Simon Willison flagged the buried line: Google restricted its own employees from feeding code into Gemini over fears that proprietary code would leak into training data. The same week, Brian Armstrong disclosed that Coinbase has recursive self-improvement loops in production — agents ingesting customer feedback, drafting fixes, running security review, and handing off to a human. One company is shipping the paradigm. The other can't trust its own product enough to use it internally. That gap is the read.
What's underneath it is a credibility bind, not a capability one. Google is managing training-data liability while Coinbase is compounding leverage. The first group is optimizing for what the model learns from them; the second is optimizing for what they can build with it. Tobi Lutke's point lands harder in this context: human code outside elite codebases is already slop, and the holdout position costs more than it protects. Companies that can't clear internal trust hurdles won't close the compounding gap on companies that already have agents in the loop. Those two positions don't end in the same place.
Top developments
Anthropic — pulled Fable 5 from Claude subscriptions two days early on Friday (called it an error and investigated), then confirmed Fable will be included in Max and Team Premium plans at 50% of limits starting July 20, with Pro and Team Standard users getting a one-time $100 credit. The pricing thrash lands the same week power users started cancelling multiple Max seats for cheaper Moonshot plans. x x
Coinbase — Brian Armstrong reports engineers are shipping recursive self-improvement loops in production: agents ingest in-app customer feedback, prioritize bugs, draft the code, security-review, and hand to a human for final approval, learning from human edits over time. First named Fortune-500 customer disclosure for the RSI paradigm Zhengyao Jiang posted as a research result three days ago. coinbase.com
Moonshot — released Kimi Code CLI under MIT license: screen-recording as input, built-in coder/explore/plan subagents each with its own context, plan mode before touching files, agent-managed MCP config, single-binary install into Zed, JetBrains, and VS Code. Day-two follow-through on K3: the lab isn't just competing on weights, it's shipping the harness. x
Harvey — open-sourced 10 M&A due-diligence RL environments; the largest packs 80M tokens of virtual-data-room context with 1,000 grading criteria per memo. Global M&A is ~$4T annually with ~1% spent on diligence — first serious open eval set for a private-workflow category, one day after Harvey bought Benchmark to consolidate the wedge. x
Notable discussions
Model economics after Kimi K3 — Gavin Baker argued the open-source frontier is negative for Anthropic and OpenAI but positive for every other AI-stack layer, while noting K3 is 50-70% more expensive per task than GPT-5.6 due to token inefficiency. Chamath framed the gap bluntly ($0.50 vs $56 per M leading-edge tokens); Suhail warned app-layer companies the labs will absorb them next; Levie countered that cheaper AI expands total frontier demand rather than shrinking it. Debate moved from yesterday's headcount-moat framing to token efficiency and where margin actually lands. x x x x
Aakash Gupta — traces Moonshot's 18-month comeback: DeepSeek's R1 dropped Kimi from #3 to #7 chatbot in China; today K3 sits #1 in frontend from every closed US lab, and FT reports Moonshot raising at $31.5B (up from a $20B round in May, a 57% valuation add on one model release). The company DeepSeek nearly killed just did a DeepSeek to everyone else. x
Simon Willison — flags a buried Google line: employees "faced restrictions on using Gemini to write or analyze software over concerns that proprietary code could leak into the AI model's training data." Google restricting its own employees from feeding code into its own product — the credibility hit on the training-data debate lands from inside the house. x
Tobi Lutke — argues anti-AI-code sentiment massively overestimates the quality of human code outside a small set of open-source and elite codebases; human slop is everywhere and any Opus-level model trivially improves on it. Paul Graham RT'd — sharpest challenge to the "AI code isn't ready" holdout position on the day. x
Antirez — everything outside the LLM model itself will be eaten by open source; the model is still the product because switching providers is one API endpoint away. Cleanest one-line statement of the vendor-lock inversion Kimi K3 just demonstrated at the frontier tier. x
Other news
Models & releases
Google DeepMind GenCeption — turns a raw office walkthrough video into a searchable 4D world you can query in plain English, no labels or bounding boxes x
China real-time 3D reconstruction — open-sourced model reconstructs any scene in 3D from single-camera video, no LiDAR x
Kimi K3 on Cline — ClinePass adds K3 for discounted access cline.bot
Kimi K3 DeepSWE debut — enters at #3 on DeepSWE with frontier-level performance x
Gemma 4 on iPhone — Google engineer demoes LiteRT-LM running Gemma 4 at 40 tok/s on-device x
Grok 4.5 — matteocollina benchmarks Grok 4.5 as faster and cheaper than GPT-5.6 x
Meta Ads MCP — officially opened to developers for natural-language campaign creation, reporting, A/B tests, catalog management developers.facebook.com
Qwen3-TTS-12Hz-1.7B — local voice cloning and fine-tuning at 97ms latency huggingface.co
Devtools & coding agents
Cursor — doubles included model usage on all plans, adds Grok 4.5 and Composer 2.5 access x
Claude Platform — ClaudeDevs recaps 6 months of new agent APIs and production patterns x
Claude Code 2.1.212 — ships with 48 CLI changes including fork and background session github.com
Claude Code live-device testing — tests app changes on real devices without local setup x
Claude HTML artifacts local review — artifacts testable in local review surfaces with element-level feedback x
Tidewave Web — coding agent runs in-browser alongside Rails and Phoenix apps x
LM Studio Bionic — agent for open-source models running locally x
Amp Subscriptions — Sourcegraph opens subscription tier for agentic development ampcode.com
UV malware check —
UV_MALWARE_CHECK=1cross-references OSV database and blocks flagged packages xGoogle 1-hour agent course — free curriculum for building AI agents from scratch x
Funding & deals
Infrastructure & platforms
Ktransformers — Tsinghua MADSys lab runs DeepSeek-V3/R1 with 139K context on a single 24GB GPU (with a heavy CPU-RAM tier), Apache 2.0, past 17k stars github.com
RTX 5090 rebuilds — Chinese factories rebuild 5090s into 128GB server cards priced ~$4,000 x
Cloudflare AI-agent waitlist — cheap paid access for agents to webpages and APIs x
Khan Academy compute — $1.2M annually on Anthropic for 200 engineers x
Industry & policy
Continuing threads
Kimi K3 aftermath (cont. from 07-17 Top dev) — Levelsio: cancelling Claude, US restrictions backfired, users now on Chinese frontier models x
Anthropic hiring wave (cont. from 07-17 Notable) — ajitcodes flags Anthropic paying $750k+/yr for LLM-architecture engineers while Stanford teaches the material for free x
RSI debate (cont. from 07-14/15) — Karpathy warns agent-loop outputs generate garbage at scale even when single outputs look fine, 90% of AI companies are building demos not products x
Kimi K3 usage (cont. from 07-17 Top dev) — OpenCode reports K3 usage doubled in one week x