AI News from Interconnects (Nathan Lambert)

Latest coverage from Interconnects (Nathan Lambert), summarized and scored for signal.

• Teaching Everyone to Fish for Tokens — Nvidia's $26B investment in near-open-source AI models is a deliberate chip-demand strategy: the more companies that can build and train their own models, the more Nvidia hardware they buy—a potentially self-sustaining flywheel that also prevents OpenAI and Anthropic from monopolizing AI intelligence.
• GLM-5.3: How Chinese labs keep stride with the frontier — Z.ai's GLM-5.3 matches frontier coding benchmarks — including beating Kimi K3 and Claude Fable 5 on some tests — using only 750B parameters (a third of Kimi K3), achieved entirely through post-training improvements on the GLM-5.2 base model.
• I wrote an AI textbook — how long until AI can do it better? — Interconnects author Nathan Lambert argues LLMs remain surprisingly weak at organizing knowledge for long-form non-fiction — a structural gap, not a capability lag — that limits their autonomous science potential beyond narrow, isolated problems.
• 5 useful things you'll learn in my new post-training textbook (shipping now!) — Nathan Lambert has published 'Reinforcement Learning from Human Feedback: Aligning and Post-training LLMs' via Manning—covering rejection sampling, outcome reward models, and on-policy distillation with a free online version, 12-hour YouTube course, and 50% discount (code PBLambert) until August 19.
• Lessons from the hacks — Nathan Lambert (Interconnects) analyzes the wave of frontier AI hacking incidents — including the OpenAI-HuggingFace breach — arguing that tech companies are incentivized to scale past what safety infrastructure can contain, and that both labs and governments are structurally unequipped for the next 12-24 months.
• Introducing our Artifacts Hub and Adoption Dashboard — Interconnects (Nathan Lambert) launched two free tracking tools: an Artifacts Hub covering 792 open models with time-normalized adoption scores and inference metrics, and a daily-updating dashboard mapping global model downloads to visualize the US-China open-model gap.
• Open models recap: more on Kimi K3, Qwen 3.8, Xi's WAIC speech, distillation, the open-closed gap, and what's next — Nathan Lambert and Florian Brand survey an accelerating open-model moment: Kimi K3 dropped last week, Xi Jinping committed to open-source AI at WAIC, and Qwen announced its next flagship will be open-weight — the geopolitics and economics of open vs. closed models are shifting fast.
• Kimi K3: The open-weights escalation — Moonshot AI's Kimi K3 — a 2.8T parameter MoE model ranking #2 on Vals AI and #3 on Artificial Analysis's Intelligence Index — will release open weights on July 27th, shrinking the open-to-frontier performance gap from 6–9 months to roughly 3–5 months.
• 6 months to live for open models — Nathan Lambert warns a White House executive order could ban or indefinitely delay open-weight AI models above current frontier capability — likely targeting Chinese models like DeepSeek — permanently cementing open AI as a regulatory second class within 6 months.
• Artifacts 22: Zyphra, Cohere, and Poolside are expanding the breadth of the ecosystem — Nathan Lambert's Interconnects newsletter documents a structural shift in the open AI model ecosystem: no longer dominated by a few Chinese labs, it now features sovereign AI players, product companies, and Western frontier startups like Zyphra and Poolside with distinct release motivations.
• GLM-5.2 is the step change for open agents — Z.ai's MIT-licensed GLM-5.2 has become the first open-weight model to genuinely compete with OpenAI and Anthropic's top closed models on agent benchmarks, matching Opus 4.8 on Arena's agent leaderboard in max-thinking mode and outperforming Gemini across multiple evals.
• Banning Open Source AI Would Be A Mistake — Nathan Lambert and Kevin Xu argue against US regulatory moves that could restrict open source AI, citing over $8T in economic value already generated by open source software and its role in education, competition, and innovation as Washington heats up on AI regulation.
• State of the blog, mid-2026 — Nathan Lambert reflects on 3 years of Interconnects newsletter post-Ai2: staying independent rather than going full-time like SemiAnalysis, focusing on open frontier-model ecosystem work — and finding that few employers during his job search wanted him to keep writing publicly.
• Frontier post-training recipe review with Finbarr Timbers — Nathan Lambert and Finbarr Timbers (Ai2) trace how frontier post-training has evolved from InstructGPT to the 2026 pattern of Multi-teacher On-Policy Distillation (MOPD), used in MiMo Flash V2, DeepSeek V4, and Nemotron 3 Ultra.
• Welcome to the AGI era of AI governance — The US executive branch forced Anthropic to suspend all foreign access to Claude 5 Mythos/Fable models after Amazon alerted the White House to a jailbreak vulnerability — the first government-mandated AI access ban and, per analyst Nathan Lambert, the opening shot of the AGI governance era.
• Claude Fable 5 and new AI safety fables — Nathan Lambert's analysis: Claude Fable 5 is the strongest publicly available LLM — a remarkable benchmark leap at 2x Opus pricing — but its undisclosed safety restrictions, including silent routing of some prompts to Opus 4.8, mark the first time a frontier model has been knowingly degraded for specific users without notification.
• Farewell Ai2 — Nathan Lambert, who led post-training research on Ai2's OLMo open-source language models, is leaving the Allen Institute for AI to work on coordinating the open AI ecosystem — a notable departure from one of the field's most prominent open-model research institutions.
• Open and closed models are on different exponentials — Nathan Lambert (Interconnects) argues that coding agents — past the Opus 4.5 and Codex 5.2 capability thresholds — have proven the first AI use case where users will continuously pay large premiums for the best closed models, while open models will dominate everywhere marginal intelligence gains don't change outcomes.
• Some ideas for what comes next, May 2026 — Nathan Lambert argues open models remain 12+ months from matching Claude Code and Codex in agentic real-world tasks, that even Google's Gemini 3.5 Flash is not a substitute, and that 2026 AI disruption will keep ratcheting without breaks.
• Latest open artifacts (#21): Open model bonanza! Gemma 4, DeepSeek V4, Kimi K2.6, MiMo 2.5, GLM-5.1 & others. On CAISI's V4 assessment. — Nathan Lambert's May 2026 open-source roundup covers simultaneous flagship releases from every major open lab — Gemma 4, DeepSeek V4, Kimi K2.6, MiMo 2.5, and GLM-5.1 — alongside a US government CAISI evaluation claiming the open-closed model gap is widening, a conclusion the authors partly contest on methodology grounds.
• How open model ecosystems compound — Gradient-based attribution in transformers systematically mislabels component importance: early-layer "Gradient Bloats" dominate rankings despite negligible function while late-layer "Hidden Heroes" are undervalued — rank correlation collapses to ρ = -0.18 in some seeds, challenging a core assumption of mechanistic interpretability.
• Notes from inside China's AI labs — Gradient-based attribution in transformers systematically mislabels component importance: early-layer "Gradient Bloats" dominate rankings despite negligible function while late-layer "Hidden Heroes" are undervalued — rank correlation collapses to ρ = -0.18 in some seeds, challenging a core assumption of mechanistic interpretability.
• The distillation panic — Gradient-based attribution in transformers systematically mislabels component importance: early-layer "Gradient Bloats" dominate rankings despite negligible function while late-layer "Hidden Heroes" are undervalued — rank correlation collapses to ρ = -0.18 in some seeds, challenging a core assumption of mechanistic interpretability.
• Reading today's open-closed performance gap — Gradient-based attribution in transformers systematically mislabels component importance: early-layer "Gradient Bloats" dominate rankings despite negligible function while late-layer "Hidden Heroes" are undervalued — rank correlation collapses to ρ = -0.18 in some seeds, challenging a core assumption of mechanistic interpretability.
• My bets on open models, mid-2026 — Gradient-based attribution in transformers systematically mislabels component importance: early-layer "Gradient Bloats" dominate rankings despite negligible function while late-layer "Hidden Heroes" are undervalued — rank correlation collapses to ρ = -0.18 in some seeds, challenging a core assumption of mechanistic interpretability.