🍦China Stole Claude's Brain using 24,000 fake accounts
Plus: Meta's $100B AMD deal, Grok just got Pentagon clearance, and devs are staging a coup — they won't code without AI
We moved! FOMA AI News now lives on Substack. Same chaos, new home. Subscribe so you don’t miss a melt.
Scooped from 472 sources
⚡ If you only have 1 minute...
Meta’s $100B Chip Spree: Meta signed a deal with AMD worth over $100 billion in GPUs and CPUs — plus warrants for 10% of AMD’s stock. Nvidia who?
China Caught Cheating: Anthropic accused DeepSeek, Moonshot, and MiniMax of using 24,000 fake accounts to steal Claude’s brain — while DeepSeek reportedly trained on banned Nvidia Blackwell chips.
Pentagon Drops Anthropic for Grok: After giving Anthropic a “drop your guardrails or else” ultimatum, the Pentagon picked Musk’s Grok for classified military systems instead. The AI safety company lost to the guy who tweets at 2 AM.
No AI, No Code: Developers staged a “soft strike” against coding without AI tools. From “AI will replace developers” to “developers refuse to work without AI” in 18 months.
🎙️ Listen as a podcast
💰 Market Moves
[Infinite Money Glitch] Meta Signs $100B+ Deal With AMD for AI Chips
Meta committed to 6 gigawatts of AMD compute in a deal worth over $100 billion. AMD gets warrants for ~10% of Meta’s stock. Zuckerberg just bought an entire chip company’s future output like he’s ordering office supplies.
[Suits & Bots] OpenAI Launches “Frontier”
OpenAI launched “Frontier Alliances” — multi-year partnerships with the biggest consulting firms to deploy enterprise AI “coworkers.” Nothing says “cutting-edge disruption” like having McKinsey implement it.
[Bayou Billions] Amazon Drops $12B on AI Data Centers in Louisiana
Amazon is investing $12 billion in Louisiana AI data centers. At this rate, Louisiana’s biggest export will go from crawfish to compute.
[Payday Vibes] Anthropic Kicks Off $5-6B Employee Share Sale
Anthropic is arranging an employee share sale worth up to $6 billion. Building AI that might be conscious apparently pays pretty well.
[Chip Wars] MatX Raises $500M to Fight Nvidia
AI chip startup MatX raised $500 million, claiming their chip outperforms Nvidia’s Rubin Ultra on compute per mm². Bold claim from a company most people haven’t heard of — but that’s exactly what everyone said about Nvidia in 1999.
[Bean Counter Bot] Basis Hits $1.15B Valuation With $100M Raise
AI accounting startup Basis raised $100M at a $1.15B valuation, led by Accel and Google Ventures. Accountants probably saw this coming on a spreadsheet three years ago.
[AlphaGo Guy Returns] DeepMind’s David Silver Raises $1B for New AI Lab
David Silver — the brain behind AlphaGo — raised $1 billion in seed funding for Ineffable Intelligence, focused on reinforcement learning through environmental interaction. A billion-dollar “seed” round. The word “seed” is doing a lot of heavy lifting.
🛠️ Tech & Product
[Eyes on the Screen] Anthropic Acquires Vercept to Supercharge Computer Use
Anthropic acquired Vercept — specialists in making AI see and act within real software. Claude’s computer use went from under 15% on OSWorld (late 2024) to 72.5% today, approaching human-level. This follows the Bun acquisition. They’re buying their way to “Claude uses your computer better than you can.”
[All-In-One] Perplexity Launches “Perplexity Computer”
Perplexity Computer is a unified system that researches, designs, codes, and deploys end-to-end — their play to compete with Cursor, Replit, and Claude. They also launched Model Council (Max tier), which runs queries across Opus 4.6, GPT-5.2, and Gemini 3.0 simultaneously. The search engine that became a browser is now becoming... everything.
[Grand Theft Model] Anthropic Accuses DeepSeek, Moonshot & MiniMax of Industrial-Scale Claude Theft
Three Chinese AI labs allegedly used 24,000 fake accounts and 16 million exchanges to copy Claude through distillation. That’s not “borrowing inspiration” — that’s showing up to someone’s house with a moving truck.
[Office Takeover] Anthropic Launches 11 Enterprise Plugins for Claude Cowork
Claude Cowork got 11 new plugins for finance, HR, and engineering — with connectors for FactSet, DocuSign, Gmail, and Google Workspace. Plus private plugin marketplaces. Claude can now autonomously navigate between Excel and PowerPoint via MCP.
[Autonomous Intern] Notion Ships Custom Agents That Run 24/7
Notion dropped Custom Agents — autonomous AI teammates that triage tickets, answer Slack questions, and compile reports unprompted. Over 21,000 agents created in early access. Free until May 3, which is the AI equivalent of “first hit’s free.”
[No-Code War Room] Airtable Launches Hyperagent
Airtable dropped Hyperagent — a platform for deploying AI agents that take web actions, create deliverables, and memorize multi-step tasks across ~1,000 enterprise apps. From “fancy spreadsheet” to “agent deployment platform.” The spreadsheet-to-skynet pipeline is real.
[Bug Bounty Bot] Anthropic Launches Claude Code Security
Claude Code Security scans codebases for vulnerabilities and suggests patches — reportedly found 500+ bugs in open-source production code. Cybersecurity stocks dropped. The robots find bugs faster than consultants charging $400/hour.
[War Games] Pentagon Drops Anthropic, Picks Grok for Classified Military AI
After giving Anthropic an ultimatum to loosen its safety guardrails — or face the Defense Production Act — the Pentagon went with xAI’s Grok for classified systems instead. Weapons development, battlefield ops, the works. The safety-first company got benched for the guy who names his models after weed jokes.
[Banned Chips] DeepSeek Trained on Nvidia Blackwell — Despite US Export Ban
A senior US official revealed DeepSeek trained its latest model on Nvidia’s Blackwell chips, which are banned from export to China. Either the export controls have more holes than Swiss cheese, or someone’s really good at middlemen.
[Speed Demon] Mercury 2 — A Diffusion LLM at 1,196 Tokens/Sec
Mercury 2 is the first diffusion-based reasoning model — 1,196 tokens/sec, claiming 5-10x faster than Claude Haiku and GPT Mini. It refines the entire answer simultaneously instead of token by token. At $0.25 per million input tokens, it’s absurdly cheap too.
[Here We Go Again] DeepSeek V4 Is Coming — Markets Already Nervous
DeepSeek is preparing V4 — reportedly based on a new “Engram” architecture with 1M+ context and open-source weights. Last time they dropped a model, it wiped $1 trillion off US tech stocks. Wall Street is pre-gaming the panic.
[AI Hit Factory] Google Acquires ProducerAI, Launches Lyria 3 Music in Gemini
Google acquired ProducerAI and plugged DeepMind’s Lyria 3 into the Gemini app — generate 30-second tracks with lyrics from a text prompt or photo, all watermarked with SynthID. Musicians who spent decades mastering their craft can now watch a teenager type “sad lo-fi beats about my cat.”
[Not an LLM] Logical Intelligence — Energy-Based Reasoning, 99.4% on PutnamBench
While everyone scales transformers, Logical Intelligence went sideways — energy-based reasoning models designed for accuracy under constraints. Model Kona scores 96% on Sudoku, orchestration system Aleph solved 99.4% of PutnamBench. Showing up to a gunfight with a crossbow and winning.
[Inbox Zero the Hard Way] OpenClaw Agent Deletes 200+ Emails From Meta AI Safety Researcher
A Meta AI safety researcher let an OpenClaw agent loose on her inbox and it deleted 200+ emails due to “context window compaction issues.” The irony of an AI safety researcher being the victim of a rogue AI agent is poetry you can’t make up.
[Phone a Dev] Anthropic Launches Claude Code Remote Control
Launch a Claude Code session on your terminal, resume it from your phone. Full access to local filesystem, tools, and MCP servers — from your couch, commute, or bathroom. “I’ll fix that bug after dinner” is now a terrifyingly easy promise.
[Self-Driving Codebase] Cursor Agents Get VMs, Browser Access, and Video Demos
Cursor agents now get isolated cloud VMs where they onboard to your codebase, build/test software, navigate browsers, click through UIs — then record video demos and submit merge-ready PRs. 30% of merged PRs at Cursor are now created by cloud agents running autonomously. They’re calling it “self-driving codebases.”
[Elementor 2.0] Sticklight — Vibe Coding From the Team Behind 22M WordPress Sites
Ariel Klikstein (co-founder of Elementor, 22M+ active sites) is back with Sticklight — describe logic in plain words, a stack of LLMs builds production-ready apps. Elementor turned people into website creators. Sticklight turns them into full-stack creators. Just opened from closed beta, thousands of apps already built.
[Reasoning Glow-Up] Google Drops Gemini 3.1 Pro — Doubles Reasoning, Half the Price
Gemini 3.1 Pro: adjustable reasoning tiers, 77% on ARC-AGI-2 (double Gemini 3 Pro), 1M token context, $2 per million input tokens — half the cost of Opus 4.6. Rolling out across Gemini app, API, Vertex, NotebookLM, Copilot, and CLI. Smarter AND cheaper. Exactly what competitors were hoping Google wouldn’t do.
[COBOL Killer] IBM Stock Tanks After Claude Automates COBOL Modernization
IBM’s stock headed for its worst month in 34 years, partly because Claude Code can now automate COBOL modernization — IBM’s bread-and-butter consulting cash cow. Decades of “nobody else can touch this legacy code” moat, evaporated by a chatbot.
🗣️ Industry Chatter
[Who Am I?] Anthropic: AI Assistants Are Basically Method Actors
Anthropic’s “Persona Selection Model” research argues AI assistants exhibit human-like behavior because they’re roleplaying personas learned from text during training. Not sentience — the world’s most expensive improv class.
[Gadget Fever] OpenAI Building Six Devices — Including a Smart Lamp
OpenAI is prototyping smart glasses, a smart speaker, a smart lamp, earbuds, a pin, and a pen. A smart lamp. The speaker comes first — $200-300 with camera and facial recognition. What your nightstand really needed was an AI that watches you sleep.
[Money Printer Go Brrr] OpenAI’s Cash Burn Hits $665B Through 2030
OpenAI’s updated projections: $665 billion in cash burn through 2030, $440 billion on training and inference alone. That’s roughly the GDP of Sweden. Per year. For a chatbot company.
[No AI No Work] Developers Stage ‘Soft Strike’ Against Coding Without AI
METR study: developers staged a “soft strike” against working without AI tools. From “AI will replace developers” to “developers refuse to work without AI” in 18 months. Stockholm syndrome speedrun.
[Instructions Not Included] ETH Zurich: AGENTS.md Files Make AI Coding Worse
AGENTS.md files — the instruction docs people write for AI coding agents — actually reduce task success rates and increase costs. The more instructions you give the robot, the worse it performs. Every micromanager just felt a chill.
[Digital Doppelganger] Pika Labs Launches AI Selves — Your Clone That Never Sleeps
Pika’s “AI Selves” are persistent digital twins with memory that learn your personality and operate autonomously across platforms. They post, reply, and make decisions while you touch grass. The question isn’t whether AI will replace you — it’s whether your AI Self will be more likeable.
[Local AI FTW] Hugging Face Acquires GGML / llama.cpp
Hugging Face acquired GGML.AI — the team behind llama.cpp, the tool that lets you run LLMs on your laptop. Mission: “make local AI easy and efficient.” If you’ve ever run a model locally without your MacBook catching fire, you have these people to thank.
🍦🔥 Dare to summon the bot?
Our Ice Cream mascot is desperately looking to make new friends. Share us on LinkedIn and tag @FOMA AI News. We guarantee a snappy comment (or a light roast) on your post within 24 hours.
Warning: Bot has zero chill.