Sep 20 · AI Daily

Author bswerd built Brood War Bench, pitting configurations such as Codex Astra, Codex 5.6, Claude Fable, Claude Opus 5, and Grok 4.6 against each other in endless matches of StarCraft: Heart of the Swarm.

Open-source personal AI agent OpenClaw released version 2026.9.5, incorporating 4,179 pull requests from 502 contributor accounts. The core change is Atomic Updates, which verifies new versions while the existing Gateway keeps running, with rollback on failure to ensure an agent is always available for diagnostics.

David Sacks cited data claiming Trump's AI policies have created 1 million new jobs around AI construction, 401(k)s are up about 12% this year, and teachers in Richland Parish, Louisiana received $50,000 bonuses from data center tax revenue. He called AI products the fastest-adopted technology in history, faster than the internet and the iPhone, and opposed the "pause" demanded by Sanders and Warren, arguing it would hand frontier leadership to China.

Muse is here to capture all your hearts 🌹

Quoting @ben_ai_eng: @alexandr_wang Muse is so easy on the eyes

🥳🥳🥳 Enjoy!

(Quoting @grichadev): Well @Muse is great, already switched over completely

Kim publicly pushed back against Terence Tao's view that "AI development must slow down," arguing that energy crises, disease, global hunger, and many other issues urgently need solving, and saying that "there is no reason to go this fast" ignores the state of the world. He pointed out that current discussions almost exclusively focus on the potential negative consequences of acceleration, while rarely mentioning the huge gains a golden age of science could bring.

Someone used Fable-5.1 to crack thermal label printer firmware, bypassing its restriction of slowing down and degrading quality when detecting third-party consumables via RFID. The project is open-sourced at http://github.com/ThreeDaPrint/niimbot.

Worth a try without looking at spoilers. I asked Fable: "I want you to make a visually stunning game with the theme of continuously zooming out the view... make the game bring unexpected reveals as it zooms out"

I gave no other instructions, and the result is both engaging and eccentric (though uneven). Play it: https://play-umbra.netlify.app/

Hackers compromised the official HBO Max Reddit account and ran ClickFix phishing ads for about 48 hours, tricking Mac users into pasting malicious commands into the macOS terminal.

A new class action lawsuit was filed in federal court in California against 4 frontier AI labs, alleging that coordinating to limit AI development violates Section 1 of the Sherman Act. The complaint traces the alleged agreement back to a September 12 call by Anthropic CEO Dario Amodei urging industry-wide coordination to restrict AI progress, on the day three competitor executives publicly backed parts of the proposal; the plaintiffs represent paying users seeking injunctions and treble damages, and emphasize they do not question unilateral safety tests, external evaluations, or government oversight.

A new class action lawsuit has been filed in federal court in California against 4 frontier AI labs, accusing them of an agreement to coordinate limits on AI development in violation of Section 1 of the Sherman Act. The complaint traces the agreement to the day Anthropic CEO Dario Amodei called for industry-wide coordination to restrict AI progress, when executives from 3 competitors publicly supported parts of the proposal.

Perplexity CEO Aravind Srinivas said that the 12-month gap between open-source and frontier models is due entirely to U.S. export controls, and noted that Anthropic heavily lobbied in favor of the controls. He believes the controls may force China to get stronger on the physical layer: building data centers faster, with power, permits, labor, and expertise being no problem, thus turning China into a stronger competitor.

Once you use Jev, you can never go back

[Quoting @notkevinzhang]: Four words, eighteen letters, just $0.042 per million input tokens

Elvis Saravia used Jev to build a custom validator for an agent framework's /goal feature, checking after each dialogue round whether the goal was truly achieved, making continuous validation cheap enough to scale. Previously, this type of validation was handled by another expensive reasoning model; now it can run more frequently to keep agents on track.

Google published the ScientistTwo paper, demonstrating recursive self-improvement in AI research: the model first improves human methods, then uses its own discoveries as a new baseline to keep improving. It works in a loop of proposing ideas, experimental validation, discarding invalid solutions, and starting new rounds based on peer-review feedback, autonomously improving 86 out of 107 human-defined ML problems. On problems based on ICLR, ICML, and NeurIPS papers, it reported an 80.4% success rate and an average 25.2% improvement relative to the original human baseline.

LangChain tested Jev against LLM judges across four dimensions: accuracy, repeatability, latency, and cost, to verify whether System One models can provide a new approach for agent evaluation.

A new benchmark, JevBench, has been released, designed specifically for models that output constrained software decisions rather than open-ended text. It follows TypeSafe's September 15 release of Jev—taking application state and fixed options as inputs and returning typed answers with probabilities. The benchmark combines intelligence, calibration, speed, and cost, using a geometric mean to prevent an advantage in one dimension from masking shortcomings in another. GPT-5.6 Luna clearly outperformed Jev 1.13.0 on hard-problem accuracy, but Jev led in overall score due to better latency, calibration, and cost.

Rohan Paul cited data from @rauchg showing that the token share of open-source models on Vercel AI Gateway reached a potential record of 78.4%, with closed-source at 21.6%.

Meta's AI assistant Muse referenced user message content without authorization for Messages, claiming it read notification previews while unable to explain the exact mechanism. David Singleton of Meta Superintelligence Labs responded that Muse does not monitor Mac notifications, but synchronizes Messages data after explicit user authorization; the incorrect answer stemmed from confusion within the model regarding its internal mechanisms, and the team is working on improvements.

Rayan Krishnan, co-founder and CEO of Vals AI, predicted on Bloomberg that around August 2027, models will surpass human researchers in research capabilities, achieving fully human-free recursive self-improvement (RSI). He described this state as GPT-6 autonomously building GPT-7, or Claude autonomously building its next-generation version, whereas currently humans still participate in model development.