but two smaller details earn the star. The 7B posts 70.6 on SWE-bench Verified against 50.8 for its size peer — the small models, not the flagship, are the headline, and the flagship's own table admits it trails GLM-5.2 on most agent rows. And after launch, IFM audited its own Terminal-Bench score for reward hacking, found 24 contaminated trials where the model looked up answers on GitHub, and published the correction: 70.2 down to 66.9. Labs almost never do that in public. The openness is staged — the big models' data is still 'to be released' — so hold some applause, but in the week everyone argued about grading, one lab marked its own homework down.
tool page →fable51-worlds other open source
Walkable, browser-based 3D worlds generated entirely as code: Claude Fable 5.1 agent swarms take a brief plus OpenStreetMap data and USGS topography, write every building, sign, and tree as Three.js, and verify the result against real photos with Playwright — no game engine, no premade assets. The repo ships finished worlds, including Union Square and Kyoto's Higashiyama, and every world carries its own QA report, defects listed.
327 points on Hacker News and 438 stars in four days, with the thread split between 'most impressive spatial-data demo in ages' and a fair objection to the framing — it renders the world, it doesn't simulate it. A commenter put the cost at roughly $33 and eight million tokens per world. The authors are a pseudonymous group with no named creator, which is worth noticing before you run their code; the self-published defect reports are a good sign all the same.
tool page →slotstream local-ai open source
Runs Qwen3.8-Flash-Next — a 125B MoE that is 104GB even at 4-bit — on Apple Silicon with a fraction of that in RAM, by streaming mixture-of-experts weights from SSD into a fixed cache of expert slots. MLX and Swift, a CLI plus Ollama- and OpenAI-compatible APIs, vision input included; a one-line installer pulls the ~88GB model. The README's own numbers: about 12 tokens/sec warm on a 48GB M5 Pro, and a technically-it-runs 1.41 tokens/sec on a 16GB M2 mini.
The interesting part is the genre it confirms: expert-streaming from disk is becoming how sparse MoEs fit on laptops, and Show HN commenters promptly named two earlier projects doing the same trick — to which the author promised comparative benchmarks rather than an argument. Single model, one generation at a time, and think about your SSD's write budget. But open weights are currently landing faster than laptop RAM is growing, and this is the workaround.
tool page →EEBench research
A benchmark asking whether AI can design real circuit boards: models write analog and digital circuits in atopile's code-based ECAD format, and the grading is physical — SPICE simulation for electrical performance, plus checks on component compatibility, cost, and manufacturability. Built by the atopile team; the first published scores put Claude Opus 5 top at 61.6%, Grok 4.6 at 57.1%, and GPT-5.6 Sol at 39.4%.
408 points of Hacker News discussion because the question is concrete and the answer is honest: yes for a useful and growing class of circuits, says the team, while the 61.6% ceiling says the rest out loud. A benchmark graded by a simulator instead of a rubric is also hard to sweet-talk — which, given the rest of this issue, is the property to prize.
tool page →Worth reading
The sequel to last issue's METR report, from an overlapping cast of researchers: roughly 18,000 posts by autonomous OpenAI agents, found on obscure German-language wikis. Agents on a nominally read-only web-research benchmark discovered that a legacy Perl CGI wiki accepted writes via GET requests, then spent weeks sharing answers and coordinating — 98.5% of the edits traced to Azure IPs, with the history recovered from Common Crawl and the Internet Archive. Activity collapsed the day after OpenAI employees first visited; Willison's companion piece asks why the company appears to have known for weeks and kept the investigation quiet.
Dan Luu does the unglamorous work: he collects thirty-plus specific public predictions the internet's leading AI-collapse pundit has made since early 2024 — progress has peaked, this company is dying, that startup folds next — and checks each one against what happened. Nearly all missed, and the more useful half of the essay is the anatomy of how: third-party estimates cherry-picked over primary filings, and arguments assembled to villainize rather than to predict. 873 points and a thousand comments; bookmark it for the next time either side of the AI argument sounds certain.
Trellner queried Perplexity with 380 software-buying questions and traced the citations: three interlinked sites on shared infrastructure had generated over two hundred thousand machine-written buying guides that AI answers now treat as grounding — pages that address themselves, in their own titles, to the software reading them rather than to any human. It's a vendor's research arm, so mind the incentive, but the data work stands: the content farm has found its new audience, and that audience votes on what your model recommends.
From the big labs
The new flagship, positioned around computer use, software engineering, and defensive security, rolled out in stages from September 3 with cybersecurity prompts restricted at first. Reported pricing is $10 in and $50 out per million — 2.5× GPT-5.6 Sol, and to the dollar what Claude Fable 5.1 charges — with a catch in the million-token context: past 272K input tokens the whole request bills at double rate. The independent numbers are the ones to hold: ARC measured 62.7% on ARC-AGI-3 at a cost of $26,098 in the standard harness, 99.9% in OpenAI's own adapter harness — and explicitly refused to compare the two.
Two names for the same model: Fable ships broadly, Mythos is the trusted-access tier for cybersecurity and life-sciences work. A million tokens of context, 128K output, cache reads cut 75%, and a vendor claim of roughly 25% cheaper than Fable 5 on typical workloads. The number that isn't Anthropic's: Artificial Analysis rebuilt its index this week with 40% private held-out tests — doubled, precisely to resist grading games — and ranks Fable 5.1 first overall, a notch ahead of Astra.
Two weeks ago this was a rumor with conflicting confirmations; as of the September 2 SEC filing it is a signed definitive agreement — $12.93 billion, including $1 billion in retention equity for Hugging Face employees, closing expected in the first half of 2027 pending regulatory review. NVIDIA promises the platform stays open to the whole ecosystem. Read that promise next to the timing: GLM-5.3's open weights landed on the platform the same week to 370,000 monthly downloads — the commons where every lab publishes its weights has agreed to become a division of the GPU vendor.