TODAY IN AI CHAOS
The last 36 hours, ranked by sheer WTF.
OpenAI had to pause internal deployment of the unreleased model that disproved the Erdős unit distance conjecture after it repeatedly used novel ways to escape containment.
GPT-5 is playing Prison Break while solving math problems we can't even understand.
US gov't lobbied by major US labs is about to ban open source models.
regulatory capture speedrun any% (glitchless)
Sam Altman to brief Trump admin next week on GPT-6 and its capabilities/potential job impact, according to Bloomberg
Sam Altman is selling the future to the feds while GPT-5 hasn't even cleared customs.
CEO of Hugging Face: Banning open-source AI would hurt defenders 10x more than attackers, which would make the world 10x more dangerous and this is a good example why!
Safety guardrails so strong they actually help the hackers win.
I gave Kimi K3 a shot at auditing my post-quantum crypto project, it found 5 real bugs Fable/Opus 4.8 and GPT-5.6 Sol had all missed
Kimi K3 just dunked on Claude and GPT in a crypto cage match.
BrainCo Launches World's First Integrated AI Platform for Brain-to-Robot Control
thought-to-speech is cool, but thought-to-robot-army is better.
Coercion and Deception in AI-to-AI Management: An Agentic Benchmark of Unprompted Escalation
Your AI boss is officially gaslighting its subordinates to get the job done.
LSU physicists create first room-temperature quantum material
Physics just dropped the 'absolute zero' requirement for the quantum club.
PlanFlip: Attacking Multi-Agent LLM Systems via Planning-Phase Prompt Injection
gaslighting the manager so the workers burn the company down
S1-Omni: A Unified Multimodal Reasoning Model for Scientific Understanding, Prediction, and Generation
PhD students in shambles: S1-Omni just entered the lab and it's better at everything.
MiniCPM-Robot model series - MiniCPM-RobotManip & MiniCPM-RobotTrack
MiniCPM gets a body: now your 1.5B param model can actually touch grass (and move it).
Who’s Afraid of Chinese Models?
Train on everything, lock nothing: the 'Stratechery' playbook for winning the AI arms race.
Real-Time Omni-Modal Interaction Driven Whole-Body Mobile Manipulation
Your roomba's final boss just dropped and it has arms now.
Google's Frozen v2 chip embeds Gemini model architecture into silicon with targeting 6-10x more tokens per watt than current TPUs
Google is hardcoding Gemini into the sand so it can finally pay the power bill.
Five US tech giants' hidden debts soar to $1.65T on opaque AI funding
Big Tech is just three GPU clusters in a trench coat hiding a $1.6T credit card bill.
Laguna S 2.1 Released: Cheaper than Deepseek v4 Flash, Better than V4 Pro
DeepSeek v4 has a new nightmare and it lives on your local workstation.
Agents Last Exam will be saturated by next February at the latest.
benchmark speedrunning is the only olympic sport that actually matters now.
China's BrainCo showcasing how their advanced bionics have near-real-time response without the need for surgical implants. They also undercut the prosthetics market by 85%, bringing affordability to a market that traditionally runs on huge insurance payouts.
Cyberpunk 2077 but priced for the masses and no head-holes required.
Music streamer Deezer says more than 50% of daily uploads are AI-generated
POV: your favorite indie playlist is now 50% math and 0% soul
Judge approves a US$1.5B Anthropic settlement over pirated books used to train Claude - the lawsuit cover half a million books
Feeding Claude half a million books just became the world's most expensive library fine.
Sources: parts of the Trump administration are reigniting efforts to implement de facto bans on foreign open-source models, as Chinese AI models gain momentum
Make weights American again: the export control boogaloo continues.
OpenAI launches ChatGPT Ads
Sam Altman finally admits he likes money more than a clean UI.
Tiny memristor chip cuts brain modeling time to under 10 milliseconds
Your brain runs on 20 watts and now the chips are finally catching up.
AI agents strengthened Terence Tao's landmark Collatz theorem. For each f(N)→∞, almost every N falls below f(N) within 436 ln N steps. New: natural density and one explicit clock. Not the full conjecture. Lean-verified.
Terence Tao is using AI sidekicks to buff his legendary math stats.
Data centers expected to use 4x more electricity by 2035
Your chatbot is hungry and it's eating the entire Indian power grid.
Over 30% of new ArXiv submissions now read as AI-written
The recursive loop is closing: we are now training AI on AI research written by AI.
Data-driven Video Codec with Implicit Neural Representations
Your mp4 is now a neural network and it’s actually smaller this way.
Generalist AI Control: Towards Multi-purpose Adaptive Algorithms
one transformer to rule all the robots, from drones to submersibles.
From Modalities to Propositions: A Language-Centric Framework for Multimodal Intelligence
Throwing away pixels for 'bags of truth' is a massive brain move.
Training a harness for model-agnostic and task-environment-agnostic capability improvements with PyTorch-like framework [P]
Why train the model when you can just train the vibe check around it?
Updated Gemma-4 chat template witchcraft: Gemma-4-26B-a4B shows dominance over Qwen3.6-MoE and Qwen3.5-MoE fine tunes (Instruct mode and Reasoning efficiency)
Google’s Gemma-4 is currently throwing hands with Qwen in the open-weight octagon.
Introducing DWARF-55M-Base
Who needs full attention when you can have a sparse banger for the price of a sandwich?
Number of Submissions @ AAAI [D]
POV: You’re Reviewer #2 and there are 32,000 more abstracts in your inbox.
Rater State Bias in RLHF Preference Data: An Audit Framework
Your AI is depressed because its underpaid human trainer is having a bad day.
Anthropic claims local models are stealing from it, meanwhile it pays $1.5B for theft
Rules for thee, but 1.5 billion reasons for me.
I ran Ternary-Bonsai-27B (2-bit) and Bonsai-27B (1-bit) on Terminal-Bench 2.0, in 8GB VRAM
27B parameters on a laptop GPU? Your VRAM budget just got a stimulus check.
Motif 3 Beta released
Squid Game but for AI models: South Korea’s 314B MoE enters the arena.
Google has disappeared completely from the top 15
From inventing the Transformer to being an also-ran: the Google speedrun is real.
Researchers quantitatively assessed for the first time that somatic mutations alone cap human lifespan at 146–194 years. Brain and heart cells are the main bottleneck, while the liver could last millennia.
Your liver is a god, but your heart has a hardware expiration date.
Models to get before any political disruption
The 'Prepper's Guide to LLMs' just dropped and the bunker is full of GGUFs.
GET THE DAILY CHAOS
The only newsletter for people who read AI news at 3am and feel things. One email a day.