Interconnects (Nathan Lambert)

After a few long years of finding time to document my lessons from training open models, my post-training book is done!

Musings on model alignment, what determines safety, and where we go from here.

Scaling our curation and measurement of the open ecosystem.

Despite predictions of consolidation, the open model ecosystem continues to thrive. This roundup covers Thinking Machines' Inkling, Tencent's Hy3, Poolside's Laguna S2.1, DeepSeek-V4-Flash, and Moonshot AI's Kimi K3, along with several other releases, showing how open models are finding utility on the Pareto frontier.

In this podcast, Nathan and Florian discuss recent developments in open AI models, including the release of Kimi K3, Qwen's open-weight strategy, Xi Jinping's speech at WAIC supporting open source, the performance gap between open and closed models, and the distillation controversy. They delve into why Chinese models are performing well, the state of the US open model ecosystem, and predictions for the future.

An assessment of the open ecosystem and the motivations behind releasing models, highlighting the growing diversity of model makers and recent releases from NVIDIA, Cohere, Zyphra, Poolside, and others.

GLM-5.2, released by Z.ai, represents a significant leap for open-weight models, matching or exceeding closed-source models in agent and coding benchmarks. Its release amid the ban on Claude Fable highlights economic and geopolitical implications, sparking debates on open vs. closed models.

This article argues that banning or over-regulating open source AI would be a grave mistake. Open source software has been crucial for education, innovation, and competition, generating trillions in economic value. In AI, open source models provide a counterweight to monopolies and are more transparent and secure. Concerns about China should not lead to restrictions on open source; instead, support for domestic open source should be strengthened.

The author reflects on the blog Interconnects three years into weekly writing, discussing its role in their career goals, recent advising roles with Arcee AI and Mercor, and plans to evolve the blog's operations including paywalled comments and more paid articles to maintain a high-quality, niche audience.

This podcast dives into the evolution of post-training recipes, from InstructGPT to the 2026 multi-teacher on-policy distillation (MOPD) era. Nathan Lambert and Finbarr Timbers reflect on challenges in open-source models like OLMo-3 and analyze how frontier labs leverage specialized teachers and distillation to push performance boundaries.

One step further into the power politics of frontier AI systems.

Nathan Lambert reflects on his time at the Allen Institute for AI (Ai2), where he worked on the Olmo models and led projects like Tülu 3. He emphasizes the importance of open research and shares his journey from a relatively unknown researcher to a prominent voice in AI.

2026 continues to accelerate AI progress with open models lagging in agentic capabilities, Google's Gemini not yet competitive with Claude Code/Codex, American open models rising, a fierce competition between Anthropic and OpenAI, and power structures asserting control.

An eventful month with one flagship release after another. CAISI assessment shows open models lagging behind the US frontier, but methodology is questioned. Highlights include MiMo-V2.5-Pro, Gemma-4, Kimi-K2.6, Laguna-XS.2, and DeepSeek-V4-Flash.

The article explains that 80% of compute for frontier models is R&D, not final training. Open ecosystems like China's reduce duplicated R&D costs. Open models lower future development costs but not immediate deployment. The author argues for an open model consortium to sustain cost advantages.

An inside look at Chinese AI labs reveals a culture of humility, practical fast-following, and a focus on building rather than philosophical debates. Chinese researchers, many students, excel at meticulous LLM development with less ego, while the ecosystem lacks a developed data industry but shows early domestic AI demand.

The performance gap between open and closed models is nuanced and not captured by a single number. Benchmarks evolve, trust diminishes, and frontier labs face economic pressure to constantly innovate. Chinese open models are competitive but may focus more on benchmarks, while real-world robustness still favors closed models.

This post recaps the author's recent efforts including the updated ATOM Report, completion of the RLHF book, creation of a post-training lecture series, and involvement in two research papers.

This article analyzes the wave of fear surrounding open-weight AI models after the announcement of Claude Mythos. The author argues that the concerns are similar to past overblown fears and calls for nuanced study rather than a general ban.

The article explores the competitive landscape of open models in 2026, the key factors for their success (performance, provenance, license, tooling, finetunability), and analyzes Google's latest Gemma 4 series. It argues that success depends more on usability and ecosystem support than benchmark scores.

This issue covers a diverse range of open models spanning OCR, RAG search, audio transcription, computer use, code editing, math theorem proving, and more. Models come from a broader set of builders including NVIDIA, Cohere, Sarvam, Mistral, and others, highlighting the industry's push for domain-specific, cost-effective models.

The article argues that AI progress, while significant, is better described as 'lossy self-improvement' rather than recursive self-improvement. Frictions such as narrow automatable research, diminishing returns from parallel agents, and resource bottlenecks suggest a more linear trajectory rather than exponential takeoff.

Despite incremental benchmark gains, GPT 5.4 in Codex offers real improvements in usability, speed, and context management, though Claude still wins on charm.