Agentic Brew Daily
Your daily shot of what's brewing in AI
Fresh Batch
- Kimi K3's launch triggered a selloff that briefly let Apple overtake Nvidia as the world's most valuable company before Nvidia reclaimed the title.
- Apple's trade-secret suit against OpenAI and Anthropic's Senate testimony about Alibaba's 25,000-account Claude-distillation scheme surfaced the same week, both alleging stolen AI capability.
- Anthropic is negotiating to lease $10 billion in compute from Meta the same week xAI's Colossus faces a DOJ-backed fight over unpermitted gas turbines.
Bold Shots
Today's biggest AI stories, no chaser
Moonshot AI shipped Kimi K3 on July 16 — a 2.8-trillion-parameter open-weight model with a 1-million-token context window that Moonshot calls the largest open-weight release to date. It placed third on GDPval-AA v2, beat Claude Opus 4.8 and GPT-5.5 on the Artificial Analysis Intelligence Index, and topped the Frontend Code Arena outright, all while running at $3/$15 per million tokens and shipping full weights under a Modified MIT license by July 27. The launch triggered a broad selloff — Z.ai fell ~30%, TSMC dropped 7%, SoftBank sank 9%, and Nvidia briefly lost its crown as the world's most valuable company.
Why it matters: If a Chinese lab can get this close to the frontier without frontier-scale chip access, the premium-API business model Anthropic and OpenAI depend on gets a lot shakier — and it's already wired into Cloudflare and Vercel's AI gateways, so you can try it today.
I made a game paying homage to Slay the Spire 2 using Kimi K3, called Slaughter Three Kingdoms... K3 did it all in one go, in just 8 hours, with 1830 art assets!
Master Tibet's Kimi K3 review is here — this time, it's genuinely badass! ... direct comparison test with Opus 4.8.
Meta is in early talks to lease AI computing power from its own data centers to Anthropic, a deal reported by NYT as worth up to $10 billion over two years with monthly payments and early-exit options on both sides. It would be Meta's first move into selling compute externally — turning its roughly $145 billion annual AI capex into a potential revenue line and putting it in the same business as AWS, Microsoft, and Google Cloud. Meta shares whipsawed on the report, falling over 5% intraday before closing down 2-3%.
Why it matters: Anthropic is already juggling a ~$100B AWS deal and a ~$45B SpaceX/Colossus arrangement — stacking a third major compute source shows just how compute-constrained frontier labs still are, even as Meta's own frontier ambitions look increasingly secondary to just selling capacity.
Apple filed a roughly 41-page lawsuit against OpenAI and its hardware subsidiary io Products on July 10, alleging trade secret misappropriation and breach of contract over unreleased product details. The complaint names OpenAI's Chief Hardware Officer Tang Tan — a 24-year Apple veteran — and former Apple engineer Chang Liu as individual defendants, and Apple says over 400 former Apple employees now work at OpenAI, with legal preservation letters already sent to about 40 of them. OpenAI denies wrongdoing and says it has seen no evidence backing the claims.
Why it matters: This lands right as OpenAI moves toward a public listing north of $100 billion, forcing a material legal risk into view for underwriters — and it's a sharp reversal from the Apple-OpenAI partnership that once built ChatGPT into Siri, before Apple quietly switched to Google Gemini.
xAI reportedly ran 59 unpermitted natural gas turbines at its Colossus 2 site in Southaven, Mississippi — roughly double what was previously acknowledged — prompting the NAACP, represented by Earthjustice and the Southern Environmental Law Center, to file an emergency motion to halt the pollution. The Trump administration's DOJ has moved to defend xAI, arguing the site's compute supports U.S. military operations and counts as a national security matter. More than 10,000 Southaven residents are also part of a separate class-action suit over turbine noise hitting 60-70 decibels near their homes.
Why it matters: Colossus has become the go-to symbol of the data-center backlash — the neighboring Boxtown community already faces four times the national cancer risk average, and New York's governor cited Colossus directly when signing a statewide AI data-center moratorium.
Kimi K3's release, timed near Xi Jinping's WAIC keynote on China's open-source AI strategy, reignited claims that the US is losing the AI race. But Transformer News argues K3 isn't actually at the frontier and doesn't warrant the alarm, and points out China faces the same incentive as the US to restrict advanced open-weight models once they become a national-security risk — which is exactly what's happening, as China's Ministry of Commerce met with Alibaba, ByteDance, and Z.ai to discuss restricting overseas access to their own models.
Why it matters: The 'China is winning' framing has become as much a Washington lobbying tool as a technical read — US firms invoke it to land contracts and dodge regulation, while Beijing simultaneously tightens its own export rules, undercutting the simple story on both sides.
Slow Drip
Blog reads worth savoring
A concrete four-question framework for scoping agentic AI risk instead of defaulting to "zero risk."
Data showing AI infrastructure demand is actually outpacing AI spend, complicating the simple "AI bubble" story.
The exact egress-proxy setup for letting a sandboxed coding agent call the GitHub CLI without ever seeing your real token.
Swapping "inject every tool schema into the prompt" for a BM25-searchable catalog once you pass ~50 tools.
The actual retrieval strategy and benchmark results behind a coding agent's self-evolving memory, not just the pitch.
The Grind
Research papers, decoded
A 975B-parameter MoE (41B active), 1M token context, trained from scratch on 45T tokens of text/image/audio/video with a controllable "thinking effort" dial. Full weights are on Hugging Face (including an NVFP4 checkpoint for Blackwell), fine-tunable on Tinker, and already served on Together, Fireworks, Modal, Databricks, and Baseten.
Compresses a 27B model (built on Qwen3.6-27B) into 1.125-1.71 bpw binary/ternary weights while keeping 89.5-94.6% of original quality, with a lossless speculative-decoding drafter; ships Apache 2.0 on llama.cpp and MLX, so a 27B-class reasoning model can now run on a laptop or high-end phone.
An open image understanding/generation family that approaches closed systems like Nano-Banana-Pro, trained on 208.62M images for a claimed ~$400K total cost via an agentic prompt rewriter and a dataset that captions its own flaws instead of filtering them out.
A robot-control World Action Model that decouples video-supervised training from action-only inference, hitting 85ms latency on an RTX 4090 with 0.85-0.89 task success rates and a 33% gain on long-horizon tasks over baselines.
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.
Multi-platform SDK for integrating GitHub Copilot Agent into apps and services
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.
HFZ Image Turbo is a Hugging Face Space tagged with gradio, mcp-server, region:us. It has 3561 likes on Hugging Face.
The Counter
Voices from the AI bar today
21K viewsA rigorous look at whether current AI systems can genuinely originate ideas versus recombine existing ones.
24K viewsA wide-ranging roundup connecting a massive new open model release, post-transformer architecture research, and AI regulatory moves in finance.
6.6K viewsA lecture-style deep dive on world models explaining why sample efficiency remains AI's biggest unsolved problem versus biological brains.
506 engagementAndrew Ng announces a new course built with Cerebras covering fast-inference hardware design and memory-bandwidth vs. compute tradeoffs.
2.5K upvotes · 585 commentsAnthropic disclosed to Congress a large-scale "distillation attack" where Alibaba allegedly used fake accounts at industrial scale to extract Claude's reasoning/coding capability and train Qwen — bigger than DeepSeek/Moonshot/MiniMax combined, and murky under current law.
Roast Calendar
Your AI week, day by day
Last Sip
Parting thoughts
Every story today traces back to the same resource — compute. Whether it's Moonshot training a 2.8-trillion-parameter model, Meta weighing whether to start selling its own GPUs, or a Memphis neighborhood pushing back on the gas turbines that power Colossus, the fight for scale keeps landing on real people and real power grids. Next time a new-model headline lands, it might be worth asking who controls the hardware underneath it, not just how the benchmarks look.