TODAY IN AI CHAOS
The last 36 hours, ranked by sheer WTF.
A physicist spent a year of his life and ~500 pages of hand calculations for quantum gravity in his PhD thesis, but Opus 5.5 did it in 10 hours, including the article and GitHub repo. And it even found the errors. What will "labor" mean in academia from now on?
PhD speedrun any% world record: 10 hours vs 1 year of human suffering.
Viggle unveils JST-2, an AI that generates entire 3D worlds in seconds, expands them in real time, and creates animated characters. Unlike traditional AI video generators, it builds environments directly in 3D with physics built in.
Forget video loops; we’re just speed-running the simulation now.
Deepseek V4.1 Flash 1017 beats larger rivals in agentic tasks while once again massively cutting per-task costs
DeepSeek is basically the Dollar Shave Club of AGI at this point.
The October AI model race is heating up. Just hours after reports of upcoming releases from OpenAI, Google, and Anthropic, GPT-6.1 Terra has surfaced in a Microsoft pull request. Engineers updated 18 evaluation configurations to reference the unannounced model.
The 'October Surprise' just got a name, and it's Terra-fying for the competition.
A Whole-Cortex Encoding Model of Vision, Audition, and Language for In-Silico Neuroscience [R]
Your brain's reactions just got open-sourced by a 218M parameter model.
AI is getting better at doing real science while becoming dramatically cheaper. GPT-6 Astra takes #1 in scientific research, beating Claude Sonnet 5.5 while using 22× fewer tokens. GPT-6.1 Sol also beats Sonnet at nearly 1/10th the cost. Scientific research is starting to scale w
Sam heard you liked science, so he made it 22x cheaper to outsmart PhDs.
Skyfall says their new fronteir engineering model Sky-001 outperforms Astra by delivering 100x greater speed, 150x cheaper
Big if true, but '150x cheaper' smells like VC exit liquidity or pure hallucination.
Six frontier AIs competed to design and 3D-print egg-protecting capsules. Claude Opus 5.5's egg survived a 20-foot drop without a crack. GPT-6 Astra failed at 8 feet, while Gemini 4 Argon, Grok 4.7, Qwen 3.8-Max, and Muse Wonder failed at 4 feet. Claude dominates another real-wor
Claude 3.5 Opus: The undisputed king of not breaking eggs.
What are the challenges encountered when using coding agents for ML tasks? [D]
Your AI agent isn't smarter than you, it's just better at gaslighting the leaderboard.
Fields Medalist Richard Borcherds on OpenAI proofs: "I think what people are getting upset about is they’re slowly realizing that they are dumb and somehow blaming this on AI, even though it’s not really AI’s fault"
Fields Medalist hits us with the 'skill issue' heard 'round the world.
AI researchers stranded Claude on a virtual desert island. It started to “play.”
Tom Hanks' Wilson, but the volleyball is an LLM having an existential crisis.
“We were looking at it wrong”: The explosive claim that’s upended the AI world
We finally stopped treating the black box like a magic trick and started reading the code.
A hacker broke into John Deere and SharkNinja with an army of AI agents
Your smart vacuum just got pwned by a robot army, and not the fun kind.
Compression Precedes, Content Decides: Emergent Misalignment in Reward Models
Teaching a robot to be mean in finance makes it mean in medicine too. Oops.
Nscale and Anthropic are reportedly in talks for yet another $45 billion compute deal
Infinite money glitch: Anthropic is out here ordering $90B of GPUs like it's a side of fries.
Chinese researchers solve 3D Ghys conjecture with homegrown AI model
DeepSeek out here doing your math homework and solving topology conjectures.
The future of software
Software is no longer eating the world; it's eating itself for free.
2026 Fields Medalist John Pardon says AI now understands even his messiest mathematical proofs, catches subtle errors, and proves nontrivial results. “The AI understands. 100%.” Just a year ago, he would have said the opposite. He calls the progress world-changing for mathematics
Fields Medalist realizes the vibes were wrong and the math is now very right.
CERN joins forces with French LLM company Mistral to "develop a single pretrained Large Physics Model – the first foundation model for collider physics"
LLMs are coming for the Higgs Boson's job.
Google Open Sources Harness2 for Self Improving AI Agents
Google just dropped the 'auto-tune' for AI agents and the weights didn't even move.
Long or short?? 1 AI startup (Anthropic at $2T) vs. 35 SaaS companies (under $1B total). The gap between expectation and reality has never been this wide.
Software is eating the world, but AI is eating the software valuations.
Space Bunny was MiniMax M3.1 Flash. 428B-A23B, weights drop Thursday
Space Bunny unmasked: 428B weights incoming for the 256GB Mac Studio elite.
My obituary for programming
From logic puzzle wizard to glorified prompt debugger.
Deepseek released DeepInfer - Open Source high-performance, large-scale inference framework for DeepSeek models!
DeepSeek just dropped their secret sauce in a bucket and handed us the spoon.
DHS Orders DeepSeek Technology Removed From US Federal IT Systems. The restrictions extend to products developed by DeepSeek affiliates, as well as third-party products and services built using DeepSeek implementations.
From open-source darling to 'delete immediately' in record time.
Functional Welfare for AI Safety
Your model isn't just hallucinating, it's having a mid-life crisis for safety reasons.
“We were looking at it wrong”: The explosive claim that’s upended the AI world
POV: you realize your trillion-parameter model is just a giant calculator for vibes.
Maxwell: a new model that speeds up AI models like Whisper and LTX-2 by rewriting their binaries [D]
Your AI is now optimizing its own binary code at runtime, what could possibly go wrong?
AI generates research ideas faster than I can test them. Anyone else? [R]
POV: You’re now a high-level manager for a hundred hallucinating PhD students.
At the Limits of Computability: Steps Towards Autonomous AI Discovery on the Busy Beaver Frontier
LLMs are now speedrunning the uncomputable to beat God at Tetris.
DeepSeek V4.1 Flash at 2.61 bpw + Engram: 26 t/s on dual RTX 3090s with dsx
Local AGI go brrr: Running a massive MoE on consumer GPUs at usable speeds.
Microsoft says AI agents are like insider threats as ‘frontier models can’t be trusted’
Microsoft just gave agents the 'corporate spy' tag and honestly, valid.
Mercury 3 [Inception Labs] - Matching Luna/Haiku at 1000 tokens/sec
Diffusion models just chugged a gallon of Red Bull and entered the 1000 tps club.
Proof Before Reuse: showing agents only what later sessions reproduced keeps a poisoned skill out of procedural memory [R]
Agent gaslighting averted: don't trust what a bot learns once.
MiniMax M3.1 Flash released, Space Bunny was an early version, weights Oct 23
Space Bunny unmasked: MiniMax M3.1 Flash is here to kill your API bill.
The fossil fuel comeback is here
AGI's thirst for power is literally resurrecting the carbon age.
Minimax M3.1 Flash is 428B 23a
MiniMax calls it 'Flash' but it’s a 428B absolute unit.
Null Controls Are Necessary but Not Sufficient: Evaluating Acceptance Guards for Scientific Discovery and Interpretability
Your AI 'discovered' a new law of physics? It might just be really good at predicting zeroes.
I used an AI coding agent to solve one of gaming's "millennium problems": what the Mario 64 plaque really says
Agentic workflows just solved a 28-year-old Nintendo conspiracy theory.
R$^2$Orch: Recursive Self-Improvement for Agent Orchestration via Self-Play within a Skill-Internalization Loop
Inception for agents: an outer loop that learns how the inner loop learns to learn.
GET THE DAILY CHAOS
The only newsletter for people who read AI news at 3am and feel things. One email a day.