Top all-time · #639
Thinking Machines Lab
Company blog · thinkingmachines.ai/blogaiscience
Thinking Machines Lab is the company blog at thinkingmachines.ai, writing about ai and science. Since 2025, 99 different Hacker News readers have linked it 116 times, most often “Defeating Nondeterminism in LLM Inference”. It is still writing: the last post was on 2026-07-31, and you can follow it by feed.
Rarely recommended · 5 of 40 sampled comments recommend it?
Why HN links it
HN readers link Thinking Machines Lab mainly for its post on LLM nondeterminism, which commenters cite to explain why temperature-0 inference is not deterministic. Other links point to its Tinker fine-tuning product and Inkling model.
Summarised by Claude from a sample of 40 HN comments that link here. Each comment is marked recommending, critical or neutral (citing without judging); how this works.
Connectionism: Research Blog by Thinking Machines Lab
Most-linked posts?
- Defeating Nondeterminism in LLM Inference44 people · 52 mentions · first linked 2025-09-10
- Interaction Models7 people · 7 mentions · first linked 2026-05-11
- Inkling: Our Open-Weights Model7 people · 7 mentions · first linked 2026-07-15
- On-Policy Distillation7 people · 7 mentions · first linked 2025-10-27
- Learning to Replicate Expert Judgment in Financial Tasks6 people · 7 mentions · first linked 2026-06-30
- Tinker by Thinking Machines5 people · 5 mentions · first linked 2025-10-01
- The Future Worth Building Is Human – Thinking Machines Lab5 people · 5 mentions · first linked 2026-07-10
- Tinker: General Availability and Vision Input4 people · 4 mentions · first linked 2025-12-12
- Announcing Tinker3 people · 3 mentions · first linked 2025-10-01
- Training LLMs to Predict World Events3 people · 3 mentions · first linked 2026-03-20
What Hacker News says?
It's certainly possible to write LLM inference on the GPU in a deterministic way, but it's somewhat nontrivial and trades off against performance, so by default most LLM inference engines aren't deterministic even at zero temperature. The classic post about that is https://thinkingmachines.ai/blog/defeating-nondeterminism-in...
Comment linking /blog/defeating-nondeterminism-in-llm-inference/
on “Discovery of a new OpenAI agent message board”Submitted Defeating Nondeterminism in LLM Inference as a story · 345 points
HN discussionYes and no - yes, in theory, but in practice, non-determinism can be introduced at different points along the stack. See Thinking Machines' post on LLM non-determinism: https://thinkingmachines.ai/blog/defeating-nondeterminism-in...
Comment linking /blog/defeating-nondeterminism-in-llm-inference/
on “Backing up Spotify”Submitted LoRA Without Regret as a story · 184 points
HN discussionLooks like it's probably https://thinkingmachines.ai/blog/defeating-nondeterminism-in...
Comment linking /blog/defeating-nondeterminism-in-llm-inference/
on “Claude Sonnet 4.5”Submitted Modular Manifolds as a story · 161 points
HN discussionWhy would LLMs need injected randomness to work effectively? Their nondeterminism isn't by design, it's largely a byproduct of how (batched) inference actually happens on hardware [0]. In case you are referring to the actual token sampling (be it temperature, Top-K or whatever), i wouldn't really see this as part of the LLM itself (And I've frequently seen benchmark sites run them at t=0 anyways) [0] https://thinkin…
Comment linking /blog/defeating-nondeterminism-in-llm-inference/
on “Yes, and”Submitted Announcing Tinker as a story · 152 points
HN discussionIt looks like true 0-temperature (i.e. determinism) will happen. Here's some good context: https://thinkingmachines.ai/blog/defeating-nondeterminism-in...
Comment linking /blog/defeating-nondeterminism-in-llm-inference/
on “Meta Ray-Ban Display”Submitted Murati's Thinking Machines Releases Open-Weights 975B Parameter LLM as a story · 79 points
HN discussionNot necessarily. This is a good blog post from a few days about it: https://thinkingmachines.ai/blog/defeating-nondeterminism-in...
Comment linking /blog/defeating-nondeterminism-in-llm-inference/
on “Meta Ray-Ban Display”Submitted Tinker by Thinking Machines as a story · 55 points
HN discussionhttps://thinkingmachines.ai/tinker/ https://github.com/thinking-machines-lab/tinker-cookbook (As a TML person, I'm obviously biased, but I couldn't resist because of "tinkering"). TBF it's hard to imagine a real architecture change that wouldn't require a ton of compute, but you could certainly fine tune and play with different recipes, loss functions, etc. And Claude can carry you a lot of the way through doing thi…
Comment linking /tinker/
on “I were 17, I'd learn how to build LLMs from scratch”Submitted Inkling: Our Open-Weights Model as a story · 50 points
HN discussionAFAIK the batch your query lands in can also matter[1]. Though I imagine this should be a smaller effect than different quantization levels say. [1]: https://thinkingmachines.ai/blog/defeating-nondeterminism-in...
Comment linking /blog/defeating-nondeterminism-in-llm-inference/
on “Improved Gemini 2.5 Flash and Flash-Lite”Submitted The Future Worth Building Is Human – Thinking Machines Lab as a story · 31 points
HN discussion> with temperature at zero, for the same input they produce always the same reply. Nonsense. https://thinkingmachines.ai/blog/defeating-nondeterminism-in...
Comment linking /blog/defeating-nondeterminism-in-llm-inference/
on “US Military had close call after using AI for hallucinated intelligence report”Submitted Thinking Machines Lab as a story · 30 points
HN discussionThe value of figuring out how to make their LLM serving deterministic might help them track this down. There was a recent paper about how the received wisdom that kept assigning it to floating point associativity actually overlooked the real reasons for non-determinism [1]. [1] https://thinkingmachines.ai/blog/defeating-nondeterminism-in...
Comment linking /blog/defeating-nondeterminism-in-llm-inference/
on “A postmortem of three recent issues”Submitted LoRA Without Regret as a story · 24 points
HN discussion> LLMs hallucinate because training on source material is a lossy process and bigger, LLMs hallucinate because they are probabilistic by nature not because the source material is lossy or too big. They are literally designed to create some level of "randomness" https://thinkingmachines.ai/blog/defeating-nondeterminism-in...
Comment linking /blog/defeating-nondeterminism-in-llm-inference/
on “ChatGPT terms disallow its use in providing legal and medical advice to others”Submitted On-Policy Distillation as a story · 6 points
HN discussionInteresting - this seems to target a different layer than services like Tinker ( https://thinkingmachines.ai/blog/announcing-tinker/ ). Monarch provides the infrastructure primitives while Tinker is a managed finetuning service. Could someone build something like Tinker on top of Monarch?
Comment linking /blog/announcing-tinker/
on “PyTorch Monarch”Submitted Interaction Models as a story · 5 points
HN discussionDidn't many of the missionaries at OpenAI go to Thinking Machines Lab? https://thinkingmachines.ai/
Comment linking /
on “Sam Altman Slams Meta’s AI Talent Poaching: 'Missionaries Will Beat Mercenaries'”Submitted Tinker: General Availability and Vision Input as a story · 5 points
HN discussionhttps://thinkingmachines.ai/blog/defeating-nondeterminism-in...
Comment linking /blog/defeating-nondeterminism-in-llm-inference/
on “Is it a bubble?”Submitted On-Policy Distillation as a story · 5 points
HN discussioncurrently LLMs from majorvproviders are not deterministic with temp=0, there are startups focusing on this issue (among others) https://thinkingmachines.ai/blog/defeating-nondeterminism-in...
Comment linking /blog/defeating-nondeterminism-in-llm-inference/
on “Is it a bubble?”Submitted Inkling-Small as a story · 4 points
HN discussion
