Is Your Open-source LLM Really Open?

Slides for my lightning talk at the 86th PyData London meetup (July 2024)

There are more and more Large Language Models (LLMs) being available on the current landscape, with many of them released as "open source". We've discussed some definitions of openness and potential implications.

Transcript

1. Is Your Open-source LLM Really Open? @MarcoBonzanini PyData London
2. © Bonzanini Consulting Ltd — BonzaniniConsulting.com • “Top ranked open source AI model” (TTI - Falcon) • “Open access” (Stability AI - Beluga) • “We have the best open source models” (Mistral) • “Today, we’re introducing Meta Llama 3, the next generation of our state-of-the-art open source large language model” (Meta - Llama 3) 2
3. © Bonzanini Consulting Ltd — BonzaniniConsulting.com 3 Free software Nonfree software Released software Private software User has unlimited rights to the source code. Open source software Public domain with source Lax permissive licenses Copyleft licenses GNU GPL, GNU AGPL, GNU LGPL Other free licenses User doesn't receive the source code, or receives it under a nonfree license or with no license. Nonfree open source “Freeware” Shareware Gratis download https://www.gnu.org/philosophy/categories.html
4. © Bonzanini Consulting Ltd — BonzaniniConsulting.com 4 “Open” “Closed” Ease of Implementation 🤷 🤷 Avoid Hallucinations 🤷 🤷 Ensure Explainability ✅ 🤷 Avoid Copyright Infringement ✅ 🤷 Avoid Data Leakage ✅ 🤷 Ensure Reproducibility ✅ 🤷
5. © Bonzanini Consulting Ltd — BonzaniniConsulting.com 5 “Rethinking open source generative AI: open-washing and the EU AI Act” by Lisenfeld and Dingemanse
6. © Bonzanini Consulting Ltd — BonzaniniConsulting.com 6 EU AI Act (approved May 2024) Non-statutory EU AI Of fi ce No new regulators (“Central Function” within DSIT) Horizontal Cross-sector Vertical Sector-speci fi c Exceptions for open-source 🤷 “Open-source” TBD by EU AI Of fi ce “AI”, “adaptivity”, “autonomy” TBD by individual regulators 🇪🇺 🇬🇧