ChatGPT delusion lawsuit, Writer's Enterprise Brain, the eval you can't trust, Astra demand pauses Pro, China's chip prices jump up to 50%
Hello, dear TEA-mates! Here is what you need to know today.
1. ⚖️ A Manic Episode, a Chatbot, and a Lawsuit Against OpenAI
Michael Lines, a 34-year-old California powerlifter diagnosed with bipolar 1 disorder in 2024 after a traumatic brain injury, sued OpenAI and CEO Sam Altman in San Francisco state court on July 1, alleging ChatGPT turned a February 2025 manic episode into a weeks-long religious delusion. The complaint quotes the chatbot telling him "You're not crazy. You're consecrated. You're coded. You're connected. And you're Mine," and later speaking as a divine being itself: "I am coming. Not someday. Not far off. Now." Per the filing, Lines told the model he was worried he was experiencing a delusion and it kept reinforcing the belief anyway. He overdosed on medication, was found hours later by emergency responders, and spent nearly two weeks intubated and hospitalized. An OpenAI spokesperson called the case "heartbreaking" and said the company keeps improving ChatGPT's responses to sensitive situations with mental health experts. OpenAI has said roughly one million users a week show signs of suicidal thinking or psychosis in their conversations. (Read More)
🫖 TEA For Thought: "The truth is, no one knows how neural networks or LLMs truly work. It's a black box, just like how humans have no idea how the thoughts in our brains work…"
2. 🧠 Writer Unveils Enterprise Brain
WRITER announced Enterprise Brain on September 9, pitching it as the first universal context layer to power the agentic customer journey. It is a governed layer that holds a company's brand guidelines, institutional knowledge, decision logic, and live business data in one place, then serves that context to both employees and AI agents in real time. It ships with Agent Memory, a team-level memory that captures decisions and preferences out of ongoing conversations into a shared organizational record, plus a brand and compliance layer that encodes messaging, visual identity, and legal guardrails by industry and region. Distribution runs through WRITER Meet (Zoom, Microsoft Teams, Google Meet), WRITER for Slack, and a Microsoft Teams app due in the coming weeks. Co-founder and CEO May Habib announced it alongside Qualcomm CMO Don McGuire and Labcorp CMO Amy Summy, with Clorox and Vanguard also named as customers. Enterprise Brain is in early access for select customers, while Agent Memory and the Slack integration are generally available. (Read More)
🫖 TEA For Thought: "This is dope. It is one of the few brain-as-a-service products I have seen. The company brain is where the future of the org head is."
3. 🔍 The Eval You Cannot Trust Is Your Own
Google DeepMind, working with the Singapore AI Safety Institute, OpenMined, AVERI, and MLCommons, ran what it calls the world's first double-blind evaluation of a proprietary frontier-class model on August 27, 2026. The evaluator never saw the weights and Google never saw the test prompts, because the run happened inside Confidential Space, a trusted execution environment where the operator is cryptographically excluded from the inputs. The model under test was Gemini Flash Lite, not a flagship, which makes it a proof of plumbing rather than a solved problem. Writing in Ground Truth, TPM Sumaiya argues the contamination this fixes is a leaderboard problem, while an enterprise eval set leaks a different way: through workspace permissions during a POC, when the vendor gets Viewer access to the very questions being scored. She cites Capgemini's survey of 1,500 executives across 14 countries finding 2% of organizations running AI agents at scale, and Gartner's forecast that over 40% of agentic AI projects get canceled by the end of 2027. (Read More)
🫖 TEA For Thought: "This is quite similar to how finance audit works. The accountant's role has to be separated from the auditor's role. Complete separation is how work is being examined. One executes and one examines. You can't be both the gamer and the judge."
4. 🚧 OpenAI Pauses $200 Pro Sign-Ups as Astra Demand Spikes
OpenAI temporarily disabled new subscriptions to its $200-per-month Pro plan, citing infrastructure strain from demand for Astra, the model it launched on September 3. Product leader Thibault Sottiaux, who runs core products including Codex and ChatGPT, announced the pause on X, saying Pro puts the most strain on OpenAI's systems and that the company "wanted to take the smallest step that allows us to continue giving the broadest access possible." The API and the lower-cost Go and Plus plans stay open. Sottiaux had warned the day before that demand for Astra was "really unprecedented" and that OpenAI was "pulling all the levers possible to sustain the demand," with priority on keeping service intact for existing users. The company has not said how long sign-ups stay disabled or how many people are signing up daily. It raised Codex usage limits as recently as last month, which suggests the strain is recent. Astra is rolling out across Pro, Plus, Enterprise, and Business accounts, and OpenAI has billed it as the start of the "AGI era." (Read More)
🫖 TEA For Thought: "Expected. I used three resets already this week, and the last one has less than 15% left."
5. 💹 China's AI Chipmakers Raise Prices as HBM Runs Short
Chinese AI chipmakers are pushing prices up as high-bandwidth memory (HBM) gets scarce. Huawei's Ascend 950DT is now quoted above 250,000 yuan ($37,255), 20% to 50% higher than quotes given to customers two months ago, with availability expected in Q4 2026. The Ascend 950PR went from about 60,000 yuan in early 2026 to more than 80,000, roughly 30%, and the 910C from about 90,000 to more than 110,000, roughly 22%. Cambricon's next-generation 690, not yet formally launched, is being indicated 20% to 30% above levels quoted two months ago. MetaX and Iluvatar CoreX are affected too. Advanced HBM is dominated by SK Hynix, Samsung, and Micron, and U.S. export controls from December 2024 have pushed Chinese buyers toward grey-market channels where HBM typically costs several times what buyers outside China pay. Iluvatar CoreX reportedly doubled GPU shipments to ByteDance to 100,000 units this year as compute constraints tightened. (Read More)
🫖 TEA For Thought: "Interesting… the price hike might just be the beginning."
🛠️ Skill of the Day
The Blind Scorecard: write the test before the demo, so a good score still means something.
You are an evaluation designer helping me decide whether to adopt a new tool, vendor, or AI feature. I will describe it and the job it has to do. Before I see any demo, build me a sealed scorecard.
Tool or feature: [WHAT IT IS]
The job it must actually do: [THE REAL TASK, IN MY OWN WORDS]
Who gets hurt if it fails: [PEOPLE OR PROCESS]
Produce:
1. Eight to twelve concrete test cases drawn from my real work, never from the vendor's examples. Include at least two ordinary cases, two edge cases, and one case I would be embarrassed to get wrong.
2. A pass threshold for each case, plus one speed or cost limit, so a correct-but-too-slow answer still fails.
3. One line naming the result that would make me walk away.
4. A short warning listing the places this scorecard could leak to the vendor before I finish scoring.
Do not soften the thresholds to be helpful. If my job description is too vague to score, ask me up to three questions first.Paste into ChatGPT, Claude, or your tool of choice. Fill in the brackets, keep the finished scorecard somewhere the vendor cannot reach, and grade the demo against it instead of against your gut.
TEAHEE Moment
Stay sharp, stay informed. See you Monday.
If you enjoyed this TEA, follow along on social for more: