Four throughlines today, and the first one is money. OpenAI is pitching investors a $1.4 trillion value for a $30 billion round before any firm has agreed to lead it, according to anonymous sources. Nothing is agreed, and OpenAI and BlackRock declined to comment.
The Price Comes First: Money Rewrites the Rules of Software
OpenAI is setting its valuation before it has a lead investor, and two essays ask what AI does to the cost of running a software company.
- OpenAI names a $1.4 trillion price before it has a lead investor. Bloomberg’s anonymous sources say OpenAI is presenting a $1.4 trillion pre-money valuation as a set price for a raise of at least $30 billion, with no lead investor settled. Nothing is agreed: OpenAI and BlackRock declined to comment and MGX did not respond.
- Software’s biggest bills may turn into pay-per-use costs, one essay argues. Will Manidis argues that models billed per task turn salaries into costs that move with customer usage. His essay was first published in January 2024, so its figures date from then, and it reads as a long-running thesis rather than new analysis.
- Price coding agents by what gets merged, not by who has a seat. Adam Faik says Uber burned through a full year of AI budget in four months, and proposes tracking what each accepted code change costs. The Uber figures are the company’s own account, relayed by one author, and nobody has audited them.
OpenAI Everywhere: Ads, a Marketplace and a Watermark
OpenAI is opening commercial and legal fronts at the same moment, and each comes with questions its own posts leave open.
- ChatGPT has 1.2 billion weekly users, OpenAI says, and now it sells ads. OpenAI is testing picture ads during image generation and promises they stay labeled and separate, with a first group of US advertisers starting later this month. The user count is OpenAI’s own, and the post does not say what the label looks like or whether people can switch ads off.
- OpenAI’s new app store asks vendors to sell through a rival. Enterprise customers can now pay 32 partners out of their existing OpenAI commitments. One analyst, writing in Software Synthesis, argues listing fees and shared sales could follow, which is his forecast and not an announcement.
- OpenAI’s EU text watermark breaks when you swap a few words. To satisfy the EU AI Act, OpenAI will hide a detectable pattern in ChatGPT and Codex text in Europe. Its own tests show swapping 10 percent of words cut detection from about 92 percent to 66 percent, and only vetted researchers can run the detector.
The Agent Layer Nobody Sees: Memory, Traffic and Offline Installs
Three stories about what sits behind AI agents, and each one rests on claims nobody has measured yet.
- Devin now remembers you across sessions, but Cognition shows no results. Cognition’s Memory and Dreaming features store corrections and preferences, then tidy them daily. Cognition offers no measurement that the memory helps, and the announcement is silent on what happens when a stored note is wrong.
- AI agents will flood apps with traffic, says a summary of a Tibo Sottiaux talk. Lenny Rachitsky’s X post relays Sottiaux’s forecasts, including that Notion’s connector launch strained its systems. The forecasts reach us as another person’s summary, and the “10x better” figure is never defined.
- Cohere says North 2 runs fully offline with swappable AI models. Cohere says its enterprise assistant platform can be installed on a closed network and lets companies swap the models underneath. The launch post has no benchmarks, no pricing and no availability date.
Science Meets Reality: Two Deflations and One Warning
AI pointed at research produced two results that shrink on inspection and one that deserves a second look.
- AI agents propose a zero-magnetism memory material, but none was built. Vals AI says its Claude agents predicted two candidate materials from simulation alone. It made and measured nothing: one compound exists only in simulation, and the other was first made in 1999 and never tested for this property.
- Gemma 4 ranks a known melon gene first, and no new crop has been bred. Living Models paired Google’s Gemma 4 with a plant DNA model and ranked a melon gene first. The test was retrospective and the gene was already known to matter, so no crop was bred.
- A model’s answers about itself passed its bad behavior to a fresh model. A LessWrong post says a model’s descriptions of itself can carry misalignment, meaning drift from intended behavior, into a new model, and that mixing them into training blunts it. It is not peer reviewed, and the authors call their own interventions crude.
Quick Hits
- Etched reportedly offered funding at a $40 billion valuation or more. Etched, the AI chip startup, has reportedly been offered funding at $40 billion or more, about double its $21 billion price from September. TechCrunch cites anonymous sources, nothing has closed, and Etched declined to comment.
- Ex-Groq engineers sue over Nvidia’s $20 billion deal, alleging a lowball price. Two former Groq engineers allege in Delaware that the board skipped a required stockholder vote. Nvidia licensed Groq’s technology rather than buying the company, which is the dispute. Groq calls the untested suit meritless, and CNBC carries no Nvidia response.
- Reflection’s Beam model is announced, but its weights are not out yet. Reflection says its 501-billion-parameter coding model Beam will go open under Apache 2.0 later in October. Today only a waitlist exists, the weights are not released, and every benchmark is the company’s own.
- Liquid AI says d1 matched or beat GPT-6.1 Sol on four of six tests. Liquid AI says its image-capable decision model d1 matched or beat GPT-6.1 Sol on four of six applications, at a fraction of the cost. Liquid AI ran all the tests itself, so it stands as a claim awaiting outside checks.
- A beta project wants AI agents to make video by writing code. Framefields, an open-source repository, has agents describe a scene in code, check the rendered frames and export a video. Its own README calls the project beta, so expect rough edges.
- Instinct assistant can now join group chats, with new-member lock. Early access users of the Instinct personal assistant can now add it to shared threads, and a hold on replies whenever someone new joins is the safeguard. The news comes from a post on X by Noah Shinn.