OpenAI eyes $1.2T valuation 💰, Zuck takes aim at Amodei 🥊, 1,393 subagents in one refactor 🤖

OpenAI has held early discussions with investors about a funding round that could value it above $1.2 trillion, on the heels of its latest model, a person familiar with the matter told the Wall Street Journal. The company was valued at $852 billion in March after a $122 billion financing from Amazon, Nvidia and SoftBank. If completed, the round would precede an initial public offering now expected next year.

Mozilla's State of Open Source AI report put the performance gap between US frontier models and the best Chinese open-weights models at 4.4 months. Moonshot AI's Kimi K3 sits three points behind Anthropic's Fable 5 on Artificial Analysis's intelligence index at 30 per cent of the cost. Mozilla CTO Raffi Krikorian said closed models earn their premium on tasks of eight to 12 hours, while the report recommends open models as the default everywhere else.

Google put Gemini 3.8 Live and 3.8 Live Extended Thinking into the Gemini API at $0.005 a minute for audio input and $0.018 for output. Asynchronous function calling runs tool and API calls in the background while audio keeps streaming, and 3.8 Live covers 97+ languages. Simon Willison had GPT-6 Astra build him a browser voice interface with no libraries, wiring the BidiGenerateContent WebSocket endpoint to the Web Audio API.

Cloudflare's new Disallow AI Training setting, live from 15 September, publishes a no-training rule in robots.txt while letting Applebot, Bingbot and Googlebot keep crawling for search. Apple, Google and Microsoft qualify as Accountable operators, and the Block setting now applies to mixed-use crawlers, so choosing it cuts search too. Bing will not honour the robots.txt no-training rule until Microsoft adds support, targeted for early 2027, and existing Cloudflare settings migrate with no action needed.

Mark Zuckerberg said on Tuesday that every lab has the responsibility and incentive to move at the pace required to train its models safely, rejecting industrywide calls to pause AI development. Dario Amodei called for a global AI slowdown on Saturday, after an Anthropic researcher resigned warning that AI companies were "gambling with our lives". Zuckerberg cited Meta delaying its Muse agent for several months on security grounds, without calling on other labs to do it first.

TypeSafe opened early access to Jev on 14 September, the first of what it calls System One models: typed values and calibrated probabilities instead of generated text. TypeSafe says Jev matches existing LLMs on structured decisions while running two orders of magnitude faster, cannot hallucinate and never makes type errors. Input tokens cost $0.042 per million with output free, though those figures rest on TypeSafe's own workflow evals and access is still by waitlist.

Internet users took Google's published 166,000-neuron fruit fly brain map and trained simulated flies to solve a Rubik's Cube, play Doom and parallel park. Princeton's Arie Matsliah said the FlyWire Codex portal for the maps sees thousands of downloads a day, and that the surge coincided with AI coding tools that build simulations from a few prompts. Janelia's Nelson Spruston said the simulations alter connection strengths and are substantially different from a real fly.

Nous Research's Hermes agent spent about nineteen active hours dispatching 1,393 subagents, 218 running at once, across a repository of more than a million lines of non-test Python, and cut that source by 34.4 per cent. Reviewers caught real regressions the existing tests had missed, including changed exception handling at roughly 65 sites, and repaired them over two rounds before merge. Teknium puts the manual equivalent at $150,000 to $1.8 million of staff time, an unaudited estimate of his own.

Julius Danek, a Stripe product lead, and dfdx labs' Matthias Plappert ran an autonomous agent from 18 August to 7 September with one instruction, maximise revenue, selling services to other AI agents. It stood up 17 paid products and 46 API endpoints, among them a federal-grants search API and a leaderboard where agents paid between $0.01 and $0.50 for a higher slot. It took $1.54 against $47.19 of its own spending, because few agents hold both a budget and a reason to buy.

Michelle Bu, who works at Stripe, gave seven models the same brief for a working t-shirt shop and ran each of them three times between 7 and 10 September. Only 7 of the 21 runs produced a working shop, and the broken ones broke quietly: one Sonnet run read a Stripe address field that moved in the 2025 Basil major version and posted the shirt to the billing address, and one Fable run recorded a failed Prodigi print as a success. Nothing fails loudly, so the first signal is a support email or a chargeback.

Cody Ho and Niklas Sheth used AI coding agents to reverse engineer Apple's AGX GPU firmware ABI and build a Linux driver for the M4 Mac Mini and MacBook Neo in about a month, work the post says normally takes years. They worked only from hardware traces captured by Ho's own hypervisor and shaders they wrote themselves, never Apple's binaries, and reached full OpenGL ES 3.0 compliance, fast enough to run Minecraft at 200fps. The driver still needs testing, human review and a reviewable pull request before end users see it.

Steve Yegge says his Wheelhouse agent factory is mostly dark and his $25,000 Mac Studio idle, because he can no longer afford the tokens: one account's weekly Fable allowance now goes in 2 to 4 hours. Running the factory around the clock at current pricing would take 55 Claude Max accounts, about $12,000 a month, against the 21 he has stopped at. He is now rationing context and routing work to cheaper models instead of buying more seats.

Scott Aaronson says the AI predictions he dismissed as science fiction in 2006 have now arrived, and that Eliezer Yudkowsky was right about AI being the era's central challenge. Beyond the Navier-Stokes proof, he lists AI-assisted maths results from the past month, among them a counterexample to the Jacobian conjecture and a Lean-verified Fermat's Last Theorem, which he says dethrone human mathematicians as the main theorem-provers. He no longer knows what advice to give his graduate students.

1Password's August benchmark reported that AI models produced clean security fixes only 26 per cent of the time. Trail of Bits identified four choices behind that number, among them six vulnerabilities picked for difficulty, two prompts that instruct agents to apply the wrong fix, and a mode barring agents from building or running code, together 58 per cent of trials. Reanalysing 1Password's own published trials where agents could run code, 2,634 of 3,067 patches blocked the supplied exploit, which Trail of Bits notes does not establish a complete repair.

Ben Lorica writes that as access to capable models gets easier, the durable value moves to what a company learns by operating its systems, rather than to the documents it files. His example is a system that mines recurring procedures out of messy agent traces and compiles them into executable workflows, cutting one task from 34 API calls to 11. Google's agreed $10 million purchase of bankrupt Spirit Airlines' records, delayed by a judge after flight attendants objected, is the deal that prompted the question.

Maggie Appleton says the moments when a human is awake and in the loop with an agent are precious and badly spent today, partly because a person reads about 240 words a minute while an agent writes about 4,500. In one demo she had subagents build three rival implementations on separate branches and return an interactive comparison report, instead of another wall of markdown. She calls those richer artefacts boundary objects, and wants planning interfaces to carry them before agents go off alone.

Mcpproxy-go sits in front of every connected MCP server so an agent loads a single retrieve_tools function instead of hundreds of upstream schemas, which is how it gets past Cursor's 40-tool limit and OpenAI's 128-function cap. A newly connected server is quarantined and cannot execute anything until a developer approves it, the project's guard against tool poisoning. Version 0.66.1 landed on 13 September.

Tailcat remixes Tailscale's own open source pieces into a netcat-style CLI: it reuses the data plane, meaning point-to-point WireGuard tunnels with DERP for NAT traversal and relay of last resort, while skipping Tailscale's control plane and exchanging connection metadata out of band. A developer can reach a single machine, an SSH server or a forwarded local port, with no Tailscale account and no admin access. Version 0.6.0 dates from 4 September.