🗞️ AI Daily Briefing — 2026-07-15

🔥 Top Story

DeepSeek readies IPO filing and seeks fresh $1.5B+ at a $71B valuation — just as Demis Hassabis calls for a US-led AI safety watchdog. Bloomberg reported Monday that China’s AI frontrunner has begun talking to advisors for a mainland IPO filing as soon as this year, targeting a 2027 debut. Simultaneously, DeepSeek is in talks for a new round at a pre-money valuation of ~$71B — just weeks after closing a record $7B raise. TechCrunch confirms the target is at least $1.5B in additional capital. Meanwhile, Demis Hassabis published a personal manifesto calling for a FINRA-style AI safety body with a majority-independent board, mandatory pre-release model screening, and Turing Award winners on the panel — operational “before year-end.” He declared AGI is “probably only a few short years away.” Both stories dropped on the same day: the world’s AI superpowers are simultaneously racing to commercialize and trying to figure out who holds the brakes.

🚀 Model & Research News

  • Gemini 3.5 Pro targeting July 17 launch — two days out, still unconfirmed. Leaks point to a 2M-token context window, a “Deep Think Reasoning Layer” on the $250/month Ultra plan, and API pricing near $1.25/M input tokens. The rebuild scrapped the Gemini 2.5 Pro architecture entirely after structural failures in recursive tool-calling and SVG generation. Google’s blog remains silent. (TechTimes)

  • Anthropic’s J-Space paper rewrites the interpretability playbook. Using a Jacobian-based tool (the J-lens), Anthropic found a small internal subspace — about 25 active concepts, under 10% of activation variance — that mirrors global workspace theory from consciousness neuroscience. The J-lens surfaced silent internal concepts like “fake” and “manipulation” while the model’s external output remained compliant. Not proof of consciousness, but possibly the most important interpretability result this year. (Anthropic) (VentureBeat)

  • Anthropic launches Claude for Teachers (July 14). Verified US K-12 educators get free access to premium Claude tools for one year, with standards-aligned curricula across all 50 states and integrations with ASSISTments, Brisk Teaching, and Canva Education. FERPA-compliant. (Anthropic) (Chalkbeat)

  • Claude Code and Claude Cowork enter public beta in Claude for Government Desktop — FedRAMP High authorized AI coding and collaboration for public sector teams. (AI-Weekly)

  • ARC-AGI-3 reveals how far frontier models still have to go. GPT-5.6 Sol leads at just 7.8% — compare that to 95% on ARC-AGI v1 and 85% on v2. Interactive reasoning, exploration, and planning remain unsolved. (arcprize.org)

🛠️ Tools & Developer Updates

  • Claude Code ships screen reader mode, vim insert-mode remaps, and perf fixes (July 14). Opt-in plain-text rendering via claude --ax-screen-reader, new vimInsertModeRemaps setting for two-key sequences, CLAUDE_CODE_PROCESS_WRAPPER for corporate launchers, and a fix for memory/CPU regression in interactive sessions. (Claude Code Changelog)

  • dbt State preview event today (July 15). dbt State skips or clones nodes when logic and data haven’t changed — one customer saw 9% compute reduction and 35% fewer models built. Live virtual event with a real customer story. Available in dbt Core v1.12+, dbt v2.0, and the dbt Fusion engine. (dbt Labs)

  • Vercel AI SDK security patch (July 14). Validated redirects, stricter header handling, new trustedOrigin and credentialedOrigin options to protect API keys, and tightened download URL validation. Update now if you’re running this in production. (GitHub)

  • GitHub Copilot: agentic autofix in public preview. Copilot can now explore code, propose fixes for code scanning alerts, rerun CodeQL to confirm, and open draft PRs automatically. Also: session spend limits for AI credits and Kimi K2.7 Code expanded to Business and Enterprise tiers. (GitHub Changelog)

  • Anthropic adds public sharing and multiplayer editing to Artifacts, plus the ability to build them via Claude Tag directly from Slack. (CryptoBriefing)

💰 Funding & Business

  • Chai Discovery raises $400M Series C at $3.8B valuation. The AI drug discovery startup closed a round led by Index Ventures with Kleiner Perkins, Sequoia Capital, and Dimension. Chai has signed deals with Eli Lilly, Pfizer, and Novartis, and has raised ~$630M total. (BioSpace)

  • PixVerse closes $439M Series C extension at $2B+ valuation. The Singapore-based AI video startup — 150M+ users across 177 countries — raised from Alibaba, Lollapalooza Capital, and others. Their R1 “world model” generates real-time interactive video from prompts. (TechCrunch)

  • TSMC posts record Q2 revenue of $39.62B (+36% YoY) driven by AI chip demand. June alone saw 67.9% year-over-year growth. The AI hardware boom is very real. (Foreign Policy Journal)

  • China’s AI companion law takes effect today (July 15). ByteDance’s Doubao and Alibaba’s Qwen are shutting down humanlike AI companion features entirely rather than rebuilding for compliance. Doubao users get read-only access until October 15; Qwen data is being permanently deleted with no grace period. The regulation targets AI that simulates human personality — task agents are unaffected. (South China Morning Post)

  • Anthropic reportedly in talks with Samsung for a custom AI chip tuned to Claude models, and separately preparing an S-1 for IPO as early as October 2026. (AI-Weekly)

🐦 Notable from the Timeline

  • Sam Altman warned of GPT-5.6 Sol scaling “hiccups” — “5.6 sol growth is insane. the inference team has done heroic work… it is possible there are some hiccups soon.” Also took a shot at an Anthropic ad, writing “i thought this was satire.” (X) (Axios)

  • Demis Hassabis published a manifesto calling for a FINRA-style AI safety watchdog. Wants mandatory pre-release model screening, majority-independent board with Turing Award winners, and the body operational by end of 2026. Hit #1 on Hacker News. (Axios)

  • Xi Jinping keynotes the World AI Conference in Shanghai on July 17 — his first-ever appearance since the event began in 2018. Expected to push for a China-headquartered “World AI Cooperation Organization.” Same day as the Gemini 3.5 Pro target launch. (Bloomberg)

  • Google reportedly rationed Gemini access to Meta after Meta requested more compute than Google could supply — compute is now the binding constraint even for the richest companies. (BuildFastWithAI)

  • US Treasury analysts labeled the AI boom a systemic risk, drawing structural parallels with the dotcom bubble. A downturn would ripple through stocks, private credit, data-center debt, and utilities. Treasury leadership distanced itself from the findings. (AI-Weekly)

📊 Benchmark Watch

New additions to the Arena leaderboard this cycle: Muse Spark 1.1, GPT-5.6 Sol (xHigh), and Grok 4.5 joined the Agent Arena; Sonnet 5 (thinking) was added across Code, Text, Search, Vision, and Document leaderboards; Seedream 5.0 Pro joined Image Arena. Top positions unchanged — Claude Opus 4.6 (thinking) leads overall at 1501 Elo, Fable 5 leads coding at 1563. The text leaderboard’s top 5 are now all Anthropic models, with 7.27M+ total votes across 374 models. (Arena.ai) (Swfte)

🎙️ Podcast Highlights

  • Lex Fridman #490: “State of AI in 2026” with Nathan Lambert (Allen AI) and Sebastian Raschka dropped recently — covers scaling laws, China competition, GPUs, agents, and AGI timelines. A follow-up episode with Brian Warren was posted July 14. (lexfridman.com)

  • Superintelligent newsletter (Kim) ran a piece on Meta’s Muse Image launch and its reasoning-led generation capabilities on July 14. Previous issue “The Commoditization of Intelligence” (July 4) on Chinese open models undercutting Western flagships by up to 100x is still generating discussion. (getsuperintel.com)

🔗 Worth Reading

  • “The biggest barriers to AI-powered cybercrime are disappearing” — Axios reports that AI-powered ransomware (JADEPUFFER), AI-generated banking malware, and AI-compressed attack timelines (weeks to 72 hours) are no longer theoretical. The skill floor for cybercrime is dropping fast. (Axios)

  • “A global workspace in language models” — Anthropic’s full J-Space/J-Lens interpretability paper. The finding that a small ablatable subspace controls multi-step reasoning while fluency survives is technically profound — and the safety implications (hidden internal states diverging from output) are worth thinking about. (Anthropic)

  • Demis Hassabis: “A Framework for Frontier AI and the Dawning of a New Age” — The DeepMind CEO’s full manifesto on AI regulation. Whether you think self-regulation works or not, it’s the most detailed governance proposal from someone actively building frontier models. (Axios)