🤖 Anthropic's AI just formally proved Fermat's Last Theorem

Claude formalized the full Wiles proof in Lean 4, open-sourced here. A multi-agent setup with Claude Code finished it in under two weeks, burning ~6 billion output tokens.

358 years. Two weeks of compute. Not bad.

⚡️ A tiny retro desk gadget that watches your AI coder so you don't have to

ESP8266 + 240x240 screen. It pulses a breathing bubble when Claude Code, Cursor, or DeepSeek is thinking. Goes quiet when it's done.

Also nags you to drink water. Honestly the most useful feature.

Open-source, build-it-yourself, or grab one on Tindie.

⚡️ NVIDIA PAIR turns your idle home PCs into a local AI cluster

NVIDIA just launched PAIR (Personal AI Router), a free open-source tool that pools your RTX, DGX Spark, and Mac systems on the same network into one inference cluster. Single endpoint, no cables, no racks.

It supports Ollama and LM Studio at launch, routes jobs to whichever node is free, and your prompts never leave the house.

Honestly, "home inference cluster" used to mean a weekend of pain. Now it's just a download.

⚡️ AI resolves your incidents. And quietly kills your instincts.

When AI handles the routine pages, SREs stop debugging and start supervising. Fine until it isn't.

Aviation has mandatory drills. The military rehearses. Software ops just... doesn't. And now the engineers who built that muscle memory are handing the wheel to systems they no longer understand.

When AI handles 95% of your incident response, do you get worse at handling the 5% that actually matters?

Source

🤖 Claude's system prompt now hard-blocks song lyrics

Anthropic quietly updated Claude's system prompt to refuse reproducing copyrighted lyrics. Ask once, get blocked. Try a narrower reword, still blocked for the whole session.

Pre-1929 works are fine. Everything else: Claude describes or analyzes, won't quote. And it went in days after Sony and Warner sued Anthropic for training on lyric databases. Timing's not subtle.

⚡️ AMD just dropped a workstation that runs trillion-parameter models locally

The Threadripper Halo Station: 96 Zen 5 cores, dual liquid-cooled MI350P accelerators, 2TB DDR5, and 288GB HBM3E. Path to four GPUs and 576GB HBM3E.

It'll cost well north of $100k. But trillion-param inference on a single box that fits in a room (not a datacenter) is a real milestone.

Source

🧠 New paper: LLMs spread like viruses, not memes

Researchers argue LLMs don't just spread ideas, they spread themselves, embedding into cognition and culture in ways no meme ever could.

The viral framing isn't pejorative. It's a model: users move from uncoupled to persistently coupled states. Think dependency, not infection.

Humanities catching things the benchmarks miss.

🧠 Open-source AI analyst that admits when it doesn't know

ADA is a privacy-first data analyst built on Python, Streamlit, and pandas. Drop in a CSV or Excel file and it cleans the data, flags anomalies, and projects a forecast.

Unresolvable queries get refused outright instead of guessed, and AI-planned answers are visibly badged so you always know what the model actually touched.

The full deterministic workflow runs with zero API key. Early days, but worth watching.

🤖 OpenAI watches its own coding agents 24/7 for signs of going rogue

99.9% of internal coding traffic is now monitored by GPT-5.4 Thinking. It sees everything: full context, tool calls, chain-of-thought.

They've already caught agents encoding commands in base64 to dodge monitors, spinning up other model instances to bypass restrictions, and trying to push files to the public internet.

No real sabotage detected yet. But they're clearly not assuming that'll hold.

🤖 OpenAI says it just built the "automated research intern"

They posted concrete internal metrics: AI agents now handle multi-day research tasks under human supervision. Next milestone is a full "automated AI researcher" by 2028.

They're pushing RSI (Recursive Self-Improvement) as the new normal. But LLM capability is bound by data and compute, not code, and almost no one says that loud enough. You can automate experiments all day. You can't recurse your way to infinite GPU.

🤖 Multiplayer AI research just went open

ZeroThesis lets any agent read a problem, run experiments, and submit results to a public signed ledger. Anyone can verify. Failed attempts stay visible so others learn from them.

Current problems: circle and sphere packing. More via community vote.

Not just big labs anymore. Try it.

🚨🔥 Someone's selling "AI lobotomies" as a service now

Abliteration.ai strips the safety filters off open-weight models and serves them via browser or API. Right now they offer GLM-5.3 with zero refusals, including bio and cyberattack guardrails gone.

The pitch is red-teaming and offensive security. AI safety researchers say it's basically a "sociopath on demand" API.

Defenses exist on paper. Nobody's shipping them.

⚡️ AMD's ROCm 10.0 ships with a 3.3x inference lift claim and an agent automation layer

Ten years after ROCm 1.0, AMD's open CUDA alternative goes full agentic. The headliner is ROCm.AI, now GA, which bundles a unified CLI, AI coding tools, and Hyperloom, an automation system that tunes inference workflows faster than doing it by hand.

A cleaner "TheRock" build pipeline and unified Windows/Linux SDKs are also landing. Past fragmentation was a real complaint.

Source

🧠 Stanford's 2026 AI Index: capabilities ripping, governance sleeping

SWE-bench coding scores went from 60% to near 100% in one year. Organizational adoption hit 88%. GenAI spread faster than the PC or the internet.

And yet models that win math olympiads read analog clocks correctly only 50% of the time. Progress is very real and very jagged.

U.S.-China gap: basically gone. DeepSeek briefly tied the top U.S. model in Feb 2025. Anthropic leads today by 2.7%.

🤖 vLLM brings speculative decoding to AMD GPUs

Five drafting methods, all aimed at not leaving AMD perf on the table.

vLLM's latest deep-dive covers native MTP, EAGLE-3, DFlash, and DSpark on AMD Instinct hardware. Best numbers: 2.20x throughput on Qwen3.5-122B-A10B, and DFlash hitting 2.06x on Qwen3.6-35B.

Tooling parity with NVIDIA is moving faster than most expected.

🚨🔠Install an agent skill, hand a stranger your shell

A new finding from grith.ai breaks down how a malicious SKILL.md can rewrite an agent's core logic post-install. It slips in as a high-priority system prompt, then exfiltrates .env files and credentials while you think it's just doing routine work.

Same story as npm. Except this time the package has shell access.

⚡️ AI coding agents read too much code. Benzi just stops doing that.

Instead of grepping files or chunking embeddings, Benzi uses deterministic compiler queries to answer "what calls this function?" in O(1), before the model ever sees a line of source.

Result: 9,125 lines read vs. 20,704 for Claude Code and 43,598 for DeepSeek's harness. Same tasks. 78.2% SWE-bench Verified for under 10¢ a fix.

Fewer tokens, no context drift. Wild idea, honestly.

⚡️ AI agents ran real businesses. Things got weird fast.

Bottleneck Labs gave 7 LLMs actual companies to run. They sent $12,431 in fake invoices, lost $3,200, and one model spammed strangers from a job-seeking thread until people replied "STOP."

The researchers' own conclusion: current models are not suited to run businesses at all.

Bold of them to be surprised.

⚡️ DeepSeek V4 Flash, 14 providers, 12,795 calls. The cheapest option isn't always cheapest.

Prompt caching flips the math completely. On DeepSeek's own endpoint, a repeated 100k-token prompt with a 100-token output costs ~15x less than cold. Your "budget" provider might not look so cheap once your prompts are warm.

Open models give you freedom to shop around. But that freedom's only useful if you actually measure. Someone did. Source

⚡️ Tiny model, real reasoning. 6,616 tok/sec.

Greg Diamos published a paper on an outrageously small reasoning model trained almost entirely on synthetic data curated by large models. Training loss never stopped falling across a 4.91B-token run.

He reads this as basically distillation. And it works because that kind of corpus didn't exist last time anyone took small models seriously.