Llama 3 degrades more than Llama 2 when quantized.

Probably because Llama 3, trained on a record 15T tokens, captures extremely nuanced data relationships, utilizing even the minutest decimals in BF16 precision fully.

Making it more sensitive to quantization degradation.…

This is the next big thing (for nerds): your own home cluster of Mac Mini or Mac Studio for distributed local LLM inference.

Made possible by the unified memory architecture of Apple Silicon.

Should I create a distributed version of Pico AI apps and developer tools for iOS and macOS MLX server just for fun?

😱

tl;dr Waymo robotaxis will stop when they see a person wearing a t-shirt with a stop sign on it. Surprising?

What will they do if they see a person wearing a 'Speed Limit 60' t-shirt in the middle of a San Francisco street?

thestreet.com/automotive/the…

Some divulgative information about the Hatedemics project to fight #HateSpeech and #disinformation with AI and LLMs. Article in English available here: magazine.fbk.eu/en/news/hatede…
Interview available here: buff.ly/49Otnw3 (in Italian but automatic translation is quite good).

Published in Nature Machine Intelligence today, our new article explores the trade-offs of personalised alignment in large language models ⚖️ Personalisation has potential to democratise decisions over how LLMs behave, but brings its own set of risks...
nature.com/articles/s4225…

Honored to coordinate this project. A proposal accepted, a dream come true: now we should all walk together northward - as Dave taught us. 😃 Thanks to all the wonderful people who participated!

What??

So many young talented students visiting our group recently. Offices are a bit overcrowded, but it is worth it!
Nicolás Benjamín Ocampo welcome onboard!

Today we release a paper on #NLProc for CounterSpeech that is much more than a survey, it is an how-to-guide for all those who want to experiment with the topic. I really enjoyed writing it with wonderful co-authors: Helena Bonaldi Yi-Ling Chung gavin abercrombie #AI4socialgood

Trenitalia really have lost the plot when it comes to regional ticketing, book a ticket on their own app and you receive a 'not valid for travel' pdf.
You then need to retrieve the ticket in the app, and finally check-in before your planned train departure.
Talk about simplicity!

We are doing some exciting work in understanding how we can make the multimodal hate speech detection explainable ...

Please apply if you are interested. You can DM me if you have any questions ...

New HiTZ zentroa (UPV/EHU) paper! Check out our new Basque and Spanish CONAN-EUS dataset for counter-narrative generation against hate speech. Paper, data and fine-tuned models available Hugging Face
Paper: arxiv.org/abs/2403.09159
Dataset: huggingface.co/datasets/HiTZ/…

This is LLM temperature 2 decoding visualized…

We won Best Demo Paper! at #EACL2024 for multi-person human-robot conversation using Generative AI / LLMs Interaction Lab The National Robotarium Edinburgh Robotics Heriot-Watt University #GenAI #NLProc #HRI #Conversational_AI #AI aclanthology.org/2024.eacl-demo…

With LLMs on the rise, there is a growing number of empathetic systems. But are they truly empathetic? We survey 21 systems & find limitations wrt important aspects of empathy—specificity, reflection level, diversity—&offer guidance for future work. Work led by Andrew Lee Andrew Lee

😨

It's been a long journey, but it was worth the effort. Well done Daniela Occhipinti! And thanks to Serra Sinem Tekiroğlu for being by our side. As a side note: endowing LLMs with a personality is still a long way to go, but from today it's a little shorter. 👇👇

For those who missed Nicolò Penzo's talk yesterday on using contextual information for text classification, here's a summary 👉 x.com/penzo_nicolo/s…
#EACL2024

This is wild. First empirical evidence and controlled study of GPT use in CS reviews. I suggest reading at least the abstract to find many additional striking findings.