🗞️ AI Daily Briefing — April 7, 2026

🔥 Top Story

The New Yorker drops a bombshell Sam Altman profile. Ronan Farrow and Andrew Marantz published an investigation today citing ~70 pages of internal memos compiled by Ilya Sutskever and Dario Amodei before Altman’s 2023 firing, alleging “a consistent pattern of lying.” This lands while OpenAI is in the middle of its $122B raise and just five days after acquiring TBPN — expect the governance question to dominate the week. (Semafor, Inc.)

🚀 Model & Research News

  • Gemini 3.1 Pro takes ARC-AGI-2 crown at 77.1%: Google DeepMind’s Gemini 3.1 Pro now leads the ARC-AGI v2 leaderboard at 0.771, the largest jump on the benchmark since launch. (llm-stats)
  • Google’s TurboQuant lands at ICLR 2026: A two-step KV-cache compression algorithm (PolarQuant rotation + Quantized Johnson-Lindenstrauss) that meaningfully cuts inference memory overhead. (devFlokers)
  • Anthropic’s MCP crosses 97M installs: Model Context Protocol is officially infrastructure — every major lab now ships MCP-compatible tooling. (crescendo.ai)
  • Energy-efficiency paper claims 100× cut: Researchers published an approach that reportedly slashes training/inference energy ~100× while improving accuracy. Worth scrutinizing the methodology before believing it. (ScienceDaily)
  • GPT-5.4 momentum continues: OpenAI’s March release is still posting record numbers — 75% on OSWorld-V (above human baseline) and 83% on GDPval. (llm-stats)

🛠️ Tools & Developer Updates

  • MCP everywhere: With 97M installs, the practical takeaway for builders — if you’re not exposing your tools/data via MCP yet, you’re now the outlier. (crescendo.ai)
  • DSPy stays the lightest framework: A fresh RAG framework benchmark pegs DSPy overhead at ~3.5ms vs LangChain ~10ms / LangGraph ~14ms. Useful ammo if you’re picking a stack. (aimultiple)

💰 Funding & Business

  • OpenAI’s $122B round: The April 2026 mega-round is still being papered — the largest private funding event in history, reshaping the entire VC stack. (Vertu)
  • Unconventional AI raises $475M seed: Naveen Rao’s two-month-old neuromorphic-computing startup co-led by a16z and Lightspeed, with Sequoia in. Bio-inspired hardware is back. (blog.mean.ceo)
  • Nexthop AI lands $500M Series B: AI-optimized networking infra round led by Lightspeed with a16z. The picks-and-shovels trade is still hot. (blog.mean.ceo)
  • OpenAI ↔ TBPN closes: Five days in, the acquisition is already the case study for “what happens when a model lab buys media.” Stratechery argues it’s a token-distribution play, not a content play. (Stratechery)

🐦 Notable from the Timeline

  • @sama: Conspicuously quiet today as the New Yorker piece circulates — no rebuttal yet on X.
  • @fchollet: Reframing ARC-AGI-3 as the “real” target now that frontier models are saturating v2 — highlighting that v3 requires online learning inside novel environments.
  • @DrJimFan: Riffing on NVIDIA’s National Robotics Week posts — physical-AI foundation models are “the next frontier where benchmarks lie.”
  • @arena: LMArena leaderboard updated today (April 7) — Grok-4.1 Thinking holds #1 at 1483 Elo. (Arena)
  • @hwchase17: Continued push on LangGraph as the “agent-runtime” layer rather than competing with DSPy on optimization.

📊 Benchmark Watch

  • LMArena (text): Grok-4.1 Thinking #1 at 1483 as of April 7. (llm-stats)
  • ARC-AGI v2: Gemini 3.1 Pro #1 at 0.771 as of April 5 — 13 models now evaluated. (llm-stats)
  • ARC-AGI-3 (Kaggle 2026): Open competition still wide-open — no model has cracked the interactive-learning regime yet. (Kaggle)

🎙️ Podcast Highlights

  • TBPN: First episodes under OpenAI ownership are out — Coogan & Hays leaning hard into the “editorial independence” framing. The Artemis II + Eddy Cue interview episode is the highlight. (TechCrunch)
  • Stratechery (Ben Thompson): “OpenAI Buys TBPN, Tech and the Token Tsunami” — the sharpest analysis of why this deal is really about distribution for ChatGPT, not media. (Stratechery)

🔗 Worth Reading