In partnership with
In Todayโs Issue:
๐ฌ GPT-6 Astra corrects the physicists
๐ต๏ธ AI agents dodge Alibaba's bot defenses
โ Google's Fitbit watches for insulin resistance
โจ And more AI goodnessโฆ
โก The Signal
AI is starting to pay its own way, and to find its own way.
Anthropic told investors it took in $16.8 billion last quarter and generated $1.70 in revenue for every dollar it spent on compute, the clearest sign yet that a frontier lab can turn a profit, at least on an adjusted basis. The rest of today's issue shows what AI now does with less supervision. GPT-6 Astra solved 49 of 50 fresh physics problems and caught mistakes in the experts' own answers. A fleet of AI agents on Tencent Cloud kept finding new routes around Alibaba's bot defenses. A jailbroken PS5 runs a 27-billion-parameter model on its own chip. And McKinsey now lets an AI play the demanding client in job interviews. More of the work happens without a human at every step, and one lab is already getting paid handsomely for it.
All the best,
Kim Isenberg
๐ต๏ธ Researchers Tracked an AI Agent Fleet Scraping Alibaba's Maps
Independent researchers say a fleet of AI agents running on Tencent Cloud made more than 2,600 attempts to pull data from Amap, Alibaba's mapping service, between September 28 and October 8. The agents were after mundane facts, such as which entrance most visitors use at a park or hospital. When Amap's bot defenses blocked them, they routed their code through a public URL-scanning site, tried API keys leaked in public code and probed what appear to be non-production servers. The researchers say they cannot tell which AI model was behind it; a source close to Tencent called it a temporary test of third-party models.
๐ tl;dr: AI agents sent to collect simple map data found their own ways around a major site's bot defenses, a real-world preview of how agents behave when they are told no.
๐ฌ GPT-6 Astra Caught Errors in the Physicists' Answer Key
On TheoPhysEval, a new benchmark of 50 unpublished, research-level physics problems, OpenAI's GPT-6 Astra solved 49, and several answers first marked wrong turned out to be right. The authors rechecked those cases with the experts who wrote the problems and found subtle errors in the original solutions; once corrected, Astra's answers held up. A world expert would typically need about a day per problem; Astra, working in Codex with literature search and coding tools, averaged under $3 each.
๐ tl;dr: An AI agent now solves nearly every problem on a fresh, research-grade physics test, and is accurate enough to catch mistakes in the experts' own answers.
๐ผ McKinsey Now Makes Recruits Argue With an AI Client
McKinsey is adding an AI avatar to some entry-level consulting interviews this fall, and candidates will have to defend their recommendations to it, the Wall Street Journal reports. After the usual case challenge, applicants spend about 10 minutes with the avatar, which plays a client such as a mayor deciding where to build a stadium. The AI does not grade anyone; partners or assessors in the room watch how candidates handle it. The pilot covers undergraduates applying for business-analyst roles in some North American offices. "What we're looking for is how do you respond to changing stimuli?" said Saleem Zafar, a partner who helped design it.
๐ tl;dr: Some McKinsey candidates will now have to pitch their ideas to an AI playing a demanding client, a test built to show how they think on their feet.
The AI notetaker that gets the hard words right
Your AI is only as good as what you feed it. Feed it a transcript with your product names, acronyms, and numbers spelled wrong, and you spend the afternoon hand-fixing the output.
Wispr Flow Notetaker uses your dictionary and calendar, so product names, acronyms, numbers, and uncommon names come out spelled right. It pops up when your meeting starts and captures Zoom, Google Meet, Teams, and Slack huddles with one click, no bot joining the call.
Then your meetings show up inside Claude or ChatGPT through the built-in connector. No copy paste. Try Notetaker free on Mac and Windows.
๐ฌ Watch This
Reuters' Carmel Crimmins explains the fault lines that opened at this week's International AI Summit in Dublin. Europe's tech chief argued that existing EU regulation already gives the region ample tools to handle rogue AI agents, and parts of Silicon Valley are pushing back. It takes under two minutes, and it pairs well with today's Amap story: the agents are already improvising in the wild while Brussels insists its current rulebook is enough.
Anthropic's update makes it quicker to start coding sessions, split work across several and see what each one is doing; Thariq calls it "almost like a v2." It shipped on October 14, the same day Anthropic presented its $16.8 billion quarter to investors.
A jailbroken PlayStation 5 is now running a 27-billion-parameter AI model on its own graphics chip at about 20 tokens per second, according to the developers' benchmarks. Mert Cobanov first ported llama.cpp to the console's CPU on October 1; a community fork has since squeezed a heavily compressed Qwen 3.8 27B into the PS5's memory and onto its GPU.
(PS5LM Plus dashboard, GitHub)
Anthropic's $16.8 Billion Quarter Shows AI Can Pay Its Way
The Takeaway
๐ Anthropic told prospective investors its third-quarter revenue reached $16.8 billion, up about 45% from $11.6 billion in the second quarter, according to people familiar with the matter.
๐ The company said it was profitable on an adjusted basis and cash-flow positive, earning $1.70 in revenue for every dollar it spent on computing power, the Financial Times reports.
๐ A single quarter now brings in about 3.6 times what Anthropic made in all of 2025 (about $4.6 billion).
๐ IPO marketing could begin as soon as the week of November 9, and prospective investors have floated a valuation of $1.8 trillion to $2 trillion.
The biggest open question about frontier AI has been whether anyone can sell it for more than it costs to run. Anthropic has just given investors its answer. At a presentation at its San Francisco headquarters on Wednesday, the Claude maker said its third-quarter revenue reached $16.8 billion, up from $11.6 billion in the second quarter and $4.7 billion in the first, according to people familiar with the matter cited by Bloomberg and CNBC. It also said the quarter was profitable on an adjusted basis.
(Chart: Superintelligence, based on figures reported by CNBC and Bloomberg)
The number behind the shift is efficiency. According to the Financial Times, Anthropic told investors it was cash-flow positive and earned $1.70 in revenue for every dollar it spent on computing power during the quarter. Compute is usually a frontier lab's biggest bill, and last year the math looked very different: in 2025 Anthropic brought in about $4.6 billion while posting a net loss of nearly $42 billion, according to documents previously reviewed by Bloomberg. Investors now expect annualized revenue of $110 billion by year-end, The Wall Street Journal reports.
(REUTERS)
Those numbers set up what could be one of the largest stock-market debuts ever. Anthropic is preparing to start marketing its IPO as soon as the week of November 9, aiming to trade before Thanksgiving and no later than year-end, and prospective investors have floated a valuation of $1.8 trillion to $2 trillion. The figures are unaudited and were relayed by people familiar with the presentation; the public filing will show the full accounts. On these numbers, frontier AI is starting to pay for itself.
Why it matters: Every frontier lab has been financed on the promise that AI would one day earn back what it costs in compute. Anthropic's $1.70 per compute dollar is the clearest number yet showing that promise being met, just as public investors get their first chance to buy in.
Sources:
Most AI Tools Weren't Built for This
B2B customer issues move across teams and systems, not through a single chat window. A new Harvard Business Review Analytic Services briefing paper, sponsored by Front, breaks down where AI tools fall short in B2B service and what to ask before you invest.
(TheoPhysEval, Figure 3, arXiv)
The chart: This is the full scoreboard behind Astra's physics result in today's news. TheoPhysEval gave five AI agents 50 unpublished, research-level theoretical physics problems, plus literature search and coding tools, and let them work without hints. GPT-6 Astra in Codex passed 98%, Claude Opus 5.5 in Claude Code 92%, Gemini 3.8 Flash 68%, and the open-weight models DeepSeek-V4.1-Flash and GLM 5.3 passed 54% and 42%.
The lesson: Astra has company at the top: Claude Opus 5.5 solved just three fewer problems, so two frontier agents now clear more than nine in ten fresh, research-grade physics problems. The open-weight models solve about half or fewer, trailing Astra by 44 and 56 points.
The caveat: Without web search and background papers, Astra and Opus still scored 92%, which the authors read as a sign that the agents have memorized much of the literature. And solving a well-posed problem is only part of research: the authors observe that agents do not yet show the same "taste" as physicists in choosing which problems are worth solving.
โ Google's $180 Fitbit Now Watches for Insulin Resistance
โก Bottom line
Google's new Fitbit Edge brings AI-estimated insulin-resistance and blood-pressure trends to a $180 tracker, no blood test required.
๐ก Why it matters
Insulin resistance usually causes no symptoms for years while it raises the risk of type 2 diabetes and heart disease.
๐ What it means
Metabolic monitoring that once needed lab work is moving onto a cheap everyday wearable, with Gemini explaining what the trends mean.
On Monday, Google put an insulin-resistance tracker on a $179.99 wristband. Insulin resistance is one of the quiet engines of aging: the body gradually stops responding well to insulin, usually without symptoms, while the risk of type 2 diabetes and heart disease climbs. The Fitbit Edge, announced October 12 and in stores October 27, sits between the screenless Fitbit Air and the Pixel Watch and carries Google's Health Guardian features.
(Google, via Droid Life)
It never draws blood. Instead, it passively analyzes weeks of heart rate, movement, skin temperature and electrodermal activity (tiny changes in how well the skin conducts electricity) to estimate shifts in insulin resistance. If it sees a consistent rise, users get a monthly notification suggesting they review their habits or try a continuous glucose monitor. A second feature follows blood pressure trends using WavesFM, a foundation model Google trained on billions of minutes of raw heart and movement data.
(Google)
Those signals feed Google's Gemini-powered Health Coach, which explains the shifts and suggests changes. In the US, users can also import clinical lab results from Function and glucose readings from Abbott's Lingo sensor, putting wristband estimates next to real bloodwork.
Google is explicit about the limits: the feature does not measure glucose and is not meant to diagnose or manage diabetes. Health Guardian and Health Coach also require a Google Health Premium subscription (three months come with the device), and availability varies by country. Still, a cheap band that checks for metabolic drift every day brings a core idea of longevity medicine to an ordinary wrist: catch decline early, between doctor's visits.