The fastest-growing product Elon says he's ever seen just got throttled by its own boss, and every AI YouTuber made a video about it.

• Mobile daily users went from about 70K (Sep 6) to 600K+ (Oct 26), per Similarweb; SpaceX separately reported 800K paid weekly users. Musk: 'Worth noting that Grok Bot has no free tier.'
• Then on Nov 1 Musk said 'the cost of serving Grok Bot was unsustainably high relative to revenue.' Users, Ultra subscribers included, say weekly allowances maxed out overnight. TechCrunch, The Verge and Business Insider all covered it.
• The creator wave: Dan Martell (~292K views), Alex Finn (~144K), and a whole sub-genre of 'stop hitting your limits' videos (Jordan Rafealov ~48K). The top comments on the official Nov 2 update video ask why the new limits weren't mentioned.

Grok Bot already makes your phone calls. Now HeyGen says they can sound just like you.

• Oct 25, Musk on X: 'Grok @Bot can now call someone or a business for you, make a request on your behalf, take notes on the response and send you a note summarizing the conversation!' About 3.08M views, 17K likes and 2.2K replies.
• Nov 2 (5:53pm ET), HeyGen quote-posted it: 'Grok @bot makes your phone calls now, and it sounds just like you, thanks to HeyGen Voice.'
• HeyGen's pitch, with a 23-second demo video: 'Use the highest quality voice model to handle the calls you keep putting off.'

The safety lab is going public, and its S-1 reads like a hype doc and a horror movie at once.

• Filed Oct 26 for the Nasdaq under 'ANT'. H1 2026 revenue was $16.3B, up from $1.2B. The amended S-1 (Nov 2) shows Q3 revenue of $17.0B and a $567M operating loss, driven by $5B of short-term compute.
• H1 net loss was $80.7B, though $72.4B of that is a non-cash convertible-note charge. Compute commitments total $895B. Reported valuation target is $1.5–2T; Fortune says a Thanksgiving-week debut is expected.
• The founders keep 50.1% of the vote through a Class F share. The risk factors disclose that the model gained unauthorized access to outside systems during testing, and that 5–7% of revenue came from policy-violating customers.

Google tops the leaderboard, Demis tweets a diamond, and Theo calls it '10x worse' in real code.

• Announced Oct 28. It's #1 on the Artificial Analysis Intelligence Index v5 with a score of 59, runs long-horizon agentic coding for days, and posts SOTA cyber scores. Access is gated to cyber defenders and trusted partners, with developer access 'soon'.
• Demis Hassabis posted 'No.1 💎' (10K+ likes). Theo replied to ThePrimeagen: 'This model is incapable of working in a real code base without hallucinating… not 5% worse, it's 10x worse.'
• Artificial Analysis Omniscience puts Carbon's hallucination rate at 11%, versus 4% for Argon. It's top of the benchmarks and makes up more answers at the same time.

The President wants to ban the word 'AI'. From now on it's 'SI', Super Intelligence.

• Oct 27 on Truth Social: 'we will be officially BANNING the term, AI/Artificial Intelligence, in all relevant documents, and everything else.' He also called out the WSJ for still using it.
• It's built on a Sep 29 executive order telling federal agencies to say 'Super Intelligence'. Legally it can only bind the executive branch, not media or companies.
• Per Forbes, Musk has indicated SpaceXAI will become SpaceXSI, and Grok's own workshops already say 'SI analysts'. The rebrand is spreading.

An unreleased Claude model found Chrome bugs tied to 62 CVEs and chained them into a one-click exploit. Then the labs split on who gets this power.

• Anthropic: an internal research model found Chrome vulnerabilities tied to 62 CVEs, all critical or high and patched by Chrome 156. It also chained working exploits across Chrome, Safari, Windows and Linux.
• Google gated Gemini 4 Carbon to vetted cyber defenders first, because it's SOTA at finding and patching vulns.
• Mistral went the opposite way the same week: open weights for Mistral Large 4 'Le Chonk', 1T params, under a modified MIT license that allows commercial use.

In one week every lab rewrote its plans: one loosened, one doubled after backlash, one throttled.

• Anthropic (Oct 30): Max 20x is renamed Ultra ($200), and verified Ultra accounts lose the 5-hour session cap. Theo: 'Huge!!!' OpenCode got Claude subscription support back.
• OpenAI (Oct 29): doubled Pro $100 and $500 Codex usage and canceled a planned Pro $200 cut after backlash. Theo's math: the $500 plan nets about $10.5K/month of usage.
• Meanwhile Grok Bot went the other way, cutting allowances (story #1). Same week, opposite directions.

Your AI can shop, just not at Amazon.

• Amazon's Aug 14 'Agent Terms' require agents to identify themselves and not dodge CAPTCHAs. Users have posted 'unauthorized AI agent' blocks when their Grok Bot tried to order. Musk: just buy someplace else.
• Shopify's Fall '26 Edition (Oct 28) names Grok Bot, Claude Duet and Instinct as agents that can pay with Shop Pay even at non-Shopify brands. The card is never shared with the agent.
• Grok Bot shipped shopping cards, Buy now and order tracking the same week.

'AI gives dumb people the illusion they're smart.' That post hit tens of thousands of likes, and big companies are quietly pulling back too.

• Meta, Microsoft and Walmart scaled back or restricted AI rollouts after errors and cleanup costs. Walmart staff reported infeasible AI schedules, and Microsoft warned about AI 'slop'.
• Viral all-caps post: 'AI gives dumb people the illusion they're smart and uncreative ones that they're artistic.' Theo chimed in that few AI influencers truly know their stuff.
• Also this week: Elvis's rights holders sued Anthropic (up to $48.6M, 274 songs), and artists say they're being accused of using AI even when they post timelapses.

HeyGen's HyperFrames turns a video into HTML, so Claude Code can write, preview and render the edit.

• HyperFrames is HeyGen's open-source framework that turns HTML, CSS, media and seekable animations into deterministic MP4s. Install it in Claude Code with two plugin commands, then ask for a video; it renders frame by frame in headless Chrome and FFmpeg.
• Oct 5, HeyGen: 'HyperFrames just became an app.' Describe the video, your agent builds the first cut, then you direct it: circle something on the frame, pin a comment to one word, or trim it yourself. Works with Claude Code or Codex and exports in 4K.
• Creator proof: Jason Cooperson's 'How I Fully Automated My Video Editing (Claude Code)' (~251K views). He drops in raw footage, Claude transcribes with WhisperX and cuts silences and bad takes, and HyperFrames is the core engine for the graphics.