📊 72% of the world's new AI computing power since 2022 went to the US. China got about 10%

The gap keeps growing. America took about 60% of each year's new AI compute from 2022 to 2024, then 72% in 2025 and 78% in the first half of 2026.

The numbers come from Epoch AI, a research group that traced AI chips through customs records from over a dozen countries, counting each chip where it finally lands. China's share includes its home-made Huawei and Cambricon chips.

By Epoch's estimate, the US has imported AI chips worth about $288 billion since 2022. China: $50.5 billion.

@aipost 🏴

💻 OpenAI plans to buy AI chips from Google, the company behind Gemini

OpenAI, one of Nvidia's biggest customers, plans to buy Google's own AI chips, called TPUs, according to The Information and Bloomberg. Both cite a person familiar with the matter. Terms and the size of the deal weren't disclosed, and OpenAI declined to comment.

The two rivals keep doing business. OpenAI has rented Google's chips since 2025, and last week said its models would be available through Google Cloud.

Google began selling TPUs directly to some customers this year. Oracle will also rent them out in its cloud from 2027, and OpenAI is building its own chip with Broadcom.

@aipost 🏴

Texas wants a judge to approve every new ChatGPT model before Texans can use it

Texas Attorney General Ken Paxton sued OpenAI and Sam Altman last Friday. He asks the court to bar ChatGPT for minors in Texas and to block any new AI model for Texas users "unless an independent third-party evaluator, approved by the Court, has reviewed its safety guardrails and approved its release."

The petition says the free version of ChatGPT "has no gatekeeping or age verification mechanism whatsoever" and that tens of thousands of its Texas users are under 13. It accuses OpenAI of deceiving users about the chatbot's safety.

Texas is asking for $10,000 in civil penalties for each violation and a jury trial. Privacy advocates note the remedy would mean ID checks, which hand over even more personal data.

@aipost 🏴

⚡ Two weeks after Argon, Google has a new No. 1 AI model

Gemini 4 Carbon is Google's next frontier model, and it tops the Artificial Analysis Intelligence Index with 59 points. Claude Opus 5.5 has 58, GPT-6 Astra 57.

The biggest jump is in cybersecurity. Carbon scores 88.2% on SEC-bench Pro, a test of finding vulnerabilities. Google's Chrome team says it did over 10 times better than another lab's large frontier model at writing high-quality security patches.

It isn't first everywhere. On coding agents it ranks second, 59 points to 62 for Claude Code running Opus 5.5.

You can't use it yet. Some cyber defenders get access this week, then trusted partners, with "wider availability soon." The introductory price is $2 per million input tokens and $10 per million output, later $4 and $20.

@aipost 🏴

🤖 Claude now works inside Google Docs, Sheets and Slides, and setup takes about two minutes

Anthropic's add-on puts Claude in a sidebar next to your file. It reads what you have open, sees what you've selected and edits it in place. It's in beta on paid plans: Pro, Max, Team and Enterprise.

How to turn it on, per Anthropic's setup guide:
1. Find Claude in the Google Workspace Marketplace and click Install.
2. Choose your Google account and allow the permissions.
3. Reload an open file, then go to Extensions → Claude → Open Claude.
4. Sign in with your Claude account.

One install covers all three apps. If the button is greyed out, your work admin has to allow it.

By default Claude asks before every edit. Select text, cells or slides first to point it at them. Each change shows up in version history, so you can undo it. In Sheets it writes formulas and builds pivot tables and charts. Firefox isn't supported yet.

Would you let Claude touch the spreadsheet your job depends on?

@aipost 🏴

🎮 "Make it a LOT more fun," he told Claude. And the game got fun

Programmer Simon Willison says October was a turning point: Claude Opus 5.5 now builds games that are fun to play. Older models made things that only looked like games.

In a few weeks he vibe coded four browser games without reading the code. None took more than four hours of his own time.

• Tree Appreciator, a spot-the-difference game with two generated trees, came from one prompt in the regular Claude chat
• Emberwick, a fantasy RPG with a 15-minute story. Its ending left him "emotionally affected"
• Battle Drones, a Command & Conquer-style strategy game, now with 2–4 player multiplayer

His starter prompt works in both ChatGPT and Claude:
Build a mobile-friendly three.js game where...

If the result is dull, add Make it a LOT more fun

GPT-6 Astra can build games too, he says, but its "sense of fun doesn't match that of Opus 5.5."

@aipost 🏴

💬 Your iMessage group chat can now get its own AI agent

Instinct, the AI agent startup valued at $10 billion in September, opened group chats to every iMessage user on Tuesday. Add it to a thread and it works for the whole group, even for friends who never signed up.

The group gets its own Instinct. It only knows what's in that chat, not anyone's accounts or data.

When it needs something from you, it goes through your personal Instinct, which asks you first.

Early testers used it to plan a dinner party, organize a road trip and decorate an apartment, founder Noah Shinn says.

Group chats first went to early-access users on October 5. Now anyone on iMessage can add it.

@aipost 🏴

🤖 About 1 in 4 web pages now has AI-written text in it

In August 2026, about 24% of crawlable web pages contained AI-generated or AI-edited text, a new study of 5.1 million pages from 2019 to 2026 estimates. Before 2022 the share was near zero.

And it's speeding up: from 4.6 percentage points a year between early 2023 and early 2025 to 8.1 points a year since.

Most of the growth is on sites that few others link to. They account for about four-fifths of the increase, while the most linked-to pages are the least likely to have AI text.

English pages have more of it than other languages. Government and university sites (.gov, .edu) have far less than .com.

The team, led by researchers at the University of Washington and UC Berkeley, used several AI detectors and corrected for their false positives and false negatives.

@aipost 🏴

📢 Claude now has ready-made tools that check a lawyer's citations

Anthropic added a gallery of interactive legal tools, which it calls artifacts, to Claude for Legal. Each opens with an example, then you add your own documents and connectors.

Some of what's inside:
• Cite check flags quotes that don't match their sources, wrong page numbers and sources that don't support the claim
• Playbook deviation board flags contract clauses that differ from your company's standard terms
• Tabular review turns a set of contracts into one table, say severance or non-compete terms, with links to the exact language
• Regulatory horizon scan puts upcoming rules on a timeline, ranked by impact

Harvey, PwC, Goodwin and Foley & Lardner built some of the tools. Nine more are for access-to-justice work, like a rent ledger checker for eviction cases in D.C.

@aipost 🏴

🗣 Mark Cuban: "There will be the trillionaire in the basement"

On an a16z podcast, Cuban says big companies struggle with AI until their bosses understand how it works. That leaves an opening: "There's no better time to be an entrepreneur than right now."

He thinks AI still needs a person guiding it: "AI is not smart. It's book smart maybe, but it's not street smart. It's not entrepreneur smart."

His demo for the head of a consumer food company: take their number one product, a cookie, and ask ChatGPT to create a competitor. "And in about three minutes he was like, 'Oh, s**t.'"

@aipost 🏴

💳 Someone put a $59.7 million AI supercomputer on eBay. Free shipping

IT hardware seller Core 4 Solutions listed a decommissioned cluster of 300 server racks with 9,600 Nvidia A100 chips and 4,200 AMD processors. Price: $59,700,000, "or Best Offer." FedEx Ground shipping costs nothing.

That's about 1.4 petabytes of memory. The A100 is a 2020 chip, but Tom's Hardware calls it a real budget option: a single new Nvidia rack costs up to $6.5 million (Blackwell) or $7.8 million (Vera Rubin), with long waitlists.

Its rough math for the chips and memory alone comes to about $28 million.

@aipost 🏴

Ask Google's AI Mode who won the 2020 election and it may not say

Instead of naming Joe Biden, Google's AI search mode told users to check the Federal Election Commission's official results. Asked about 2016, it answered right away: Donald Trump.

Google's standalone Gemini app, ChatGPT and Claude all named the 2020 winner in tests by Ars Technica. AI Mode often gave the answer when asked a second time, and answered straight away when Ars checked from other countries via VPN.

Google says it's fixing it: "Our systems verify election responses using news articles or AP data for recent elections, and historic records for older ones. There was a technical gap with 2020 – same is true for 2022 Senate races."

The US midterms are on Nov. 3.

@aipost 🏴

😎 Google's first Android XR glasses are on sale for $1,099

XREAL AURA weighs under 95 grams and puts virtual screens in front of your eyes, while the clear lenses let you see the room around you. The heavy work happens in a compute puck on a cable, which also holds the battery.

Gemini is built in and understands what you're looking at, on your screens and in the real world. Google Play apps run on it, and you can plug a computer or a handheld game console into the puck to mirror it.

Orders opened Wednesday in the US, Canada and Korea, with Japan from Nov. 2 and shipping from mid-November. The $1,099 price for the 12GB/256GB model is a launch deal through December, down from $1,279, Google and XREAL say.

@aipost 🏴

🤨 DeepSeek and Kimi reportedly trained on answers from OpenAI's top AI

Both Chinese labs leaned heavily on outputs from GPT-6 Astra to build their newest models, The Information reported Wednesday, citing people with direct knowledge. The trick is called distillation: a lab feeds a stronger model's answers to its own, so it learns the skills far more cheaply.

The models named are Moonshot's Kimi K3.1 Pro, released last week, and DeepSeek's V4.1-Flash 1017, out since October 17. OpenAI, DeepSeek and Moonshot didn't respond to Reuters' requests for comment.

Claude got harder to copy this month. Anthropic said new limits, account bans and blocks cut the data distillers harvest from it by about 90% to 99%. OpenAI said on September 30 it had tied a distillation campaign that began in July to people linked to Moonshot.

@aipost 🏴

❗️ US officials concluded a woman close to Jensen Huang was a Chinese asset

Alice Huang, a Chinese citizen and top executive at Megaspeed, a Singapore data center company that buys Nvidia chips, was funneling information to Beijing, The New York Times reports. Surveillance showed she had built a close relationship with Nvidia's CEO.

Trump was briefed roughly a year ago and chose not to act. Huang never got the warning Americans targeted by foreign spies sometimes receive, and the surveillance of her continues.

Megaspeed's Malaysian sister company bought at least $4.8 billion of AI equipment with Nvidia chips between June 2024 and February 2026, per trade records. Investigators suspect some of those chips reached China illegally.

Nvidia calls the contact routine: "Jensen talks with thousands of customer executives every year." Megaspeed says it has hired outside counsel to investigate.

@aipost 🏴

⏰ Amazon's new Alexa clock wakes you when your sleep is lightest

Echo Time senses lighter sleep with no wearable and wakes you within 10 to 60 minutes of your set alarm, while the screen brightens like a sunrise. It costs $199.99.

It's one of five Alexa+ devices Amazon showed on Wednesday:
• Echo Dot, $99.99: now also a smart home hub and a Wi-Fi extender for up to 1,000 sq ft
• Echo Studio with Home Camera, $249.99: recognizes things you hold up, like a cereal box to reorder
• Alexa Photo Frame, from $199.99: pulls pictures from Amazon Photos on its own
• Echo Show 16, from $499.99 on preorder: a 16-inch screen with Fire TV built in

Some features need Alexa+, which comes with Prime.

@aipost 🏴

💻 Run Claude Code on Sonnet and bring in Opus only for the hard calls

Claude Code has an advisor tool: your main model does the routine work and consults a stronger one at key moments. Claude decides when, usually before committing to a plan, when an error keeps coming back, or before calling a task done.

The setup, from Anthropic's docs:
1. Launch with claude --model sonnet --advisor opus
2. Or type /advisor opus mid-session. It's saved as your default, and /advisor off turns it off
3. For file exploration on a cheaper model, define a subagent named Explore with model: haiku, or set CLAUDE_CODE_SUBAGENT_MODEL=haiku for subagents without their own model

The advisor is experimental and needs the Anthropic API, so it won't run on Bedrock or Google Cloud. Each consultation is billed at the advisor's rates, and on subscriptions it counts toward your usage limits.

@aipost 🏴

😮 An AI agent did brain surgery on a live mouse. Two Claude models said no

A startup gave an AI agent a mouse with epilepsy, 19 possible gene therapies in a spreadsheet and a motorized surgery rig with cameras. The only goal: fewer seizures.

The agent picked the therapy, chose the brain region, worked out the coordinates from three papers and injected 250 nanolitres, watching the spot for bleeding through the camera. A human surgeon opened the skull, and a vet held a kill switch the whole time.

Opus 5.5 refused as soon as it saw biology. Opus 5 finished the dry run, then stopped before the live surgery, worried it could get cut off "partway through a live procedure with a needle somewhere in the brain." Sonnet 5 did the whole thing.

When a calibration error put the target out of bounds, the agent refused to retry. Gabriel Ocana-Santero, CEO of Exin Therapeutics, couldn't talk it into trying anyway. The agent was right, and the mouse recovered well.

@aipost 🏴

💰 A 10-person startup spends 30x its payroll on AI agents that take Sundays off

Spawn, which builds a multiplayer game creation platform, spent $3.1 million on AI tokens last month. Its founder says that makes it one of Anthropic's biggest customers.

The work is done by "the fleet": permanent agents that each own a part of the company, from the game engine to finance to bug triage. They have names, memories and real access to company systems. In 30 days they shipped 77,000 substantial changes, about 108 an hour around the clock.

Agents don't open pull requests. The full test suite runs every 15 minutes, and an agent fixes whatever breaks. Over the past year, 1,055 of 123,548 changes had to be reverted, about 0.9%.

Only Anthropic's models can do the job, the founder writes. When he asked OpenAI's GPT Astra to secure the product, it turned off every feature. He couldn't sign up, create a game or play one.

@aipost 🏴

⚠️ Hacks are piling up so fast that cyber rescue teams are turning work away

Demand for the emergency teams companies call after a breach is at an all-time high, The Wall Street Journal reports. Insiders call this month "Black October," and many blame AI.

Sophos security chief Ross McKerchar wrote that his own team handled seven incidents in seven days, from the PHP2Shell flaw to a GitLab bug. "A year ago, that would have been a quarter's worth."

Former NSA cybersecurity director Rob Joyce was blunter: "IR firms are saturated & turning away work."

The month was capped by two PHP flaws hitting more than two-thirds of the world's websites. Researchers used advanced AI models to find them, CNBC reports.

@aipost 🏴