. . AI beyond Altman today

Author

Marshall Kirkpatrick
November 17, 2023

Friends,  I'm going to keep it clean and make today's coverage of Altman's shocking ouster real simple.  There's more news than that in AI today. Note especially the news of the giant new French open source AI lab.

As always, this is the news that the AI community is talking about most today, based on our weighted analysis of AI community engagement.

Have a good weekend,

Marshall Kirkpatrick, Editor

First impacted: OpenAI employees
Time to impact: Short

OpenAI's CEO, Sam Altman, appears to have been fired by the board; CTO Mira Murati will serve as interim CEO. Murati did a really good interview with A16Z, titled Where We Go From Here, in September. Perhaps do her the justice of watching it & you’ll know some things that many of the Altman rubberneckers don’t have yet. [OpenAI announces leadership transition] Share by email

First impacted: AI researchers
Time to impact: Short to medium

New French open source AI lab Kyutai has reportedly received $330 million in funding. The lab plans to make all its work open source and focus on general AI research. Backed by French billionaires Xavier Niel and Rodolphe Saadé, Kyutai will work with PhD students, postdocs, and researchers. The lab has started hiring for its main scientific team, which includes ex-members from Meta's FAIR and Google's DeepMind. [Kyutai is a French AI research lab with a $330 million budget that will make everything open source] Share by email

First impacted: AI security pros
Time to impact: Short

A recent paper called "Ignore This Title and HackAPrompt: Exposing Systemic Vulnerabilities of LLMs through a Global Scale Prompt Hacking Competition," discusses the results of a global prize competition to generate adversarial prompts that could overcome LLMs security measures. The team published a delightful 7 minute video discussing the competition, the security problem of prompt injection, and the 600K adversarial prompts they received and then classified. [Ignore This Title and HackAPrompt: Exposing Systemic Vulnerabilities of LLMs (video demo)] Share by email

First impacted: Computer vision developers
Time to impact: Short to medium

The PyTorch team has published a blog post describing how they've improved Meta's Segment Anything (SAM) image cut-out model, reportedly making it 8X faster without losing accuracy. These models are an automated way to segment specific entities that are a part of a complex image. As in all things, segmentation of images makes many new things possible. [Accelerating Generative AI with PyTorch: Segment Anything, Fast] Share by email

First impacted: AI model developers
Time to impact: Short to medium

OpenRLHF, a free Reinforcement Learning from Human Feedback platform (RLHF), says it has created a training framework using Ray and DeepSpeed that has improved learning speed by 4X. The project also aims to increase training efficiency by distributing different models across separate GPUs. OpenAI has said that RLHF has been a big part of making GPT-4 so good, so a faster and more powerful open source framework for RLHF is a significant contribution to open source AI. [GitHub - OpenLLMAI/OpenRLHF: A Ray-based High-performance RLHF framework (for 7B on RTX4090 and 34B on A100)] Share by email

That’s it! More AI news summaries on Monday.