🗞️ AI Daily Briefing — 2026-07-16

🔥 Top Story

China’s AI companion law took effect yesterday, forcing ByteDance’s Doubao and Alibaba’s Qwen to shut down personalized AI agents overnight — hundreds of millions of users woke up to features switched off. The Interim Measures for AI Human-Like Interaction Services, published by five Chinese agencies in April, formally took effect July 15. Providers must now disclose AI isn’t human, add anti-addiction reminders, include self-harm crisis pathways, and are barred from offering virtual companion services to minors. It’s the most aggressive regulation of AI emotional engagement anywhere in the world — and a template other governments are watching closely.

🚀 Model & Research News

  • Gemini 3.5 Pro reportedly targeting July 17 GA launch. Google’s rebuilt flagship aims for a 2M-token context window (the largest of any production frontier model), a Deep Think reasoning layer, and autonomous workflow capabilities. Google has still not officially confirmed the date or specs. If it lands tomorrow, expect a benchmark shakeup.

  • Future of Life Institute releases 2026 AI Safety Index — nobody passes. Anthropic earned the highest grade at C+, with OpenAI and Google DeepMind at C, Meta at D+, and xAI, DeepSeek, and Mistral effectively failing. The report card highlights that even the most safety-conscious labs have major gaps.

  • ARC-AGI-3 scores confirm frontier models still can’t do interactive reasoning. GPT-5.6 Sol leads at just 7.8% — a humbling result given that v1 is at 95% and v2 at 85%. Terra scored 0.8%, Luna 0.2%. The benchmark evaluates whether agents can learn unfamiliar task mechanics through action and feedback — the kind of flexible learning humans do effortlessly.

  • Andrej Karpathy reportedly joins Anthropic. The former Tesla AI director and OpenAI founding member has joined Anthropic’s ranks, along with Tom Blomfield (Monzo co-founder) who is joining the AI compute team. A major talent acquisition if confirmed.

  • South Korea unveils $880B decade-long AI infrastructure plan. President Lee Jae-myung announced 1,350 trillion won in spending — ~$518B for memory chip manufacturing, ~$550B for AI data centers, targeting 8.4GW of data center capacity by 2029 and growing humanoid robotics market share from 1% to 20% by 2028.

🛠️ Tools & Developer Updates

  • LangChain kicks off “Build More with LangSmith” Summer AMA Series. Running July 9 through August 12, the series connects developers with product experts for deep dives on LangSmith features and best practices.

  • Gemini 3.5 Pro developer preview expands on Vertex AI. Enterprise teams can now access the model in expanded preview ahead of the expected GA launch, testing the 2M-token context and Deep Think mode for production workflows.

💰 Funding & Business

  • Neko Health raises $700M Series C at ~$7B valuation. Spotify founder Daniel Ek’s AI-powered full-body health scanning startup raised from Lightspeed and O.G. Venture Partners to fund US expansion. Celebrity backers include Mark Zuckerberg, Maria Sharapova, and Alexis Ohanian. A 4x valuation leap from $1.7B in January 2025.

  • Chai Discovery lands $400M Series C at $3.8B for AI drug design. The OpenAI-backed biotech tripled its valuation in 7 months after signing Novartis, Pfizer, and Eli Lilly as pharma partners. Index Ventures led alongside Kleiner Perkins and Sequoia. AI-designed antibodies are reaching Big Pharma pipelines.

  • Anthropic-Samsung custom chip talks continue. Anthropic is exploring Samsung’s 2nm process for a custom inference chip, following OpenAI’s Jalapeño chip with Broadcom. The project is early stage — no design finalized yet — but signals every major lab now wants to own its silicon destiny.

  • TSMC posts all-time revenue records. June revenue up 68% year-over-year, Q2 at $39.6B. The AI chip supply chain is red-hot.

🐦 Notable from the Timeline

  • Jim Fan (NVIDIA GEAR Lab) gave his “Robotics: Endgame” talk at Sequoia AI Ascent — laid out a roadmap to Physical AGI mirroring the LLM success story. His bet: World Action Models where robots dream by generating video a few seconds into the future. NVIDIA’s GR00T models have now passed 10M downloads. He puts 95% certainty on Physical AGI by 2040.

  • Ilya Sutskever’s explosive deposition (from the Musk v. OpenAI lawsuit) continues to circulate — revealed that OpenAI and Anthropic nearly merged after Altman’s firing, with Anthropic’s Dario Amodei potentially taking the helm. Helen Toner has publicly disputed parts of Sutskever’s account.

  • Sam Altman appeared on CNBC to announce GPT-5.6 general availability and pitched the US government a $42.6B equity stake in OpenAI to share AI-driven economic gains.

  • The AI Safety Index scoring every major lab below a C+ is generating heated debate across the timeline. The consensus: nobody is doing enough, but at least now there’s a grading rubric.

📊 Benchmark Watch

LMArena leaderboard (July 2026): Claude Opus 4.6 Thinking holds #1 at 1501 Elo. Claude Fable 5 leads coding (1563 Elo) and vision (1327 Elo). Opus 4.6 Thinking tops math (1518 Elo). The Anthropic sweep across categories is notable — but Gemini 3.5 Pro’s imminent launch could reshuffle the board. GPT-5.6 Sol’s ARC-AGI-3 lead (7.8%) shows OpenAI still has edges on novel reasoning tasks, even if Arena voters prefer Claude.

🎙️ Podcast Highlights

  • Lex Fridman Podcast #490 — “State of AI in 2026” dropped recently covering LLMs, coding, scaling laws, China, agents, GPUs, and AGI. A comprehensive 3+ hour deep dive worth bookmarking.

  • Jim Fan’s Sequoia AI Ascent talk is the must-watch robotics content this week — 20 minutes distilling the Physical AGI roadmap.

🔗 Worth Reading