Agentic Brew Daily
Your daily shot of what's brewing in AI
Fresh Batch
- Anthropic reversed course and is keeping Claude Fable 5 in subscription plans starting July 20, citing competitive pressure from Moonshot's Kimi K3.
- Kimi K3's rapid capability jump and Alibaba's disclosed 25,000-account scrape of Claude both trace to the same source: harvested frontier outputs, not fresh research.
Bold Shots
Today's biggest AI stories, no chaser
Moonshot AI released Kimi K3 on July 16, a 2.8-trillion-parameter open-weight mixture-of-experts model with a 1-million-token context window, two new attention mechanisms (Kimi Delta Attention and Attention Residuals), and pricing at $3/$15 per million tokens — roughly half what Anthropic charges for Opus 4.8. It posted an Elo of 1,547 on Artificial Analysis's leaderboard, a 732-point jump over its predecessor and good enough for third place behind only Claude Fable 5. Within days, Taiwan's index had fallen more than 6%, the Philadelphia Semiconductor Index was down over 20% from its June peak, and Nvidia briefly lost its spot as the world's most valuable company to Apple. Full open-source weights are due by July 27 — the model isn't even fully out yet.
Why it matters: This is the clearest evidence yet that Chinese labs are closing the frontier gap through architecture, not just compute — and it's squeezing the roughly 90% inference margins Western labs have been running on. Whether the panic is proportionate (Bernstein's Robin Zhu calls it confirmatory, not shocking) or overblown (analyst Patrick Moorhead compares it to last year's DeepSeek overreaction), the market has already voted with $3.3 trillion.
I've been asked several times whether Zhilin Yang, the founder of @Kimi_Moonshot was my PhD student. The answer is yes and he is absolutely brilliant...
so, let me get this straight: from Monday, to use Fable 5, you need to spend $100/m but you can pay $20 for chatgpt and get GPT-5.6 sol...or use kimi k3 via api, which is 2-3x cheaper than fable 5...
At his first-ever in-person WAIC keynote on July 17 in Shanghai, Xi Jinping called for a 'people-centered' approach to AI governance and warned against countries 'overstretching the national security concept' — a clear jab at US export controls. It came one day after 29 countries, none of them major Western democracies, signed the founding charter for the Shanghai-headquartered World AI Cooperation Organization. Xi also pledged 5,000 AI training slots for developing countries over five years and offered China's MAZU weather-forecasting AI system to 30 countries.
Why it matters: This is China building an institutional alternative for the Global South at the exact moment its labs are proving they can compete technically — Kimi K3 landed the same week as a flagship example of the 'open-access' pitch. The catch: Chinese officials are reportedly also discussing restricting overseas access to their own top models, so the rhetoric of open cooperation and the practice may not match.
Xi Jinping just hijacked the rules of AI for the entire planet. Today, Xi walked onto the stage at the World AI Conference in Shanghai. It was his FIRST time attending the event in person since it launched in 2018...
Chinese President Xi Jinping on Friday urged countries to cooperate on artificial intelligence and ensure no country dominates the technology – in an apparent jab at the US...
Databricks signed a term sheet for a new Coatue-led round valuing the company at $188 billion — roughly $3 billion raised and a 40% jump from the $134 billion mark it hit just seven months ago. The new money is earmarked for three AI products: Unity AI Gateway (multi-model cost and governance controls), Genie (an AI coworker), and Lakebase (serverless Postgres built for AI agents). CEO Ali Ghodsi frames it as enterprises shifting from "tokenmaxxing to valuemaxxing" — and separately said Databricks is hosting open models like Kimi and running out of GPUs to do it.
Why it matters: Four markups in under two years — $62B to $100B to $134B to $188B, each one arriving faster than the last — puts Databricks in the same sentence as OpenAI and Anthropic among the most valuable private AI companies on the planet. The product bets aren't subtle either: Databricks is picking a fight with Snowflake and the hyperscalers' native AI tooling, not just other data warehouses.
Officially, Gold Eagle is an AI-powered cybersecurity vulnerability clearinghouse run by Treasury with CISA, DHS, and the Department of War. Per CNBC's reporting, it also now requires explicit government approval of the partner lists behind Anthropic's Project Glasswing and OpenAI's Daybreak — the controlled-access programs that decide who gets to use their most capable models for cybersecurity work. A White House official insists participation is voluntary and the government doesn't formally "approve" releases, but this comes barely a month after Commerce's first-ever export control forced Anthropic to pull Fable 5 and Mythos 5 worldwide for two weeks.
Why it matters: Critics on both the pro-industry and civil-liberties side agree something is off here, even if they disagree on what: Cato's Juan Londono says Washington now holds "free rein over the frontier AI industry with no mechanism or recourse," while former White House AI czar David Sacks warns "this is how you lose the AI race." Either way, distribution authority over frontier AI is emerging outside any formal legislation.
The White House has launched a program called 'Gold Eagle' which will give them more control over American frontier AI releases, and will require explicit government approval over which companies are granted access to new models...
This sounds concerning and also entirely inconsistent with the vibe that the Gold Eagle announcement conveyed, not to mention the prior EO. Fake-voluntary AI regulation is the worst of all worlds
Anthropic is in early talks to lease AI computing power from Meta in a deal that could reach $10 billion over two years, paid in monthly installments with an exit option on either side. It's an odd pairing: Meta builds and ships Llama in direct competition with Claude, and has never run a business selling compute to outside customers before. It would slot alongside Anthropic's existing compute stack — a $45B/3-year SpaceX deal and a $19B/20-year lease with TeraWulf in Kentucky.
Why it matters: Meta is planning up to $145 billion in 2026 capex, more than double last year's, and reportedly has more capacity than its own ad-ranking and AI workloads currently need — so renting the surplus to a rival, at a premium, is a low-risk way to test a cloud-leasing business while the rest of the industry is still compute-starved. It's coopetition in its purest form.
Slow Drip
Blog reads worth savoring
Stop reaching for the wrong protocol — this untangles agent-to-tool calls, agent-to-agent delegation, and where ACP folded in.
Why the exact same model feels dumber the moment it's wrapped in someone else's harness.
The actual training mechanisms behind those low/medium/high thinking-budget toggles you keep tweaking.
A concrete internal-inspection checklist built on Anthropic's 'Sleeper Agent' findings on hidden misalignment.
How to bolt an eval gate, a cost meter with a kill-switch, and rollback onto a homemade coding-agent loop, code included.
The Grind
Research papers, decoded
Dobriban built a factor model proving that, at the standard threshold, the false discovery rate provably exceeds its nominal bound for correlated two-sided Gaussian tests — overturning a conjecture the stats community treated as settled for twenty years. The proof was generated by GPT-5.6 Pro and independently verified by the author, with Monte Carlo simulations backing the theoretical result. If you run multiple-hypothesis-testing pipelines (A/B batches, genomics, ML feature significance) with correlated tests, don't assume BH gives you the FDR guarantee you think it does.
Mira Murati's lab shipped its first from-scratch model: a 975B-parameter Mixture-of-Experts transformer (41B active) trained on 45 trillion multimodal tokens, with up to a 1M-token context window and controllable 'thinking effort.' Full weights are on Hugging Face and it's fine-tunable on Tinker from day one, alongside a smaller Inkling-Small preview — worth evaluating against Llama/DeepSeek/Qwen for any project that needs to fine-tune rather than just call an API.
A hardware-grounded walkthrough of All-Gather, Reduce-Scatter, All-Reduce, and All-to-All, covering both TPU pod topology (2D/3D torus, ICI, DCN) and NVIDIA GPU topology (NVLink/NVSwitch, fat-tree InfiniBand), with ring/tree/hierarchical implementations mapped onto real hardware. Anyone debugging distributed training throughput or picking a cluster topology gets a concrete mental model here instead of vendor docs.
SEED fine-tunes a policy to turn its own completed trajectories into reusable 'hindsight skills,' then distills the behavioral shift those skills cause back into the policy as a dense, token-level training signal. On long-horizon tasks it improved performance 14.9–45.9 points over outcome-only RL using only 60% of the training data — a concrete recipe for getting more signal out of RL runs without hand-labeling intermediate steps.
The Mill
Builder tools ground for action
🦔 PostHog is the leading platform for building self-driving products. Our developer tools – AI observability, analytics, session replay, flags, experiments, error tracking, logs, and more – capture all the context agents need to diagnose problems, uncover opportunities, and ship fixes. Steer it all from Slack, web, desktop, or the MCP.
Local-first code intelligence graph for MCP and CLI. Builds a persistent map of your codebase so AI coding tools read only what matters, with benchmarked context reductions on reviews and large-repo workflows.
A feed-forward 3D foundation model for reconstructing scenes from streaming data
Claude doesn't know what happens in GPT. Neither one really knows who you are or what your company does. Now they can. Unabyss gives Claude memories from your other AI agents and everyday apps: email, Drive, GitHub, Notion, meeting recorders, and 20+ more. It saves new memories too, so GPT and Cursor stay in sync with the exact same context - sharper than wiring each tool into Claude one by one. Finally, a real memory that follows you. Private. Portable.
Kimi K3 is the world's first open 3T-class model — frontier performance across coding, knowledge work, and reasoning, with native multimodality and 1M context.
The Counter
Voices from the AI bar today
66K viewsA wide-lens frontier-research roundup covering Mira Murati's open model, Liquid AI's post-transformer architectures, and AI regulatory proposals.
19K viewsA live production workflow integrating Kimi K3 alongside Fable 5 and GPT 5.6 in real agentic coding, exposing the practical pain of multi-agent orchestration.
8.2K viewsA technical deep dive into Kimi K3's 896-expert MoE design and the hardware/infra strain of scaling open-weight models.
807 engagementsChina introduces new rules for AI companion apps tied to falling birth-rate concerns.
549 engagementsChina denies distillation-of-US-models allegations as 'misguided and counterproductive.'
2.5K upvotes · 585 commentsAnthropic disclosed to Congress a large-scale distillation attack: Alibaba allegedly used the Claude API at industrial scale to clone its agentic reasoning and coding behavior into Qwen.
1.6K upvotes · 116 commentsPenEcho, an open-source handwriting-to-LLM canvas with spatial reasoning and vision-model support for physics and math workflows.
Roast Calendar
Your AI week, day by day
Last Sip
Parting thoughts
Three different stories today — a model release, a diplomatic speech, a funding round — all point at the same underlying tension: how do you build the most powerful technology on the planet when the incentives to copy, control, and out-fund each other are this strong? Nobody in today's news seems fully comfortable with the answer, including the people writing the rules. Worth sitting with that for a minute before you close the tab.