Hacker News: lewtun

New comment by lewtun in "AlphaEvolve: Gemini-powered coding agent scaling impact across fields"

lewtun — Thu, 07 May 2026 16:14:07 +0000

Shameless plug: https://huggingface.co/spaces/smolagents/ml-intern

It’s a simple harness around Opus, but with tight integration to Hugging Face infra, so the agent can read papers, test code and launch experiments

New comment by lewtun in "I Just Want Simple S3"

lewtun — Mon, 13 Apr 2026 21:57:03 +0000

Hugging Face Buckets are pretty simple: https://huggingface.co/docs/huggingface_hub/en/guides/bucket...

Disclaimer: I work at HF

New comment by lewtun in "When models manipulate manifolds: The geometry of a counting task"

lewtun — Mon, 03 Nov 2025 11:07:14 +0000

The analogy stems from the notion that neural nets are "grown" rather than "engineered". Chris Olah has an old, but good post with some specific examples: https://colah.github.io/notes/bio-analogies/

New comment by lewtun in "The Smol Training Playbook: The Secrets to Building World-Class LLMs"

lewtun — Sun, 02 Nov 2025 13:42:46 +0000

Thanks! I expect the book will remain relevant as long as the Transformers architecture does. That’s why we mostly focus on topics we think will stand the test of time, but let’s see how that plays out :)

New comment by lewtun in "The Smol Training Playbook: The Secrets to Building World-Class LLMs"

lewtun — Sun, 02 Nov 2025 06:14:35 +0000

In the specific case of SmolLM, it originates from the meme in this dataset https://huggingface.co/datasets/bigcode/the-stack-smol

New comment by lewtun in "The Smol Training Playbook: The Secrets to Building World-Class LLMs"

lewtun — Sat, 01 Nov 2025 21:50:11 +0000

Hi, Lewis here (one of the co-authors). Happy to answer any questions people have about the book :)

PyTorch OpenEnv

lewtun — Thu, 23 Oct 2025 15:49:13 +0000

Article URL: https://github.com/meta-pytorch/OpenEnv

Comments URL: https://news.ycombinator.com/item?id=45683252

Points: 1

# Comments: 0

Scaling Laws for Reinforcement Learning

lewtun — Thu, 16 Oct 2025 16:01:16 +0000

Article URL: https://huggingface.co/papers/2510.13786

Comments URL: https://news.ycombinator.com/item?id=45607058

Points: 1

# Comments: 0

New comment by lewtun in "LoRA Without Regret"

lewtun — Sat, 04 Oct 2025 21:00:09 +0000

For those interested in playing with an implementation of these ideas, my colleagues at HF made some recipes here: https://github.com/huggingface/trl/blob/main/docs/source/lor...

New comment by lewtun in "Quantum Mechanics, Concise Book"

lewtun — Sat, 06 Sep 2025 09:47:12 +0000

“QED and the Men Who Made It” [1] might be close to what you’re after for quantum theory at least. Unlike other popular accounts, it gets quite technical and covers a lot of the historical dead ends that people had during the development of quantum field theory.

[1] https://press.princeton.edu/books/paperback/9780691033273/qe...

New comment by lewtun in "Adaptive LLM routing under budget constraints"

lewtun — Mon, 01 Sep 2025 19:59:25 +0000

> We instantiate this idea through Preference-prior Informed Linucb fOr adaptive rouTing (PILOT), a novel extension of LinUCB

Academics are pretty creative at naming their creations

New comment by lewtun in "Smollm3: Smol, multilingual, long-context reasoner LLM"

lewtun — Tue, 08 Jul 2025 18:38:07 +0000

Indeed we opted for offline methods like Anchored Preference Optimization as we found in the Open R1 project that doing multi-task RL on small models is quite a hassle to get right. With offline methods, you focus much more on dataset curation / generation, but that still provides faster iteration cycles for the model scale we’re dealing with!

New comment by lewtun in "PDF to Text, a challenging problem"

lewtun — Wed, 14 May 2025 08:39:02 +0000

> The absolute best way of doing this is these days is likely through a vision based machine learning model, but that is an approach that is very far away from scaling to processing hundreds of gigabytes of PDF files off a single server with no GPU.

SmolDocling is pretty fast and the ONNX weights can be scaled to many CPUs: https://huggingface.co/ds4sd/SmolDocling-256M-preview

Not sure what time scale the author had in mind for processing GBs of PDFs, but the future might be closer than “very far away”

DESI results show dark energy may be evolving over time

lewtun — Thu, 20 Mar 2025 17:48:46 +0000

Article URL: https://newscenter.lbl.gov/2025/03/19/new-desi-results-strengthen-hints-that-dark-energy-may-evolve/

Comments URL: https://news.ycombinator.com/item?id=43426470

Points: 1

# Comments: 0

DocumentAI with 256M Parameters

lewtun — Thu, 20 Mar 2025 17:07:50 +0000

Article URL: https://huggingface.co/spaces/ds4sd/SmolDocling-256M-Demo

Comments URL: https://news.ycombinator.com/item?id=43425862

Points: 5

# Comments: 0

220k reasoning traces from DeepSeek-R1

lewtun — Tue, 11 Feb 2025 06:12:16 +0000

Article URL: https://huggingface.co/blog/open-r1/update-2

Comments URL: https://news.ycombinator.com/item?id=43009549

Points: 1

# Comments: 0

New comment by lewtun in "Show HN: A real time AI video agent with under 1 second of latency"

lewtun — Wed, 02 Oct 2024 06:05:38 +0000

I gave the demo a spin and it’s pretty nice! One thing I noticed is that the avatar doesn’t seem to be aware of it’s surroundings- for example, I asked it why it was wearing a cowboy hat and it was adamant that it wasn’t wearing a hat at all :)

New comment by lewtun in "RLHF is just barely RL"

lewtun — Thu, 08 Aug 2024 12:51:44 +0000

> I expect language models to also get crazy good at mathematical theorem proving

Indeed, systems like AlphaProof / AlphaGeometry are already able to win a silver medal at the IMO, and the former relies on Lean for theorem verification [1]. On the open source side, I really like the ideas in LeanDojo [2], which use a form of RAG to assist the LLM with premise selection.

[1] https://deepmind.google/discover/blog/ai-solves-imo-problems...

[2] https://leandojo.org/

The largest math dataset of Olympiad problems for training LLMs

lewtun — Sun, 21 Jul 2024 18:43:20 +0000

Article URL: https://huggingface.co/datasets/AI-MO/NuminaMath-CoT

Comments URL: https://news.ycombinator.com/item?id=41027167

Points: 3

# Comments: 0

New comment by lewtun in "[dead]"

lewtun — Thu, 11 Apr 2024 23:07:08 +0000

Hello everyone, we just did a speed run with Argilla and KAIST AI to fine-tune the beefy new Mixtral model with some new techniques that came out recently. More details in the model card - enjoy!