↓Skip to main content
Zeitgeist — a spike by Chris Gathercole
  1. Creators/

Nathan Lambert — Interconnects

About #

ML researcher at the Allen Institute for AI (Ai2) who leads post-training work and has authored prominent open-model research including OLMo, Tulu 2/3, RewardBench, Zephyr-Beta, and the Open LLM Leaderboard, and wrote the first RLHF textbook. Through Interconnects he provides technical, insider-level analysis of frontier AI lab economics, training methods, and reasoning models, positioned against industry hype. A leading voice on the open-vs-closed model ecosystem, including geopolitical dynamics.

2026-06-22 — GLM-5.2 is the step change for open agents #

Newsletter · Read

  • Argues GLM-5.2 is the first open-weight model that credibly performs as a general-purpose coding agent, on par with closed-source competitors — a capability threshold rather than an incremental bump.
  • Technical notes: uses Z.ai’s SLIME RL framework, performs best at “Max thinking effort,” MIT-licensed weights on Hugging Face, and arrived ~6.8 months after Claude Opus 4.5 — release timing overlapped with Claude Fable 5 export restrictions.
  • Matches or exceeds proprietary models on agent-arena leaderboards against OpenAI and Anthropic’s latest offerings.
  • Practical takeaway: viable open alternative for coding workflows today (Lambert tested it via Fireworks API inside Claude Code); expect pricing pressure on closed-model providers and momentum for open-inference platforms (Fireworks, Together).
  • Flags a broader policy tension — growing open-agent capability strains assumptions about how advanced AI access can be controlled.

2026-06-28 — Latest open artifacts (#22): Zyphra, Cohere, and Poolside are expanding the breadth of the ecosystem #

Newsletter · Read

  • Core argument: the open model ecosystem is diversifying beyond a handful of Chinese labs into varied actors (frontier makers, big tech, specialized product builders), which strengthens it and makes restriction attempts both futile and counterproductive.
  • Round-up of releases: Nvidia Nemotron-3-Ultra-550B (LatentMoE, new OpenMDW weights license), Cohere Command A+ (218B-A25B MoE, shifted to Apache 2.0), GLM-5.2, and Zyphra ZAYA1 (74B-A4B and 8B-A0.6B, trained on AMD infra).
  • Categorizes participant motivations into three types — frontier model makers, big tech leveraging ecosystem effects, and product companies shipping specialized smaller models.
  • Practical takeaway: teams building on open models can now draw on a wider spread of architectures and training recipes, with real value in both frontier chasing and narrow domain-specific releases.

2026-07-12 — 6 months to live for open models #

Newsletter · Read

  • Core argument: open-source AI faces an existential regulatory threat within six months, driven by two overlapping fights — distillation restrictions and frontier-capability caps — that could permanently disadvantage open models.
  • Reports White House discussions of an executive order that would restrict open-weight models above a GPT-5.5-ish capability threshold.
  • Frames Anthropic’s anti-distillation push as regulatory capture dressed as safety policy: it detected foreign API scraping, restricted access, then lobbied government — without public technical evidence — while APIs (citing the Discord breach of Anthropic’s “Mythos”) aren’t inherently more secure than open weights.
  • Warns that regulation aimed at Chinese open models would collateral-damage the emerging US open-source inference/fine-tuning economy absent global coordination.
  • Practical takeaway: US labs (Microsoft, Meta) should ship competitive open models now to reframe the “China dominates open AI” narrative, and the open-source community needs to organize politically before the six-month window closes.

2026-07-20 — Kimi K3: The open-weights escalation #

Newsletter · Read

  • Core argument: Moonshot AI’s Kimi K3 marks a watershed — the open/closed frontier gap narrows from an estimated 6-9 months to roughly 3-5 months, escalating both the upside (diffusion) and downside (risk) of open AI.
  • Technical points: 2.8T-parameter MoE with Kimi Delta Attention and Attention Residuals, ~2.5x scaling-efficiency gain over Kimi K2 via a Stable LatentMoE framework activating 16 of 896 experts; ranks #2 on Vals AI, #3 on Artificial Analysis’s Intelligence Index (behind Claude Fable and GPT-5.6 Sol Max), #1 on Frontend Code Arena.
  • Argues open models are simultaneously “decelerationist” for frontier labs (margin pressure) and “accelerationist” for broader AI diffusion, with Chinese labs showing superior capital efficiency.
  • Practical takeaway: a several-month open/closed buffer is preferable policy ground to heavy-handed bans, which offer false security while merely delaying diffusion — independent evaluation capacity is the more urgent governance gap.

2026-07-22 — Open models recap: more on Kimi K3, Qwen 3.8, Xi’s WAIC speech, distillation, the open-closed gap, and what’s next #

Newsletter · Read

  • Core argument (co-written with Florian Brand): Chinese open-weight models (Kimi K3, GLM-5.2, Qwen) are closing the gap with Western frontier labs mainly through capital efficiency, focused teams, and better data practices — not primarily distillation of proprietary models.
  • Directly rebuts Ben Thompson’s claim that distillation becomes more impactful at the RL stage: grading millions of RL rollouts with an expensive API model “would be insanely expensive,” and literature doesn’t support “the strongest model” as automatically the best teacher — SFT distillation gives incremental, not transformative, gains.
  • Chinese-lab advantages cited: mid-20s researchers with singular focus, access to smuggled Nvidia chips plus domestic alternatives (Huawei Ascend), lower labor/compute costs enabling 1-2 month release cadences.
  • Practical takeaways: expect rapid third-party fine-tunes of Kimi K3’s “rough-edged” post-training; restricting Chinese open models domestically flips the cybersecurity asymmetry (defenders stop improving, attackers don’t); by year-end expect a US entrant into the “close competitor” tier, a possible Tencent surprise, and model-size scaling plateauing around ~3T parameters.
  • Flags a cybersecurity risk baked into the “6 months to live” thesis: banning open Chinese models domestically doesn’t stop attackers from using them, only stops defenders from improving with them.