In partnership with |  |
| | Good morning. It's Monday, November 9th. | | | | The voice-to-text app that built its own speech model | | Most dictation apps are a wrapper around someone else's speech model. Wispr Flow trained its own. | It's called Canto, and it was built for the way people actually talk to AI: thinking out loud, changing direction mid-sentence, half-whispering in an open office. | Wispr tested it on 10 hours of real dictations from more than 2,300 speakers. Canto had the lowest word error rate of every model tested, including models from Google, OpenAI, AssemblyAI and Deepgram. That means your prompts to ChatGPT, Claude and Cursor come through the way you said them. | Talk naturally in any app and Wispr Flow types clean, formatted text, 4x faster than typing. | Try Wispr Flow free |
|
| | | Jensen Huang admits to affair, investigation ensues | Nvidia's board opened an independent investigation on Sunday into reporting that CEO Jensen Huang "has been the target of a foreign intelligence effort." A committee of independent directors is running it with outside counsel. The board says it has no reason to believe national security or company secrets were compromised, that Huang is cooperating, and that it retains confidence in him. | Huang told employees "I was unfaithful to my wife" and "there is no excuse for what I did," in a staff memo acknowledging a relationship with Alice Huang (no relation), a senior executive at the Singapore data center firm Megaspeed. US officials reportedly believe China's Ministry of State Security recruited her to push Huang toward lobbying Washington for looser export limits on advanced chips, with surveillance tracing the relationship to at least June 2024. The New York Times reports Secretary of State Marco Rubio oversaw a classified memo warning officials about it last spring. Treat the intelligence claims as reported suspicion, not established fact. | House Republicans asked for a closed-door briefing last week and have since called Huang to testify. President Trump called him "an honorable guy" and backed him staying in the job. Nvidia went into the weekend down 8.76% for the week and roughly $1.27 trillion below its October 6 peak of $239.24, its fifth straight losing week. At AI Day Seoul this morning, Nvidia declined to say whether congressional pressure could affect its Samsung and SK hynix partnerships or its GPU deliveries to Korea. | Read more. | | OpenAI raised $11 billion, not $30 billion | OpenAI closed an $11 billion first tranche on Friday at a $1.25 trillion valuation, per Bloomberg. The round may reach $20 billion over the coming weeks. It was originally shopped at around $30 billion or more at $1.4 trillion, and was cut back as AI stocks sold off. No investors were named and OpenAI did not comment. | A company targeting $30 billion that banks $11 billion, at a valuation $150 billion below where it started asking, is not failing to raise. It is discovering the current clearing price. Quartz framed it as falling short, which is fair, and it is also the first time in two years that OpenAI's number moved down on the way through. | We asked a week ago whether OpenAI was having trouble raising money. This is the answer, and the answer is "sort of." The money is there (terms just moved) | Read more. | | METR says Anthropic is improving models twice as fast, and most of the evidence is redacted | METR published the first high-access independent assessment of AI R&D acceleration inside a frontier lab. About 2.5 full-time staff worked on Anthropic devices from August 12 to October 29, with internal benchmark data, training compute metrics, headcount, inference spend and roughly 15 staff interviews. The headline: Anthropic's internal capabilities index grew at 1.9x its pre-October 2025 rate, 95% confidence interval 1.5x to 2.8x, with a breakpoint at October 7, 2025. | Compute scaling did not accelerate, which METR says it is confident about, so the speedup has to come from somewhere else. METR's rough estimate is that about 80% of it is AI doing the work. Internal Slack polls put self-reported uplift as high as 12x for the most recent model, which METR calls very loose evidence and so should you. It also found no convincing bottleneck in algorithms, compute or data for the next twelve months, while noting its own view on that changed several times. | Anthropic held unilateral redaction rights and METR says it cut "a majority of the evidence we found most informative," including per-stage estimates, the compute breakdown, forecasts for future models and specific benchmark names. A planned verification of how compute is split between capabilities and safety was canceled midway by Anthropic. Include every developer's models rather than only Anthropic's and the acceleration estimate drops to 1.4x, because Anthropic was improving more slowly than the frontier before late 2025. METR's own October proposal for this kind of audit argued the public should not have to take the assessor's word for it. Here, largely, it does. | Read more. | | Anthropic quietly bought 1.7 gigawatts of gas generators | The Financial Times reported Sunday that Anthropic signed a $3.4 billion deal in late September for 1.7 gigawatts of large gas-fired generators. The twist is where they came from. San Antonio power supplier OnePWR had ordered them from Caterpillar about two years ago and needed money to complete the purchase. Anthropic had planned to buy power from OnePWR and instead took over the orders, the queue position, and the installation work. No acquisition, no staff moved. Both companies declined to comment. | Turbine and large-generator lead times are the binding constraint on new capacity, so buying someone else's two-year-old place in line beats ordering your own. It also means Anthropic now owns generation equipment rather than a power contract, which is a different kind of balance sheet. | Voters went against data centers in 40 of 41 ballot measures on Tuesday, and gas generators next to a data center are precisely what those measures were about. Anthropic is also mid-IPO. The roadshow was targeted for the week of November 16, and unnamed investors told the Wall Street Journal the $1.5 trillion target may need to come down toward $1 trillion, with one floating postponement outright. No price range has been set. We covered the original filing here. | Read more. | | Trump says he doesn't "mind guardrails" on AI, as long as the US isn't left in a "deep second position" (CBS News) Anthropic's IPO valuation could fall from $1.5 trillion toward $1 trillion after the AI rout, investors say (Yahoo Finance) DeepSeek says its Harness agent hit 1.54 million daily users and 92 trillion tokens a day (36Kr) OpenAI brought in security firm Hacktron to hunt vulnerabilities in its own systems and integrations (OpenAI) Nvidia, SK Telecom and Rebellions signed an MoU to put Korean NPUs in Nvidia racks via NVLink Fusion (The Korea Times) The Huang controversy has Korea worried about US chip curbs hitting Samsung and SK hynix (Korea JoongAng Daily) Two Chinese firms cleared to buy restricted Nvidia H200s appear to operate out of empty offices (The Verge) US computer and chip imports hit $86.9 billion in September, a $1.04 trillion annualized pace (24/7 Wall St.) AI companies priced more than $300 billion of stock and convertibles in 2026, with $100 billion more queued (The Information, via AI Weekly) 19.1% of US workers replaced by AI say their old employer asked them back, and half wanted a raise (Fortune) US professional and business services added 70,000 jobs in October, the biggest gain since July 2022 (The Wall Street Journal) Staff at Commerce's $10 million-a-year AI safety center were asked to report informally to a White House office (The Associated Press) A Philippine Senate bill would require six months' notice and roughly double severance for AI-driven layoffs (The Manila Times) Accel is reportedly leading a $5 billion round for Thinking Machines Lab at a $40 billion valuation (Tech Funding News) Agility Robotics went public on the Nasdaq, inviting a closer look at humanoid robot economics (Ars Technica) A 1,023-parameter direct weight edit lifted Qwen3.5 by 29.45 points on MATH-500 (Hugging Face) Inkling 2 Small would be the top-scoring non-Chinese open-weights model, once the weights actually ship (Artificial Analysis) METR also published its review of the automated R&D section of Anthropic's August risk report (METR) Ai2's StateBench tests whether models track state changes over long documents, and frontier models struggle (Ai2) Ai2 released ToyBlocks, an open simulation environment for robot learning and evaluation (Ai2) Catching up on Friday: OpenAI's first Critical-rated cyber model, and data centers losing 40 of 41 ballots (AI Breakfast)
| | | Video got cheaper. Gemini Omni 1.2 Flash stretches clips from 10 to 15 seconds, settable in one-second steps from 3, and drops 720p pricing 30% to $0.07 per second. The genuinely useful addition is multi-turn conversational editing, so you adjust a clip instead of regenerating it and hoping. Generally available in the Gemini API and Google Cloud, free for 15-second clips in Google Vids, rolling out in the Gemini app to AI Ultra subscribers. Outputs carry SynthID. | On the agent side, Perplexity launched Teammates, persistent agents that keep working between sessions, exclusive to the $200-a-month Max plan with mobile promised later. Google's Agent Development Kit hit 2.14.0 and graduated a large batch of auth, environment and toolset modules out of experimental, which matters more than it sounds if you have been pinning versions. | Models and weights: Thinking Machines' Inkling 2 Small is 594B total, 17B active, with a claimed jump in agentic coding and vision. It is closed-weight at launch, with full weights promised "in the coming weeks, pending the results of additional safety testing," so hold the applause until that lands. Mistral shipped Studio alongside its Le Chonk release, and Hugging Face put out Phonon TTS for on-device speech. Nvidia also detailed Vera CPU throughput for agent workloads, with vendor-reported numbers nobody has independently checked yet. | | Thank you for reading today's edition. | | Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter. | Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on X! | Thinking of starting your own newsletter? AI Breakfast readers who sign up with Beehiiv receive a 14-day free trial and 20% off for 3 months. |
|