Top all-time · #829
Nathan Lambert
Personal blog · interconnects.aiai
Nathan Lambert writes Interconnects AI, a personal blog at interconnects.ai about ai. Since 2023, 79 different Hacker News readers have linked it 141 times, most often “Notes from inside China's AI labs”. It is still writing: the last post was on 2026-10-30, and you can follow it by feed.
Too few sampled comments (24) to say how often HN recommends it.?
Why HN links it
HN readers link Nathan Lambert's Interconnects posts for technical explanations of reasoning models, RL for LLMs, DeepSeek's costs and distillation, and as a source on open models.
- Recommended for · 4 comments
- One commenter calls the Lambert breakdown more measured and detailed and instructive; another says the discussion of actual DeepSeek cost was helpful; another c
Summarised by Claude from a sample of 24 HN comments that link here. Each comment is marked recommending, critical or neutral (citing without judging); how this works.
The cutting edge of AI, from inside the frontier AI labs, minus the hype. The border between high-level and technical thinking. Read by leading engineers, researchers, and investors. Click to read Interconnects AI, a Substack publication with tens of thousands of subscribers.
Most-linked posts?
- Notes from inside China's AI labs5 people · 5 mentions · first linked 2026-05-08
- The American DeepSeek Project4 people · 4 mentions · first linked 2025-07-04
- /p/openais-o1-using-search-was-a-psyop3 people · 4 mentions · first linked 2024-12-22
- The current balance of power in open models3 people · 3 mentions · first linked 2026-09-21
- DeepSeek V3 and the cost of frontier AI models3 people · 3 mentions · first linked 2025-01-25
- DeepSeek R1's recipe to replicate o1 and the future of reasoning LMs2 people · 4 mentions · first linked 2025-02-02
- Kimi K3: The open-weights escalation2 people · 3 mentions · first linked 2026-07-20
- Open-source AI and open models reading list2 people · 2 mentions · first linked 2026-09-13
- GPT-4o-mini changed ChatBotArena2 people · 2 mentions · first linked 2024-08-02
- /p/openais-o3-the-2024-finale-of-ai2 people · 2 mentions · first linked 2024-12-22
What Hacker News says?
Something that's really hitting me is something brought up in this piece: https://www.interconnects.ai/p/gpt-5-and-bending-the-arc-of-... When a model comes out, I usually think about it in terms of my own use. This is largely agentic tooling, and I mostly us Claude Code. All the hallucination and eval talk doesn't really catch me because I feel like I'm getting value of these tools today. However, this model is not…
Comment linking /p/gpt-5-and-bending-the-arc-of-progress
on “GPT-5”Submitted A look at Apple's technical approach to AI including core model performance etc. as a story · 196 points
HN discussionFor context: R1 is a reasoning model based on V3. DeepSeek has claimed that GPU costs to train V3 (given prevailing rents) were about $5M. The true costs and implications of V3 are discussed here: https://www.interconnects.ai/p/deepseek-v3-and-the-actual-co...
Comment linking /p/deepseek-v3-and-the-actual-cost-of
on “DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via RL”Submitted GPT-4.5: "Not a frontier model"? as a story · 166 points
HN discussionEh, not necessarily, many of the RL vendors sell to Chinese labs, too: https://www.interconnects.ai/p/glm-53-how-chinese-labs-keep-...
Comment linking /p/glm-53-how-chinese-labs-keep-stride
on “Kimi K3.1 Pro and K3.1 Pro Highspeed”Submitted Open-source AI and open models reading list as a story · 162 points
HN discussionSee https://www.interconnects.ai/p/why-reasoning-models-will-gen... for more information.
Comment linking /p/why-reasoning-models-will-generalize
on “An analysis of DeepSeek's R1-Zero and R1”Submitted The current balance of power in open models as a story · 132 points
HN discussionThis was exactly the crux of Nathan Lambert's recent testimony to a group of US Congressional members/staff: https://www.interconnects.ai/p/the-current-balance-of-power-...
Comment linking /p/the-current-balance-of-power-in-open
on “Ember-1”Submitted Artificial Analysis v5 and the benchmarker's dilemma as a story · 73 points
HN discussionThat quote is straight up wrong to claim Deepseek has "loathed to reveal their secret sauce". The source of all the excitement is exactly how much they revealed, and I feel like that thread as a whole emphasizes why people who aren't deeply familiar with the pipeline should not get to define these things. "Nick" claims the secret sauce is in something not provided, then posts an article that demonstrates the exact "…
Comment linking /p/deepseek-r1-recipe-for-o1
on “Why DeepSeek had to be open source”Submitted One resignation turned the embers of AI fear into a wildfire as a story · 45 points
HN discussionThat is how reasoning models work: https://www.interconnects.ai/p/deepseek-r1-recipe-for-o1
Comment linking /p/deepseek-r1-recipe-for-o1
on “O3 beats a master-level GeoGuessr player, even with fake EXIF data”Submitted DeepSeek V3 and the cost of frontier AI models as a story · 32 points
HN discussionI am fairly optimistic about LLMs as a human math -> theorem-prover translator, and as a fan of Idris I am glad that the AI community is investing in Lean. As the author shows, the answer to "Can AI be useful for automated mathematical work?" is clearly "yes." But I am confident the answer to the question in the headline is "no, not for several decades." It's not just the underwhelming benchmark results discussed in…
Comment linking /p/openais-o3-the-2024-finale-of-ai
on “Can AI do maths yet? Thoughts from a mathematician”Submitted DeepSeek R1's recipe to replicate o1 and the future of reasoning LMs as a story · 22 points
HN discussionMy point is simply this: Chinese labs going to great lengths to distill Claude is evidence that distillation is useful but not that that is what makes them so good. That one doesn't even need them to admit it. Nathan Lambert has made this distinction explicitly: he thinks Chinese labs likely innovate heavily on distillation, while saying he wouldn't call it a crucial factor in their post-training capabilities, in pa…
Comment linking /p/how-much-does-distillation-really
on “Corporate America is getting hooked on open-source AI”Submitted Q* Hypothesis: Enhancing Reasoning, Rewards, and Synthetic Data as a story · 11 points
HN discussionWhether or not it's "noise" might depend on if you think Chains of Thought are causally relevant? I liked these pieces if you want to read more about CoT / O1: O1 Technical Primer: https://www.lesswrong.com/posts/byNYzsfFmb2TpYFPW/o1-a-techn... Using Search Was a Psyop: https://www.interconnects.ai/p/openais-o1-using-search-was-a... Value Attribution: https://www.lesswrong.com/posts/FX5JmftqL2j6K8dn4/shapley-va...
Comment linking /p/openais-o1-using-search-was-a-psyop
on “Watch R1 "think" with animated chains of thought”Submitted The American DeepSeek Project as a story · 6 points
HN discussionhttps://www.interconnects.ai/p/openais-o1-using-search-was-a...
Comment linking /p/openais-o1-using-search-was-a-psyop
on “Emerging reasoning with reinforcement learning”Submitted People use AI more than you think as a story · 6 points
HN discussionThere's some more commentary on their open-ness in this blog too https://www.interconnects.ai/p/olmo
Comment linking /p/olmo
on “OLMo: Accelerating the Science of Language Models [pdf]”Submitted Llama 4: Did Meta just push the panic button? as a story · 6 points
HN discussionThis may be useful: https://www.interconnects.ai/p/deepseek-r1-recipe-for-o1 but the tl;dr of the idea is that we can use reinforcement learning on a strong base model (i.e. one that hasn't been fine tuned) to elicit the generation of tokens that help the model reach a result that can be verified to be correct. That is, if we have a way of verifying that a specific output is correct, the model can be trained to cons…
Comment linking /p/deepseek-r1-recipe-for-o1
on “Simple Explanation of LLMs”Submitted GLM-5.2 is the step change for open agents as a story · 5 points
HN discussionThe big question is whether or not o3 is using any type of “meta-generation” algorithm at inference time, I.e are there multiple invocations of the LLM generation at all, or does it generate an insanely long reasoning trace in a single autoregressive stream that some somehow implicitly has search-like behavior? In other words, is the search-like behavior learned entirely in post-training and only implicitly exhibite…
Comment linking /p/openais-o1-using-search-was-a-psyop
on “Open source inference time compute example from HuggingFace”Submitted Claude 4 and Anthropic's bet on code as a story · 5 points
HN discussionJust search for "System 1 / System 2" and add word "researcher" there, filter by time <1y< and you'll find plenty. OP is right that a bunch of people made far-going predictions, & now doing mental gymnastics to justify they were right all along. Most notably, LeCun https://www.threads.net/@yannlecun/post/DD0ac1_v7Ij and Chollet https://www.youtube.com/watch?v=w9WE1aOPjHc Some argue they were kind of tricked into thi…
Comment linking /p/openais-o1-using-search-was-a-psyop
on “Reasoning models are just LLMs”Submitted Reverse Engineering OpenAI's O1 as a story · 5 points
HN discussion
