Nvidia's board has launched an independent review after CEO Jensen Huang admitted an affair with a woman US intelligence reportedly believes is a Chinese spy. Elsewhere, a startup claims an unreleased Anthropic model cracked an 18-year-old maths conjecture, and Anthropic is paying billions to jump the queue for power generators. Further down, AI agents are running three small businesses, and Anthropic has released Claude Fable 5.2.

IN THIS ISSUE · 5 SECTIONS

 1  Major AI News: Nvidia's spy-claims review, a startup's AI maths proof, Anthropic's $3.4bn generator deal.

 2  Fun AI Topics: AI judges for the 2028 Olympics, Claude agents that tidy their own memory, the creator AI labs want to test their chatbots.

 3  AI Tools: Dageno AI 2.0, Higgsfield Miracle, OpenCode Browse, Sakutto, OpenArt Product Ads.

 4  Expert Prompt: turn a long email thread into a one-minute brief.

 5  Trending: AI agents run three small businesses, an agent Siri, Anthropic's Claude Fable 5.2.

LEAD STORY

Nvidia's board has ordered an independent review into reports that CEO Jensen Huang was targeted by a foreign intelligence operation. Huang admitted to staff that he had an affair with Alice Huang, a woman US officials reportedly believe is a Chinese spy. She is a former executive at a data firm previously accused of smuggling Nvidia chips into China. China denies the allegations, and the board says it has no reason to believe company secrets or national security were compromised. Congress has requested a closed-door briefing.

Why it matters: Nvidia supplies the computing backbone for most leading AI models, so the outcome could ripple far beyond one boardroom.

ALSO TODAY

Physical Superintelligence says a research system built on copies of an unreleased Anthropic model produced a full proof of the Aaronson-Ambainis conjecture, open since 2008, then improved it twice. The result concerns where quantum computers truly outperform classical ones. A researcher set up the run and reviewed the write-up, and the proof was checked with Lean, which verifies each logical step. Some experts back the findings, but wider acceptance is still needed.

Why it matters: It is a concrete example of AI producing research a machine can check line by line.

ALSO TODAY

Anthropic reportedly agreed to pay about $3.4bn for 1.7 gigawatts of gas-fired generators, enough for roughly 1.4 million homes. The supplier ordered the Caterpillar units two years ago but needed funding to finish paying for them, letting Anthropic take over the orders and skip multi-year wait times. The report comes from the Financial Times, and neither company has commented. Anthropic also reported Q3 revenue of $16.96bn, up from $11.56bn in Q2.

Why it matters: Power is now as hard to get as chips, and the richest players can buy their way to the front.

World Gymnastics is discussing bringing Fujitsu's AI judging system to the 2028 Los Angeles Games after testing it at the World Championships. Four cameras, or eight for floor exercise, build a 3D model of each routine, measure joint angles to the degree, and deliver analysis within 15 seconds of landing. It can catch errors the human eye misses, such as a leg that is 179 degrees rather than straight. But gymnastics also scores artistry, so humans stay essential.

Why it is interesting: It shows AI being trusted with precise quantitative calls while leaving judgement and style to people.

Source: World Gymnastics

Anthropic has described a form of AI downtime in which Claude agents inspect, correct, and reorganise their own memory files so they can build on past work across sessions. The agents compare saved notes against logs to spot mistakes. In Pokemon tests, one newer model earned four badges using only 10 memory files, while an older model earned one despite using 413. Another model finished the game entirely.

Why it is interesting: It points to agents that remember ongoing jobs and improve over time, instead of starting fresh each session.

Source: Anthropic

Creator Matt Husk says AI labs started approaching him about sponsorships after his videos exposed chatbot quirks, including a bot that timed a few seconds as about 10 minutes and another that struggled to negotiate. He says OpenAI, Anthropic and xAI all contacted him, and that xAI offered him a Tesla for a week in exchange for exclusive content. He turned down OpenAI's offers.

Why it is interesting: It hints that negotiation and creative dealmaking could become real benchmarks for AI models.

Source: WIRED

[MARKETING] Dageno AI 2.0 - Shows whether ChatGPT, Gemini, Perplexity and seven other AI assistants recommend your brand or a competitor, and which sources shape those answers. Market lookups are free. dageno.ai (product page)

[VIDEO] Higgsfield Miracle - Private-beta AI video tool that splits footage into editable layers so you can recolour objects or change speeds. x.com (launch announcement)

[CODING] OpenCode Browse - Lets coding agents control your real browser, with its tabs and logins, to test authenticated flows. opencode.ai (product page)

[TRANSLATION] Sakutto - Free Chrome extension that translates pages entirely on-device using a compact local model, no account needed. Chrome Web Store (extension page)

[ADS] OpenArt Product Ads - Turns a product photo or page link into a full video advert with audio, hook, and consistent branding. openart.ai (product page)

Context: Use this when a long email thread needs to be distilled for a busy executive. It asks the AI to separate facts from opinions and surface what actually needs a decision.

Act as a senior chief of staff and executive communications adviser. I will paste a long email thread. First I will tell you the audience, the topic, my role, the length and tone I want, and my key concern. Convert the thread into a summary an executive can scan in under one minute. Clearly distinguish facts, decisions, opinions, unresolved questions, and disagreements. Prioritise business impact, deadlines, ownership, dependencies, financial implications, customer effects, and leadership decisions. Remove repetition and low-level detail. Do not invent context, resolve ambiguities, or present tentative decisions as final. Output these sections: executive summary, key decisions, open issues, risks or blockers, actions with owners and deadlines, leadership asks, and missing information. Here is the thread:

Example use case: A project lead pastes a 40-message procurement thread and gets a one-page brief showing the two decisions already made, the budget question still open, and who owes what by when.

Skyfall AI has put Steve-001, a team of AI agents, in charge of three small online businesses it has bought since July. A human owner sets the direction and approves the biggest decisions, while a set of rules decides what the agents can do alone, such as refunds of $30 or less. One business, a social media tool, grew monthly revenue from $3,649 to $8,307 in about three months. A Shopify app is about 3% below where it was when Skyfall bought it, and the third business is too new to judge. The figures are Skyfall's own and haven't been independently checked.

Why it is important: It is an early test of AI agents running a real business day to day, with a human still signing off on the big calls.

Source: The Next Web

Apple's CEO is reportedly pressing its software chief to move past incremental upgrades and redesign Siri as an autonomous agent that competes with rivals like Meta Muse, OpenAI Dots, and Claude Duet. The revamped assistant could arrive as soon as next year, though Apple has not committed to a timeline. Separately, Apple has reportedly cut engineers' daily budgets for internal AI tools to $200.

Why it is important: A credible agentic Siri would put capable AI assistants in front of a vast installed base of iPhone users.

Source: Bloomberg

Anthropic has released Claude Fable 5.2 and Mythos 5.2, the same underlying model with different safeguards. Fable is open to all users, while Mythos is limited to approved cybersecurity and life sciences organisations. Anthropic says output is up to 30% faster than the 5.1 versions, and pricing is $8 per million input tokens and $40 per million output tokens. In one test, more than 1,000 agents worked for eight days to build an operating system that boots on a laptop. Anthropic still recommends Opus 5.5 for most everyday coding and knowledge work.

Why it is important: It shows AI agents handling a single job for days, though Anthropic still points routine work to Opus 5.5.

Source: Anthropic

That's it for today's Daily Pulse. Forward this to one person who wants to stay ahead of AI. See you in the next one. - Nicolas