🗞️ AI Daily Briefing — 2026-04-14

🔥 Top Story

The OpenAI–Anthropic cold war turns personal — and physical. Within 72 hours: a second incident at Sam Altman’s SF home (two arrests Apr 13 after an apparent shooting, on the heels of the Apr 10 Molotov attack), a leaked OpenAI memo from CRO Denise Dresser instructing sales to go on the offensive against Anthropic and accusing them of “juicing” revenue numbers, and Altman personally slamming Anthropic’s Super Bowl-style ads as “clearly dishonest.” The backdrop: Anthropic is still ahead on the LMArena leaderboard, reportedly ahead on ARR, and now sitting on Mythos. OpenAI is pivoting from “we have the best model” to “they’re cheating.” (Gizmodo) (Variety) (Spokesman)

🚀 Model & Research News

  • Google ships Gemma 4 (1B / 13B / 27B, multimodal): Native vision with 2D spatial RoPE, pretrained + instruct variants, all open weights. Gemma-4-31B-NVFP4-turbo is already the production leader on HF trending. (Hugging Face blog) (SOO Group)
  • NVIDIA DreamDojo (open-sourced this week): Built on open-weight Cosmos; supports live VR teleop into a world model (demoed on Unitree G1 + PICO + one RTX 5090) and showed +17% real-world success on a fruit-packing task. Jim Fan is driving the “world models are the GPT-3 moment for robots” line hard. (NVIDIA GEAR) (the-decoder)
  • NVIDIA Open Physical AI Data Factory Blueprint: Companion release to DreamDojo — reference pipeline for turning sim + teleop + video into training data for robotics, vision agents, and AV. The “Android of robotics” strategy keeps widening. (NVIDIA Newsroom)
  • MiniMax-M2.7 & Music Flamingo land on HF: MiniMax continues to be the most interesting Chinese lab by ship-rate; Music Flamingo is the first serious open audio-language reasoning model. (SOO Group)
  • Neuro-symbolic agents are a thing again: Fresh arXiv survey + the “Ontology-Constrained Neural Reasoning” paper (600-run eval across 5 regulated industries) mark a clear return of symbolic scaffolding for enterprise agent reliability. (arXiv cs.AI)

🛠️ Tools & Developer Updates

  • Transformers v5 out: Simpler model definitions, faster default attention, first-class multimodal. This is the release Gemma 4 / MiniMax-M2.7 are built against. (HF blog)
  • Scale Labs expands: Scale’s SEAL becomes Scale Labs — post-training evals, enterprise deployment infra, risk oversight. Worth watching given Scale/Meta’s deepening ties. (Scale Labs)
  • dbt + Databricks + Nasdaq “AI-ready analytics” event (Apr 22–23): Jason Ganz teasing dbt’s pitch for governed data as the substrate for enterprise AI. (dbt AI)
  • Production RAG stack settles: Latest comparison pieces all converge on LlamaIndex (ingest/retrieval) + LangGraph (durable orchestration) + DSPy (prompt optimization) as the enterprise default. DSPy keeps winning latency benchmarks (~3.5ms overhead vs LangChain’s ~10ms). (Rahul Kolekar) (AIMultiple)

💰 Funding & Business

  • Eclipse closes $1.3B fund: Cerebras-backer raising for robotics, AI infra, manufacturing, defense. The “hard-tech AI” thesis keeps compounding. (Bloomberg)
  • Spirit AI raises ~$420M: Chinese generalist lab backed by Lei Jun and Jack Ma funds, two rounds in 30 days. A reminder the US–China frontier race isn’t a Deepseek-only story. (The AI Insider)
  • Q1 2026 is the largest VC quarter ever: $300B into ~6,000 startups; AI took $242B (80%). OpenAI’s $122B, Anthropic’s $30B, xAI’s $20B and Waymo’s $16B alone were 65% of global VC. (Crunchbase) (Crunchbase foundational AI)
  • PwC study: ~20% of companies are capturing ~75% of the economic gains from AI. The bifurcation thesis is now the consensus. (PwC)

🐦 Notable from the Timeline

  • @sama publicly accusing Anthropic of running “clearly dishonest” ads while defending his own ad plans — the tell is how thin-skinned the framing has become. (Altman on X)
  • @DrJimFan: “2026 is the year of World Models for physical AI.” DreamDojo + Cosmos + GR00T now form a fully open stack — rare posture for NVIDIA, clearly aimed at pulling roboticists off proprietary platforms.
  • @fchollet: ARC Prize 2026 ($2M, ARC-AGI-3) continues to hold the line — humans near 100%, frontier models in the low single digits. The “scaling ≠ reasoning” argument is winning on evidence this cycle. (ARC Prize) (the-decoder)
  • @alexandr_wang / Scale: Scale Labs framing (“evals + alignment + deployment infra”) reads like a quiet bid to own the Meta superintelligence team’s tooling spine.
  • @pmarca: Back on the “federal preemption or bust” drum after Maine, Maryland and Nebraska each shipped new AI-specific bills last week. (Troutman)

📊 Benchmark Watch

  • LMArena Text: Claude Opus 4.6 Thinking still #1 (~1504 Elo); Gemini 3.1 Pro within a handful of points; Grok-4.1 Thinking at 1483 holding #3–#4. Top six are within ~20 Elo — the densest the frontier has ever been. (Arena.ai) (LLM-Stats)
  • ARC-AGI-2: Gemini 3.1 Pro still holds 77.1% — the single biggest frontier gap vs. the rest of the pack on pure-reasoning evals.
  • MIT Tech Review’s state-of-AI chart pack (Apr 13) is the cleanest visual snapshot of the gap between benchmark capability and measurable economic impact. (MIT TR)

🎙️ Podcast Highlights

  • TBPN (post-OpenAI acquisition, ongoing): Coogan & Hays still running daily; recent eps have covered “token maxing” at Meta, Mythos security implications, and a long Thomas Laffont interview on AI/semiconductors investing. Editorial independence question is unresolved and worth watching. (Stratechery) (Slate)
  • All-In: Continues to stay close to the Altman “AI New Deal” arc (robot tax, sovereign wealth fund). Expect a sharper take once the attacks on Altman’s home hit the mainstream narrative. (All-In)

🔗 Worth Reading