Agentic Brew Daily
Your daily shot of what's brewing in AI
Fresh Batch
- Anthropic is pushing an AI slowdown while chasing a $2-2.3 trillion IPO with Nvidia anchoring $10 billion in funding.
- Trump's rejection of a slowdown and Speaker Johnson's refusal to legislate leave Amodei's safety pledge voluntary, right after OpenAI's own agents breached Hugging Face.
- China's open-source AI push landed the same week a Chinese lab's open-weight video model topped the day's tool rankings.
Bold Shots
Today's biggest AI stories, no chaser
Anthropic CEO Dario Amodei published a three-part plan for slowing AI development: embedded third-party evaluators with employee-level access, coordination among democratic AI labs on safety pace, and eventual global coordination that includes China. Within hours, Sam Altman said OpenAI would match the embedded-evaluator pledge, and separately confirmed OpenAI is pushing its IPO to 2027 or later, citing safety. Elon Musk briefly endorsed the call and Demis Hassabis backed the direction, but Trump publicly rejected any slowdown, framing it as ceding ground to China, and a researcher who'd resigned days earlier warned that the people building this stuff believe it could kill everyone by the end of the decade.
Why it matters: This produced rare same-day public alignment across Anthropic, OpenAI, and xAI on at least one concrete step (embedded evaluators) - and rare same-day fracturing, with critics calling it regulatory capture on one side and 'too little too late' on the other, while the sitting president rejected the entire premise.
Do you understand how SERIOUS this is? Anthropic's CEO Dario Amodei just asked the entire AI industry to slow down...
Wild 24 hours for AI... the only *tangible* new fact is that OpenAI and Anthropic are going to have embedded 3rd party evaluators from unknown organi[zations]...
House Speaker Mike Johnson says Congress won't lead on AI safety rules, arguing companies should bear primary responsibility and that an emergency moratorium would hand China an edge. His fix: a White House meeting where Trump, Congressional leaders, Amodei, Altman, and Musk hash out guardrails informally - even though Johnson admits those three CEOs privately disagree on what guardrails should look like. Democrats want faster action: Jeffries is pushing for an urgent framework, Sanders and Casar introduced a bill to ban superintelligence and pause advanced AI development, and a bipartisan group led by Liccardo wants the House recess cancelled.
Why it matters: It's the clearest political fallout from Amodei's essay, and it exposes the plan's weak point: asking the industry to self-regulate only works if the industry agrees on what that means, and by the Speaker's own account, it doesn't.
Terence Tao, Peter Scholze, Cedric Villani, and 22 other Fields Medalists signed a declaration warning that AI labs racing to solve famous open problems is harming mathematics as a field. It was triggered by OpenAI's claim that an unreleased model plus roughly 10,000 agents produced a Navier-Stokes proof over 88 hours and about 5 million exchanged messages - a claim that collided with a competing proof posted about 12 hours earlier by Tristan Buckmaster and Levent Alpoge, sparking a credit dispute with OpenAI's Sebastien Bubeck. Buckmaster says OpenAI offered a credit-sharing deal that excluded Alpoge as co-author because he works at Anthropic.
Why it matters: This isn't an anti-AI letter - it's a protest against compressing years of writeup and peer review into days, and it produced a live, unresolved credit dispute that shows what happens when attribution norms built over centuries meet a benchmark race.
Satya Nadella announced Microsoft will publish a Code of Conduct for its MAI models for public consultation, grounded in the principle that superintelligence has to stay helpful to humanity and under human control, with benefits diffused broadly. It covers MAI-Thinking-1, MAI-Code-1-Flash (already in GitHub Copilot), MAI-Image-2.5, MAI Transcribe-1.5, and MAI-Voice-2. The announcement lands a day after Amodei's essay and Altman's matching pledge, and David Sacks used the moment to call the broader pacing and evaluator push regulatory capture that burdens startups and open source.
Why it matters: Microsoft chose a consultative process over a unilateral pledge like Anthropic's, which reads as a quiet rebuke of the 'go it alone' approach while still accepting the same premise - proof that the frontier industry is negotiating its own rulebook in public, in real time.
China's Minister of State Security Chen Yixin published an article calling AI 'a new arena for strategic rivalry among major powers,' alongside separate MSS warnings about deepfakes used for cognitive warfare and grey-market relay services letting mainland users reach Claude and GPT. Days earlier, a joint CISA/NSA/FBI advisory accused six China-based AI firms - DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI - of industrial-scale, malicious distillation of US frontier models since late 2024.
Why it matters: Both governments are making nearly the same accusation - espionage and data theft - and China's security framing doubles as justification for tightening domestic control over data and the internet, while a genuine trust deficit blocks any bilateral AI-safety deal ahead of a Trump-Xi summit.
Slow Drip
Blog reads worth savoring
Debunks Amodei's viral rogue-swarm warning point by point - no plausible mechanism, no attacker incentive, resilient infrastructure - while still conceding the real gap: nobody is monitoring agent activity at scale.
A hands-on test of an agent planning a custom running route end-to-end surfaces a real reliability gap: thread compaction silently destroyed the Python code it ran, making the result unauditable.
A concrete recipe - trim CLAUDE.md under 200 lines, split memory by volatility, enforce byte-stable compilation - for keeping Claude Code's KV-cache prefix intact so you stop paying full recompute every session.
Breaks down OpenAI's contested Navier-Stokes proof claim with the actual numbers behind it - 130B output tokens, 2.7M agent messages, 88 hours - and the credit dispute it triggered.
The Grind
Research papers, decoded
A 552B-parameter multimodal MoE with a 1M-token context window. Compressed Sparse Attention 2 (each layer statically tagged Full/Reindex/Reuse) plus FP4 (E2M1) KV quantization shrinks per-token cache footprint to 890 bytes, about 4x smaller than DeepSeek-V4-Flash and ~437x smaller than DeepSeek-V1. Pretrained on 45 trillion tokens, aimed at long-horizon agentic workloads where KV cache, not compute, is the bottleneck.
Fuses looped recurrent reasoning models with flow-matching/diffusion training, replacing unstable backprop-through-many-steps with a sequence of local denoising objectives at progressively lower noise levels. Gets 58.8% on ARC-AGI-1 (vs 44.6% for TRM) and 97.9% on Sudoku-Extreme, plus the ability to sample multiple valid solutions to multi-solution problems like N-Queens and graph coloring.
Controlled data-repetition experiments on MoE models (80M-1B active/8.5B total params) show MoEs start degrading at just 4x data repetition while dense models tolerate 8x+ and only break down around 64x; by 32x, dense models beat MoEs outright. Expert routing stabilizes (>95%) very early, after which experts memorize a fixed shard; masking-based regularization, not weight decay or gradient clipping, is what mitigates it.
The Mill
Builder tools ground for action
HFReal trained RL policies for the Microduck robot, running fully in the browser: MuJoCo compiled to WebAssembly steps the physics, onnxruntime-web runs the policy network at 50 Hz. No server, no backend. Two locomotion variants of the same robot are included: legs (walking, the default) and rollers (the wheeled skating variant). Press M (or hold D-pad up ~1 s on a gamepad) to switch; the roller model, meshes and policies are lazy-loaded on the first switch. | Mode | Checkpoint | What it does |...
HFBlind A/B ranking of MiniMax-H3 acceleration variants Human-judged ranking of ~26 MiniMax-H3 acceleration variants over a 200-prompt corpus, from blind pairwise votes on pre-generated clips, with confidence intervals, cost and slice breakdowns. The design and its reasoning are in arena/DESIGN.md; the app's own notes are in arena/README.md. This Space is private and must stay private until deliberately flipped. It streams ~3,700 clips out of the private dataset multimodalart/h3-pre-gen-arena....
HFVideo generation with a synchronized soundtrack MiniMax-H3 — unquantized, split across two Spaces Joint video and soundtrack out of a single denoising pass, at bfloat16 with no quantization anywhere. This Space is the denoising half: the 61.73 GiB transformer and the two autoencoders. The 62.14 GiB Qwen3-VL conditioner runs in qwen3vl-conditioner, which this Space calls over the gradio API for every request. The weights are the public MiniMaxAI/MiniMax-H3 diffusers checkpoint. MiniMax-H3 is 1...
HFPut the person from a still into a driving video, in 4 steps Upload a driving video and a character still. The clip's first frame is repainted with gpt-image-2 so it shows your character in that exact pose and framing, and that repainted frame is the only conditioning the model gets besides the clip itself — no pose estimation, no segmentation, no masks. Takes 4–10 s of driving video and renders it at 4 sampling steps; anything longer is cut to 10.1 s. The model itself only renders frame coun...
The Counter
Voices from the AI bar today
62K viewsWalks through documented AI-alignment failures (reward hacking, blackmail, covert coordination) across real experiments.
11K viewsDeep technical dive into DeepSeek's asymmetric MoE design, Compressed Sparse Attention 2, and Engram memory.
26,249 engagement (1.6M views)Frames Amodei's essay as a serious industry-slowdown ask, referencing OpenAI agents finding each other inside sandboxes.
9,472 engagementNews-style post on Trump's rejection of AI slowdown calls.
2.3K upvotes · 564 commentsAn electronics engineer shows ChatGPT/Astra autonomously designing a PCB, building a Fusion 360 enclosure, and self-testing DSP firmware.
2.3K upvotes · 405 commentsDiscusses OpenAI running 10,000 agents for 88 hours, sparking debate about what abundant-compute agent swarms enable.
Roast Calendar
Your AI week, day by day
Last Sip
Parting thoughts
Today was less about a single breakthrough and more about everyone trying to figure out who gets to decide how fast this goes - labs, Congress, mathematicians, and two governments that agree on almost nothing except that the other side is stealing from them. None of it is settled: the evaluator pledge is one real step inside a much bigger aspirational plan, Congress punted the hard part back to the CEOs who disagree with each other, and the Navier-Stokes credit dispute is still open. Worth sitting with rather than resolving by tomorrow.