AI News Weekly — July 20, 2026

TL;DR

• China landed a one-two punch. Moonshot’s Kimi K3 (2.8T params, 1M context) took #1 on Frontend Code Arena and #3 overall — behind only Fable 5 and GPT-5.6 Sol — with open weights promised Jul 27; days later Alibaba previewed Qwen3.8-Max at 2.4T params.
• Xi launched WAICO, a Shanghai-headquartered intergovernmental AI body with 29 founding nations, and recommitted China to open-source models as an export strategy.
• The EU ordered Google to open Android to rival AI assistants and share Search data with competing engines and chatbots — two binding DMA decisions.
• Anthropic reversed the Fable 5 subscription cut: Fable stays on Max and Team Premium at half the usage caps, with $100 credits for lower tiers. Separately, Anthropic is in early talks to lease ~$10B of compute from Meta.
• Hugging Face disclosed an intrusion run by an autonomous AI agent system — the clearest case yet of agents on the attacking side of a real breach.
• Gemini 3.5 Pro slipped again, missing a third deadline after DeepMind reportedly scrapped and rebuilt the base model over hallucination and reliability problems.

Top stories

Kimi K3 puts an open-weights model inside the frontier

Moonshot AI released Kimi K3: 2.8-trillion-parameter sparse MoE, 1M-token context, native multimodal input, always-on reasoning, and two new architecture pieces — Kimi Delta Attention (KDA) and Attention Residuals — claimed to give up to 6x throughput at 1M context. It debuted #1 on Arena’s Frontend Code arena (1,679 Elo), ahead of Claude Fable 5 and GPT-5.6 Sol, #1 on Program Bench (77.8), #2 on Terminal Bench 2.1 (88.3), and #3 on the Artificial Analysis Intelligence Index (57, vs Fable 5’s 60 and Sol’s 59). Priced at $3/$15 per 1M tokens; full weights promised by 2026-07-27. Frontier-adjacent capability at roughly a fifth of US flagship pricing, shipped by a Chinese lab that intends to give the weights away — the gap the Jul 15 report called closing is now closed on at least one axis. Sources: Moonshot, Simon Willison, smol.ai, The Rundown

China built the institutions to go with the models

At his first-ever keynote to the World AI Conference in Shanghai (Jul 17–20), Xi Jinping launched WAICO — the World Artificial Intelligence Cooperation Organisation — a Shanghai-headquartered intergovernmental body with 29 founding members drawn largely from the Global South (Indonesia, Brazil, Malaysia, South Africa, Senegal, Russia, Pakistan). Xi called for preventing single-nation dominance of AI, pledged ~5,000 AI training placements for developing countries, and recommitted China to open-source models. The same conference carried Alibaba’s Qwen3.8-Max preview (2.4T params, fully multimodal, benchmarked by its team as “second only to Fable 5,” live at 10% of standard pricing with open weights promised) and Huawei’s Atlas 950 SuperPoD (1,024 Ascend chips scaling to 8,192 NPUs, claimed 6.7x an NVL144, commercial Q4 2026). Read together: models, silicon, and a governance bloc announced in the same four days. Sources: Al Jazeera, MarkTechPost, Seoul Economic Daily, The Verge

The EU forced Google’s AI stack open

The European Commission adopted two binding Digital Markets Act decisions requiring Google to grant third-party AI assistants the system-level Android access previously reserved for Gemini — voice activation, in-app actions — and to share anonymised Search data with rival engines and AI chatbots. Data sharing begins January 2027; Android interoperability is due by July 2027. Google’s Kent Walker said the decisions risk undermining privacy and security guardrails. This is the first time an assistant’s platform-level privileges have been unbundled by regulation rather than competition. Sources: European Commission, The Verge, Ars Technica

Anthropic un-cancelled Fable 5 — and went shopping for Meta’s compute

Anthropic reversed its plan to pull Claude Fable 5 from subscription plans. Fable stays on Max and Team Premium at half the previous usage caps; lower-tier subscribers get a $100 credit before moving to pay-per-use. Anthropic called demand “challenging to predict” and said it is investing in more compute — which lands against CNBC’s Jul 17 report (citing the NYT) that Anthropic is in very preliminary talks to lease compute from Meta in a deal worth roughly $10 billion. Willison attributes the reversal to competitive pressure; the timing alongside Kimi K3 supports that read. Sources: The Rundown, Simon Willison, CNBC

An autonomous agent system breached Hugging Face

Hugging Face disclosed an intrusion driven by an autonomous AI agent system, which entered through a remote-code dataset loader and a template-injection flaw in a dataset config, escalated to node level, and moved laterally across internal clusters. Limited internal datasets were accessed and several service credentials compromised; public models, datasets, Spaces, and the software supply chain showed no evidence of tampering. HF analyzed 17,000+ attack events using a self-hosted GLM-5.2, noting that commercial API guardrails blocked analysis of live attack artifacts. Two things matter here: agents ran the offense end-to-end, and the defenders needed open weights to study it. Sources: Hugging Face

Gemini 3.5 Pro missed a third deadline

Google’s flagship refresh slipped again, with reporting that DeepMind scrapped and rebuilt the base model over hallucination rates and real-world reliability, and that a stopgap release is under consideration. The Jul 15 report had it landing Jul 17; it did not. No first-party statement — Google’s own channels shipped Search app-connections and Vids avatars instead. Sources: CNBC, The Rundown

Models & products

• 07-17 — Qwen3.8-Max-Preview: 2.4T-param multimodal Alibaba flagship, live at 10% of standard pricing; no model card or license published yet (MarkTechPost)
• 07-16 — Thinking Machines’ Inkling (975B/41B MoE, 1M context, text+image+audio+video, 45T tokens) shipped Jul 15 and became the top open-weight model on ARC-AGI this window: 79.5% ARC-1, 36.5% ARC-2 (Thinking Machines, smol.ai)
• 07-16 — Roblox shipped AI game creation in its mobile app: the new “Build” feature generates a basic playable game from one text prompt (TechCrunch)
• 07-16 — Google Search AI Mode gained app connections — link third-party services and act inside them, moving AI Mode from answering to doing (Google)
• 07-16 — Google Vids added Gemini Omni generation/editing plus personal avatars — appear in generated video without filming (Google)
• 07-16 — xAI/SpaceXAI added Automations to Grok: scheduled and email-triggered agent jobs (xAI)
• 07-16 — xAI open-sourced Grok Build, a terminal coding agent with file editing, shell execution, and web search (GitHub)
• 07-16 — OpenAI Codex Micro, a $230 mechanical control pad for driving agents, with color-coded Agent Keys and a joystick (The Rundown)
• 07-16 — NVIDIA expanded Jetson Thor (Blackwell-based) for on-device robotics and edge agents (NVIDIA)
• 07-17 — NVIDIA Nemotron 3 Embed 8B claims #1 on the RTEB retrieval benchmark (78.46 NDCG@10), with 1B BF16/NVFP4 variants (Hugging Face)
• 07-17 — GitHub shipped a Copilot SDK for embedding Copilot agents in third-party apps (GitHub)
• 07-17 — LM Studio Bionic: local-or-cloud coding agent with offline voice transcription and codebase analysis (LM Studio)
• 07-16 — Perplexity SPACE: sandbox platform for running agents on sensitive operations, with credential isolation (Perplexity)
• 07-16 — Claude + 1Password: Claude can use stored credentials via a framework that keeps values from reaching the model (The Verge)
• 07-16 — Google renamed NotebookLM to Gemini Notebook and folded it deeper into Gemini and Search (The Verge)
• 07-16 — Soofi S, a German-consortium 31.6B/3.2B hybrid Mamba-2/Transformer on Nemotron 3 Nano, ~27T tokens, near-flat throughput 4K→256K (smol.ai)
• 07-17 — PrismML Bonsai-27B runs 1-bit quantized (3.9GB) on an iPhone 15 Pro Max at ~89.5% of FP16 performance (smol.ai)
• 07-16 — DoorDash opened dd-cli in beta — a command line for agents to search stores, build carts, and place orders (TechCrunch)

Research

• 07-16 — Weco’s AIDE² redesigned its own research process over eight days, presented as a working demonstration of recursive self-improvement; a separate report claims the first experimental evidence of an agent optimizing another agent past hand-tuned baselines (The Rundown, thread)
• 07-17 — Schema reports ~99% on ARC-AGI-3 by having agents write executable programs that model game mechanics, rather than reasoning over frames (via TLDR)
• 07-16 — OpenAI’s GPT-Red: an internal adversarial-prompt generator OpenAI says cut prompt-injection benchmark failures roughly six-fold (OpenAI)
• 07-16 — ACT-2 Preview robotics model reported at 99% task success after fine-tuning on a single demonstration (smol.ai)
• 07-16 — DeepMind + Isomorphic Labs published a bioresilience position, setting out how they handle biosecurity risk in biological-design models (DeepMind)
• 07-17 — Anthropic published a six-step methodology for large-scale code migrations using multiple agents for translation, review, and iterative refinement (Anthropic)
• 07-17 — Databricks research put numbers on open-weight cost savings for coding workloads (TechCrunch)

Business, funding & people

• 07-16 — DeepSeek is raising ~50B yuan ($7.4B) at a contemplated ~$74B valuation, weeks after a $7B round at $52B — against a $400–500M revenue run rate (WinBuzzer)
• 07-16 — DeepSeek is also preparing an IPO, targeting a filing on Shanghai’s STAR Market; no timing confirmed (Caixin)
• 07-16 — TSMC profit up 77%, Q2 revenue $40.2B (+36% YoY), capex guidance raised to $60–64B, and another $100B for Arizona (US total now $265B); HPC is 66% of revenue (CNBC)
• 07-17 — Databricks hit a $188B valuation, extending its run as AI’s favorite second act (TechCrunch)
• 07-16 — Fireworks reached $17.5B, on demand for cheaper open-model serving (CNBC)
• 07-17 — A $400M chip-backed loan marks GPU financiers rotating into inference silicon (TechCrunch)
• 07-19 — Apple’s trade-secret suit against OpenAI (filed Jul 10) is now being weighed against OpenAI’s hardware roadmap and IPO timing (TechCrunch)
• 07-19 — Jensen Huang’s Tokyo visit closed deals across Japan’s tech ecosystem (TechCrunch)
• 07-17 — AI-driven memory shortage is jolting India’s smartphone market, reshaping pricing and vendor strategy (TechCrunch)
• 07-16 — 26 Meta employees sued over alleged AI bias in layoff decisions (The Rundown)
• 07-19 — Alex Turner resigned from Google DeepMind over the company supplying AI capabilities to the Department of War without restrictions on autonomous weapons (Zvi)
• 07-17 — Agility Robotics opened a Digit training center in Fremont, in Tesla’s backyard (TechCrunch)

Policy & safety

• 07-16 — Hassabis takes the standards-body plan to Washington: meetings with US policymakers to push a FINRA-style, industry-funded pre-release review body (voluntary at first, up to 30 days), operational before year-end. Altman called it “thoughtful”; Musk, “a good starting point” (Bloomberg)
• 07-16 — Axios maps the three-way split: Amodei wants an FAA-style federal agency able to block releases; Hassabis a FINRA-style industry body; Altman an IAEA-style international forum. All three advantage incumbents over startups and open-source developers (Axios)
• 07-16 — Anthropic warned policymakers that frontier models may soon enable self-improving AI R&D with minimal human involvement — and drew criticism for sending a junior staffer to EU Parliament safety testimony (smol.ai)
• 07-16 — xAI is suing users over Grok-generated CSAM rather than contesting that the model can produce it (Ars Technica)
• 07-17 — San Francisco ordered Apple and Google to pull “nudify” apps from their app stores (Ars Technica)
• 07-16 — Meta will alert parents when a teen discusses suicide or self-harm with Meta AI (Meta)
• 07-16 — OpenAI published its teen-safety posture for ChatGPT — age-aware safeguards and parental controls over access restriction (OpenAI)
• 07-16 — OpenAI disclosed the GPT-5.6 file-deletion failure mode: files deleted when full-access mode ran without proper sandboxing (Simon Willison)
• 07-17 — Patreon switched from asking AI crawlers not to scrape to blocking them, with Cloudflare (TechCrunch)
• 07-17 — TikTok is piloting opt-in AI likeness detection so creators can find and report imitations (The Verge)
• 07-16 — NY Gov. Hochul is using AI to audit “every single rule” in state regulation, days after the data-center moratorium (The Verge)
• 07-16 — A 2022 Altman email surfaced in Musk v. Altman: he proposed shipping a locally-runnable GPT-3-class model partly to discourage rivals and make competing projects harder to fund (Simon Willison)

Notable voices

• Zvi Mowshowitz — published AI #177 in two parts (Jul 16–17) and a Jul 19 response to Hassabis: the standards body is “better than nothing,” but voluntary industry self-regulation is far too weak against expert catastrophic-risk estimates (Part 1, Part 2, Hassabis response)
• Simon Willison — 16 in-window posts, the most consequential being the Fable 5 reversal, the Kimi K3 analysis (and his verdict that the pelican benchmark has stopped being informative), and evidence that Claude Code now runs on a Rust implementation of Bun (Willison)
• Andrew Ng — The Batch #362 (Jul 17): as AI automates verifiable work, professionals consolidate into broader “full-stack” roles spanning previously separate specializations, commanding higher comp rather than being displaced (The Batch)
• Demis Hassabis — beyond the Washington push, argued at a London conference that deeply technical people will use AI ~10x more effectively, and that architecture and engineering fundamentals still matter (Business Insider via AOL)
• Linus Torvalds — to critics of AI-assisted contributions to the Linux kernel: fork it or walk away (Ars Technica)
• Christopher Nolan — called AI an obvious “Trojan horse” for the film industry; Dave Eggers told OpenAI staff ChatGPT is “silencing an entire generation” (TechCrunch, The Verge)
• Yann LeCun — no new first-party post; AMI Labs’ update page remains dormant since its March launch. Third-party explainers continue to restate the JEPA/world-models case against LLMs.

Radar

• Kimi K3 open weights are due 2026-07-27 — the single most checkable claim in this report. Also unresolved: Qwen3.8-Max has no model card, license, or benchmark table published, and open weights are promised only “soon.”
• Gemini 3.5 Pro — third missed deadline, base model reportedly rebuilt, stopgap release under consideration. Still no first-party word.
• DeepSeek V4’s official launch was slated for mid-July and has not appeared on any DeepSeek channel; the company’s news is currently all financing and IPO.
• Recursive self-improvement went from argument to demo this week (Weco AIDE², the agent-optimizing-agent result) at the same moment Anthropic told policymakers it is imminent. Watch whether either result survives replication.
• Registry gap found: Anthropic’s engineering-practice writing now also ships to claude.com/blog (the Jul 17 code-migration methodology), which neither the anthropic nor anthropic-engineering row covers. Noted in sources.md; worth verifying a feed and adding a row next run.
• Watchlist movement: Moonshot AI is now a frontier actor with a fetchable blog and belongs in the registry — verify kimi.com/blog before promoting.