Are you into Chip Design, EDA or just like to do RTL code for fun? Check out the largest benchmarking of LLMs for Verilog generation: TuRTLe 🐢 It includes 40 open LLMs evaluated on 4 benchmarks, following 5 tasks. And its only growing! huggingface.co/spaces/HPAI-... arxiv.org/abs/2504.01986

So many healthcare LLMs, and yet so little information! Check out this table summarizing contributions, and find more details in our latest pre-print: arxiv.org/abs/2505.04388

The Aloe Beta preprint includes full details on data & training setup. Plus four different evaluation methods (including medical expert). Plus a risk assessment of healthcare LLMs. Two years of work condensed in a few pages, figures and tables. Love open research! huggingface.co/papers/2505....

We just opened two MLOps Engineer positions at @bsc-cns.bsky.social Our active and young research team needs someone to help sustain and improve our services, including HPC clusters, automated pipelines, artifact managements and much more! Are you up for the challenge? www.bsc.es/join-us/job-...

Last week our team presented this at NAACL. Check out the beautiful poster they put together 😍

Working on a project for evaluating embryo quality using in-vitro fertilization data. A random forest using morphokinetic features of embryo evolution visually annotated by experts, and a CNN directly using static images get similar performance. Separately AND together. I find it surprising...

There are quite a lot of researchers who a so preoccupied with whether or not they could get the funding, they don't stop to think if they should. Being chased by dinosaurs and writing grants. Same thing.

How expensive 🫰 is it to get the best LLM performance? How much cash needs to burn 💸 to get reliable responses? Pareto optimal plots answer these questions. Our research shows it is economically feasible and scalable to achieve O1 level performance at a fraction of the cost. buff.ly/ji1VHiV

Our LLM safety project, Egida, reached 2K downloads 😀 It includes +60K safety questions expanded with jailbreaking prompts. The four models trained (and released) show strong signs of safety alignment and generalization capacity. Check out the 🤗 HF page and the paper for details! buff.ly/kxFVyl2

Today we release the TuRTLe leaderboard! 🐢 Are you in the Chip Design or EDA business? Wanna know which LLMs are best for the task? By integrating 4 benchmarks, TuRTLe evaluates: * Syntax * Functionality * Synthesizability * Power, Performance and Area metrics huggingface.co/spaces/HPAI-...

MIR is Spain's medical entrance exam. Best students reach an estimate accuracy of +90. Two or three every year. We took MIR, '20-'24 to test open LLMs. Llama 3.1 based models, like Aloe, reach +80 in accuracy. Deepseep R1 reaches +88. Boosted by a RAG system, 92. buff.ly/4bbbXMw buff.ly/4hLrhBV

After listening to the latest @fallofcivspod.bsky.social episode about the Mongolian Empire, by @paulcooper34.bsky.social , I realized Mongols and the Fremen from Dune share remarkable similarities. Skilled warriors adapted to harsh environments, taking over a society they don't want to adopt.

Human evaluation of LLMs is close to saturation. Models have been optimized so much for plausibility, that we are unable to tell good from bad. Only experts in expert domains can see a meaningful difference.

After a year working on LLM evaluation, our benchmarking paper is finally out (to be presented at NAACL 2025). Main lessons: * All LLM evals are wrong, some are slightly useful. * Goodhart's law. All the time. Everywhere. * Do lots of different evals and hope for the best.

Evaluating LLMs is a bit like paleontology. Trying to understand the behavior of very complex entities by observing only noisy and partial evidence. How do paleontologists deal with the uncertainty and frustration? Do they also feel like doing alchemy instead of science?

Wisdom from my 6y old daughter: "A king is a just person disguised as king."

Over and over again I keep finding @sarahooker.bsky.social papers to reference. This time about ELO rankings. She's always 2-3 years ahead...

So many keywords around LLM training, its easy to get lost. For an incoming paper, did this little visual summary. Would you change anything?

5th International Workshop on Computational Aspects of Deep Learning (CADL) to be held in conjunction with ISC-HPC 2025. 10 days to go, and an award to be decided! Submit your paper and join us sites.google.com/view/cadl2025/

Only two weeks until the deadline! Submit your paper and see you in Germany :)

Bring it on. Totally prepared for another lockdown.

Tired of not having enough time for reading cause all the writing I have to do.

Historical quotes from deep learning: "Don't be a hero" Andrej Karpathy "Attention is all you need" Vaswani et al "We Have No Moat And neither does OpenAI" Google Engineer

Trying to put some order in LLM keywords for an incoming paper. Green concepts are in a different axis, and only partly overlap with elements in blue.

AI researchers today, feeling the If poem: "If you can bear to hear the truth you’ve spoken Twisted by knaves to make a trap for fools, Or watch the things you gave your life to, broken, And stoop and build ’em up with worn-out tools"

To the editors out there, are there any serious downsides to adding an emoji to the title? The stochastic parrots paper seems to be doing alright...

"Personally I would not want to be a member of any group where you either can't wear a hat or you have to wear a hat." George Carlin

Aloe 🌱: How I Learned to Stop Worrying and Love LLMs Finishing the journal paper with all the details right now! huggingface.co/collections/...

What a day! Met the #FaccT deadline 🥵 Got good reviews from #CVPR 😃 Got a short paper accepted at #NAACL'25🥳 Research excitement all around. Need a rest now.

In 2024, we released Aloe Alpha, our 🥇 LLM specialized in healthcare, which gets 10K downloads/month 🥰 6 months later, in a huge team effort, we released four Aloe Beta variants, better than Alpha in every way test. And yet, these get at most 1K downloads/month. Why? huggingface.co/collections/...

If I had a time machine, I would travel 15 years back in time, and tell my young self, who is starting a career in AI research: "Time! The key to AI is not how to represent knowledge, but how to represent time-grounded knowledge!"

At ISC High Performance 2025, I'll be co-organizing the 5th International Workshop on Computational Aspects of Deep Learning (CADL). See: buff.ly/40qiDS4 Deadline: 28 Feb Topics: -Energy-efficient AI -Large-scale pre-training -Distributed learning approaches -Model optimization strategies

Answering "More than you have" to the question "How much data is needed for an AI to solve my problem?"

Jesus...

🚨Our paper `Reclaiming AI as a theoretical tool for cognitive science' is now forthcoming in the journal Computational Brain & Behaviour. (Preprint: osf.io/preprints/ps...) Below a thread summary 🧵1/n #metatheory #AGI #AIhype #cogsci #theoreticalpsych #criticalAIliteracy

All natural intelligence is based on detecting and reacting to change. Meanwhile all our AI systems are static, and AGI sounds like a joke.

Always remember, AI does not create, destroy, discover or is in any way responsible for anything. It's all on us. There's always people behind every marvel and awful thing achieved with AI.

Looking for fun? Wear a shirt all day that says "I'm recording this conversation and I'll use it to train an AI model"

Current AI is at the intersection between statistics and astrology

The LLM Aloe family grows 🌱 We release two new versions, based on Qwen 2.5, fine tuned for healthcare. 7B and 72B. Using 2M instructions, including a variety of medical tasks (summarization, explanation, diagnosis, text classification, treatment recommendation...)

The first time you send a team member on an overseas trip alone for the first time, there's a mix of excitement, pride, and a touch of worry. This one was a leap. From BSC@Barcelona to Meta@San Francisco. www.linkedin.com/feed/update/...

Indeed, Jurassic Park was released +30 years ago!

It used to be, data gathering and preprocessing was the costly first step towards the adoption of AI. Now is benchmarking. If you are going to spend money and effort, do it on building strong tests for AI applications that matter to you

AI Daily Dose - Day 11: Loosely speaking, a bias is a pattern in the data that can be identified and used by a model. Without biases, ML models cannot operate, but under certain undesirable biases, ML can produce wrong or dangerous outputs. Only humans can separate between good and bad biases.

AI Daily Dose - Day 10: Specifying what an AI model must learnt from a set of data scales poorly when depending on human criteria. In Machine Learning (ML) models are only told how to learn, not what. ML models exploit large volumes of data to find patterns in accordance to their programing.

you are not a serious person if your fear of A.I. is “what if scary computer decide 2 extinct humans” and not the health insurance companies using it to sentence people to death

AI Daily Dose - Day 9: The famous quote "All models are wrong, some are useful" refers to the impossibility of perfectly capturing reality, due to its unlimited complexity and granularity. Instead, models should serve a specific purpose, and model reality to that end.

The confusion around LLMs capabilities (clarified below) is increased by the introduction of engineering solutions (e.g., o1) which include one or more LLMs combined through clever tricks, and which are no longer "just an LLM".

One of our research engineers is at Meta’s Global Open Source Innovation Summit this week, presenting our family of fine-tuned open healthcare LLMs: Aloe Check out the entire model family here: huggingface.co/collections/... New, even better Aloe models based on Qwen 2.5 coming next week ;)

@bsc-cns.bsky.social is looking for a director for its new AI Institute. If you are interested in coordinating the research of 50+ researchers across 4 different scientific departments, in a leading HPC institution, consider applying! www.bsc.es/join-us/job-...