Austin covers six stories: AI labs testifying under oath before the New York City Council, OpenAI adding an invisible watermark to ChatGPT text in the EU, a Utah pilot letting an AI write acne prescriptions, former Groq engineers suing over Nvidia's roughly $20 billion deal, chip startup Etched reportedly drawing $40 to $50 billion offers, and Reflection AI's new open model Beam. Austin is skeptical of the New York hearing and the watermark, and upbeat about the Utah pilot, Etched, and US open models.
OpenAI, Anthropic, Google and Meta testify under oath to the NYC Council
Watch from 2:18
- What happened
- Executives from OpenAI, Anthropic, Google and Meta testified under oath before all 51 New York City Council members, sitting as a committee of the whole on Monday, after the council threatened subpoenas. Asked the chance of an AI catastrophe, OpenAI's Morgan Dwyer said she didn't know and that no level is acceptable; Speaker Julie Menon called that flippant. Anthropic's Logan Graham and Meta's Shane Cahill gave no number. Google's Alex Friend said no rigorous method exists yet. Former Google DeepMind researcher Alex Turner put the odds of an eventual AI takeover at roughly one in three. On liability, OpenAI, Anthropic and Meta gave no direct answer. Menon demanded written follow ups. Bills on the table include a whistleblower reward program, a right to sue over harm from AI agents, and independent testing.
- The claim
- The companies say they test their models for risks ranging from cyber attacks to loss of control. Google said that if something is illegal without AI, it's illegal with AI.
- Austin's take
- Austin thinks the hearing focused on the wrong question. He sees AI as a net benefit like the early internet, wants energy spent on helping people learn to use it, and doubts a city has meaningful jurisdiction over models built elsewhere. He worries local restrictions would mostly limit New Yorkers themselves.
- Why it matters
- Local governments are now pulling frontier labs into sworn testimony, following California's recent subpoenas and new AI laws. Liability for AI harm is still unanswered.
- What to do
- If you build or deploy AI agents in New York, watch the right to sue and independent testing bills, since they could shape your legal exposure.
- Not confirmed yet
- It is unclear whether city level bills can have any real effect on companies that build models outside New York City.
OpenAI adds an invisible watermark to ChatGPT and Codex text in the EU
Watch from 6:56
- What happened
- OpenAI said Monday that over the coming weeks, eligible ChatGPT and Codex output in the EU will carry a hidden statistical signal in its word choices. Readers can't see it; a detector can. API customers worldwide can switch it on, but it's off by default. OpenAI plans to open source the technology, and at first only approved researchers and organizations get the detector. ChatGPT joins Claude and Gemini in marking its own text.
- The claim
- OpenAI says this meets the EU AI Act's transparency rules for AI generated content. Its own tests found the mark in about 95% of 400 token passages and about 80% of 200 token passages at a 1% false positive rate. OpenAI also says swapping 25% of words for synonyms drops detection to 17%, and that a missing watermark does not prove human authorship.
- Austin's take
- Austin thinks text watermarking is close to pointless: AI writing is already easy to spot, and tools like Pangram exist. He sees watermarks as more important for deepfake video, voice and images, and views this as regulators moving slower than the market. He also called OpenAI slow to adopt safety measures.
- Why it matters
- Teachers, employers and platforms want a way to tell AI writing from human writing, but light editing defeats this watermark and its absence proves nothing.
- What to do
- Don't treat a missing watermark as proof a person wrote something. If you use the API and want marked output, you have to turn it on.
- Not confirmed yet
- Detection figures come from OpenAI's own tests.
Utah lets an AI write acne prescriptions without a doctor signing each one
Watch from 10:48
- What happened
- NOLA Health said Monday its NOLA Derm app can issue topical acne prescriptions to Utah adults under an agreement with the state's Office of Artificial Intelligence Policy, at a pilot price of $4.99 a month. Two physicians approve every prescription for at least the first 100 patients. Then the AI prescribes directly with weekly doctor review up to 500 patients, and after that doctors audit at least 10% each month. No pills and no severe acne. The waiver runs for one year. NOLA's medical advisor named flu and uncomplicated urinary tract infections as possible next steps.
- The claim
- Founder Luis Winis said that until now a licensed clinician signed every prescription and that the AI can now make the decision itself. NOLA says the AI picks among preapproved treatment protocols rather than making free form decisions.
- Austin's take
- Austin is excited to see AI in direct clinical practice and likes the staged oversight. He expects it to free doctors for more serious cases. His concern is whether it raises prescribing rates as it spreads to other conditions, and whether it lets doctors with financial incentives scale prescriptions without proportional work. Overall he expects a net positive and praised Utah for having an AI policy office.
- Why it matters
- This is a template for removing the doctor from routine prescribing. The open question is who signs off as the list of conditions grows.
- What to do
- If you're in Utah and try it, know it covers only topical acne treatment for adults. Elsewhere, watch whether other states copy this model or expand it to flu and UTIs.
Former Groq engineers sue over Nvidia's roughly $20 billion deal
Watch from 15:53
- What happened
- Former Groq engineers Benjamin Sarah Bryn and Joshua Rubin sued Groq's board Monday in Delaware's Court of Chancery, naming founder Jonathan Ross and president Sunny Madra. In December 2025, Nvidia paid about $17 billion in licensing fees plus roughly $3 billion in stock bonuses to bring over about 200 Groq engineers, a deal described as non exclusive. The suit says the board skipped a stockholder vote required by Delaware law and left Groq a hollowed out shell, and that board conflicts left common stockholders billions worse off. The Justice Department is already looking at the deal, per The New York Times in September. Both plaintiffs left Groq before the deal but kept their stock. Groq and Nvidia hadn't responded in the reports.
- Austin's take
- Austin sees this as the standard license and hire playbook, where preferred investors get paid first and common holders can lose out. He suspects the plaintiffs may be opportunistic after seeing a huge payday, and wished more luck to Groq and Nvidia for pushing technology forward.
- Why it matters
- Big tech uses license and hire deals to get startups' tech and talent without technically buying them. If a court requires a shareholder vote, that playbook gets much harder.
- What to do
- If you hold common stock in a startup, learn how a license and hire deal would treat your shares versus preferred investors.
- Not confirmed yet
- These are allegations from the complaint as reported by the Financial Times and law.com. Austin described the license as exclusive while the report called it non exclusive, and he said he doesn't know how much stock the plaintiffs held.
Chip startup Etched reportedly fields offers at $40 to $50 billion
Watch from 20:53
- What happened
- TechCrunch reported Monday, citing sources, that established investors' offers value Etched at about $40 billion and lesser known offers at about $50 billion. Etched was valued at $10 billion after a $300 million round led by Sequoia, and at $21 billion in September when Jane Street led a $700 million round. A July report said Etched had $11 billion in customer orders, and Jane Street is both investor and customer. No round has closed.
- The claim
- Etched says its chips process tokens faster and cheaper than Nvidia's.
- Austin's take
- Austin is enthusiastic, calls it proof that money is chasing alternatives to Nvidia, and wonders if a recent Invest Like the Best appearance helped. He speculates Nvidia, which spent about $20 billion on Groq, might consider buying Etched, or that Etched may prefer to stay independent.
- Why it matters
- A rapid jump in valuation shows how hard investors are betting on chips that could compete with Nvidia for running AI.
- What to do
- Treat the numbers as offers, not a done deal. Watch for a closed round or any acquisition interest.
- Not confirmed yet
- Based on unnamed sources. The report doesn't say who is bidding or whether the money goes to the company directly.
Reflection AI unveils Beam, a US open model aimed at Chinese rivals
Watch from 26:34
- What happened
- Reflection AI unveiled Beam, a mixture of experts model with 501 billion total parameters and 23 billion active per token, trained on 23.8 trillion tokens. Early access opened Monday. Weights are planned for later in October under the Apache 2.0 license after red teaming. Reflection was recently valued at $25 billion.
- The claim
- Reflection says Beam beats GLM 5.2 on some tasks using a third to a quarter of the hardware, approaches Qwen 3.8 Max, still trails frontier models like Claude, and is the first open model from a US startup to match the best Chinese models.
- Austin's take
- Austin calls this a big deal for US open source, arguing open models level the playing field and that relying on Chinese open models risks lock in. He explains mixture of experts as using only the relevant part of the model for each question, which is more efficient, and suggests it might run slowly on a personal laptop.
- Why it matters
- Open weight AI has been dominated by China. This is the strongest US attempt to change that, if the weights actually ship.
- What to do
- Wait for the weights and independent benchmarks before relying on it. If you use Chinese open models, keep an eye on this as an alternative.
- Not confirmed yet
- All performance claims are Reflection's own benchmarks; the weights aren't downloadable yet, so third party testing hasn't happened.
