Orchestrated interactions between LLMs, humans, and tools show great promise.

Today, we introduce “Semantic Decoding' - a perspective that views these interactions as optimization and search in the space of semantic tokens (thoughts).

📄 arxiv.org/abs/2403.14562

Did you know that everything is a Flow👀?

Happy to see a packed room at the workshop on developing and customizing AI workflows with aiFlows (github.com/epfl-dlab/aifl…) at AMLD earlier today!

Nicolas Baldwin

Tomorrow, with Nicolas Baldwin, we'll be hosting the 'Learn to develop and customize AI workflows with Flows' workshop at the #AMLDEPFL2024 conference! Join us to learn about the new version of aiFlows (github.com/epfl-dlab/aifl…) coming out later today!

Bob West Maxime Peyrard

When I talk with people about constrained decoding, I'm always asked: 'Can it be applied to blackbox LLMs like GPT-4?'

My response has been a bit pessimistic due to the limited logit access.🤔

🚀However, we're excited to announce Sketch-Guide Constrained Decoding (SGCD), a…

We are releasing the 1st version of 4M, a framework for training multimodal foundation models across tens of modalities & tasks, based on scalable masked modeling.
Joint effort by EPFL & Apple.

4M: Massively Multimodal Masked Modeling
🌐4m.epfl.ch
🧵1/n

Language Modeling Is Compression

Chinchilla 70B achieves compression rates of 43.4% on ImageNet patches and 16.4% on LibriSpeech samples, beating domain-specific compressors like PNG (58.5%) or FLAC (30.3%), respectively

arxiv.org/abs/2309.10668

Two important lessons:
1) Don't underestimate large language models
2) Don't overestimate large language models