Super interesting application of program search Goals are mapped to programs which are embedded in a latent space. A fitness metric is assigned to the programs and program search is done to synthesise new human-like goals.

Jorge Bravo Abad@bravo-abad.bsky.social · last yr.

Davidson et al. introduce a model that learns human-like goals by representing them as reward-producing programs and generating new, creative “games.” www.nature.com/articles/s42... & arxiv.org/abs/2405.13242

One of my big motivations is accelerating science with AI. Every discovery project had a beautiful aha moment, such as the structure of antibiotics emerging in the latent space of a model or a GFlowNet proposing new carbon capture materials. Here's some of the threads I've wrote on this topic.

Wanna try to guess which of those gets parsed as a string and which as a number? Answer in alt text. YAML parsing in python is weird.

Interesting idea to generate responses using diffusion rather than left-to-right auto-regressive models

Alexander Doria@dorialexander.bsky.social · 2y ago

One of the most interesting new research in the LLM space: introducing Large Language Diffusion Models. arxiv.org/pdf/2502.09992

What is large for a language model? Is it 400B, 70B or maybe 1T? I think focus on raw number of parameters is a less useful frame than thinking about inference speed, cost and location of inference (on-device vs cloud).

More open reasoning datasets and distilled models. It's great to see the energy of the community that got unleashed after open models that generate chains of thought!

Alex Dimakis@alexdimakis.bsky.social · 2y ago

We are releasing OpenThinker-32B, the best 32B reasoning model with open data. We match or outperform Deepseek-R1-32B (a closed data model) in reasoning benchmarks. Congrats to Negin and the whole Open Thoughts team. github.com/open-thought...

ColabFit Exchange is another great dataset curation effort that I'd like to boost. Great work by @stemartiniani.bsky.social and team to curate the most diverse materials database in the world!

Stefano Martiniani@stemartiniani.bsky.social · 2y ago

Join us for the #AI4Mat workshop at #NeurIPS2024 today and check out our spotlight on how we built the most diverse database for AI for materials in the world openreview.net/forum?id=b8q...

Neat idea! Fine-tuning using majority voting and length filtering generalises a model's capabilities. Models generalise to slightly harder versions of a problem, and the correct answers are used to bootstrap the next model and the next one and so on.

mr. TIM@timkellogg.me · 2y ago

Self-Improving Transformers They found that you can train LLMs on their own outputs by 1. generating *slightly harder* problems each time 2. filtering low quality via majority voting Is this the singularity? Maybe, but I think it might just be benchmark saturation. timkellogg.me/blog/2025/02...

Join us in creating open datasets, benchmarks and leaderboards for materials discovery.

Entalpic@entalpic.bsky.social · 2y ago

🚀 𝐋𝐞𝐌𝐚𝐭𝐞𝐫𝐢𝐚𝐥 𝐜𝐨𝐦𝐦𝐮𝐧𝐢𝐭𝐲 𝐦𝐞𝐞𝐭𝐢𝐧𝐠𝐬 𝐚𝐫𝐞 𝐥𝐢𝐯𝐞!⁣⁣ 𝖩𝗈𝗂𝗇 𝗎𝗌 𝐞𝐯𝐞𝐫𝐲 𝐬𝐞𝐜𝐨𝐧𝐝 𝐓𝐡𝐮𝐫𝐬𝐝𝐚𝐲 𝐨𝐟 𝐭𝐡𝐞 𝐦𝐨𝐧𝐭𝐡 𝗍𝗈 𝖾𝗑𝗉𝗅𝗈𝗋𝖾 𝖠𝖨-𝖽𝗋𝗂𝗏𝖾𝗇 𝗆𝖺𝗍𝖾𝗋𝗂𝖺𝗅𝗌 𝖽𝗂𝗌𝖼𝗈𝗏𝖾𝗋𝗒.⁣⁣ 📅 𝐅𝐞𝐛 𝟏𝟑 | 𝟔𝐏𝐌 𝐏𝐚𝐫𝐢𝐬 | 𝟗𝐀𝐌 𝐋𝐀⁣⁣ 📍 𝐉𝐨𝐢𝐧 𝗁𝗍𝗍𝗉𝗌://𝗆𝖾𝖾𝗍.𝗀𝗈𝗈𝗀𝗅𝖾.𝖼𝗈𝗆/𝗆𝗐𝗒-𝗎𝗒𝖽𝖽-𝗄𝗏𝖿⁣ 𝖣𝗂𝗏𝖾 𝗂𝗇𝗍𝗈 𝖫𝖾𝖬𝖺𝗍𝖾𝗋𝗂𝖺𝗅 & 𝗌𝗁𝖺𝗉𝖾 𝗍𝗁𝖾 𝖿𝗎𝗍𝗎𝗋𝖾!⁣⁣ 👉 𝐂𝐡𝐞𝐜𝐤 𝐭𝐡𝐞 𝐜𝐨𝐦𝐦𝐞𝐧𝐭𝐬 𝐭𝐨 𝐣𝐨𝐢𝐧 𝐭𝐡𝐞 𝐋𝐞𝐌𝐚𝐭𝐞𝐫𝐢𝐚𝐥 𝐒𝐥𝐚𝐜𝐤

Superb work!

Alexander Doria@dorialexander.bsky.social · 2y ago

Announcing the release of Common Corpus 2. The largest fully open corpus for pretraining comes back better than ever: 2 trillion tokens with document-level licensing, provenance and language information. huggingface.co/datasets/Ple...

The most durable motivation for research is curiosity, the desire to answer a question or understand something. Curiosity then leads you down a maze of existing answers and new questions. Eventually, you get to one that has no answer and then you start pushing at the frontier.

Interesting - 57.1% AIME24 and 94.8% MATH performance achieved using only 817 reasoning chains and STF. Adds more weight to the hypothesis that correct reasoning chains and SFT can lead to strong reasoning performance.

GitHub - GAIR-NLP/LIMO: LIMO: Less is More for Reasoning

LIMO: Less is More for Reasoning. Contribute to GAIR-NLP/LIMO development by creating an account on GitHub.

github.com

I've been reflecting today about OpenAI's five levels to measure progress in AI. GPT-4 was at Level 1, conversational AI: a model competent at 0.1-1s tasks, like holding a conversation. O1 / R1 reached Level 2, reasoners: a model solving 1-10min tasks such as basic coding tasks and math.

Agreed, we have the minimum viable scene. We now just need to amplify each other and keep going.

Marc Lanctot@sharky6000.bsky.social · 2y ago

I have been in your place several times since 2022 waiting patiently. What I can say is this recent wave (from Sept 2024 to now) has been absolutely the most successful. I think we have the nucleus. Now, we persist. Don't give up and keep contributing. We have to play the long game.

What if inference scaling is as simple as response.replace("</think>", "Wait")

mr. TIM@timkellogg.me · 2y ago

s1: Simple inference-time scaling This is a simple small-scale replication of inference-time scaling It was cheap: 16xH100 for 26 minutes (so what, ~$6?) It replicates inference-time scaling using SFT only (no RL) Extremely data frugal: 1000 samples arxiv.org/abs/2501.19393

SWE arena is going to be an interesting leaderboard to watch. It allows people to compare the code generated by LMs based on runs inside a sandbox.

every time i try uv, I'm more impressed. seems now like a tool that Just Works, reducing the complexity of the python ecosystem installed a cuda+torch+git packages and it all felt basically instant

DeepSeek-R1 has turned into such a Rorschach test for the collective psyche

Indeed, not outsourcing reasoning is an important value to ... well... reason about. How would we achieve this? It may require many individuals and groups to do RL on their own models, using their own verifiers. This may look like grading exams - not of students, but of ML models.

Ted Underwood@tedunderwood.com · 2y ago

I definitely start from default pessimism on this. But just to look at the other pan of the scales: we could plausibly justify outsourcing CMS and email. But if we fully outsource reasoning ... that's it, game over, everyone can go home. So it *should* be easier to get faculty to care about this.

Seeing A Film for the Future in 360 was a special experience. One of the most powerful parts was We Pray. The video and music match so well, hit hard, and resonate strongly with the times.

Coldplay - WE PRAY (A Film For The Future)

YouTube video by Coldplay

youtu.be

Turning the temperature up using R1 Starting to think gibberish gibberish gibberish Focus again. Calm up. 🤣

Hm, using reasoning models really feels qualitatively different (using @openrouter.bsky.social for inference). It's fun to see these aha moments and it'd be interesting to understand whether their presence helps.

Huh, interesting, Claude 3.5 sonnet seems to do hidden CoT in the app. Could not reproduce with the API tho.

Pekka Lund@pekka.bsky.social · 2y ago

Claude's internal hidden thinking steps were discovered at least 7 months ago.

Deepseek-R1 thread to gather thoughts and reactions Nice to see the technical details and MIT license for something that looks at o1 level 🥳

GitHub - deepseek-ai/DeepSeek-R1

Contribute to deepseek-ai/DeepSeek-R1 development by creating an account on GitHub.

github.com

Interesting result re evolutionary algos for inference time search

Sung Kim@sungkim.bsky.social · 2y ago

Google's Evolving Deeper LLM Thinking Provides an insightful benchmark on natural language planning tasks. Paper Link: arxiv.org/abs/2501.09891

A powerful feature of the AT Proto is that it uses domain names as the handle. If I find an interesting blog post, I can often find the author by searching for the domain name!

More details on how apple uses homomorphic encryption to search over photos in this great write-up by @boehs.org

Homomorphic Encryption in iOS 18

A mathematical miracle enables Apple's servers to process your photos while never knowing anything about them

boehs.org

Flaviu Cipcigan@flaviucipcigan.bsky.social · 2y ago

Comparing vectors of landmarks with a remote database is the first product use of homomorphic encryption I've heard of. It's a good one! Privacy-preserving RAG with local LLM and remote documents could be done in a very similar way.

A great abstraction "allows us to operate as if the underlying complexity simply does not exist". For example, TCP turns an unreliable protocol into a reliable communication channel. Otoh, indirections are justified in terms of modularity, yet often just add needless cognitive overhead.

fhur

fhur's blog

fhur.me

A critique I hear often of LLMs is that they don't have a notion of truth, that they are BS machines, in Frankfurt's sense. I don't think that's quite right. Here's two papers that helped me have a more nuanced view of this question.

There's a lot of enthusiasm in the community about transformers trained on chemical or biological data. Here's some interesting results and some thoughts on future directions.

mr. TIM@timkellogg.me · 2y ago

this is nuts a new 7B llama-style LLM for embedding of genomes & detection of pathogens in wastewater i’ve had a hunch that LLMs could lead to some big bio breakthroughs, since it feels like genes & proteins are a lot like a language