18 foundational challenges in AI safety on the path to guaranteed safe AI

Hello, fellow human! Today's issue is dedicated to a series of AI safety talks that we’ll be hosting in the next couple of weeks.

This Thursday, a researcher from Oxford University will present a massive work — massive not only in importance, but also in the number of pages and the number of contributors, including Max Tegmark and Yoshua Bengio.

The paper is titled Towards Guaranteed Safe AI: A Framework for Ensuring Robust and Reliable AI Systems. It suggests safety approaches based on mathematical descriptions that supposedly ensure the development of safe AI.

Check out this link to learn more about the virtual talk and attend.

Interestingly, researchers from Cambridge University explored foundational challenges in building safe and aligned AI. They identified 18 such challenges, which the lead researcher of this work, Usman Anwar, will discuss with our community next Thursday.

Check out the details of the virtual talk and join the conversation.

We recently hosted a talk with Aleksandar Botev, a research scientist at Google DeepMind and co-author of their recent paper on the Griffin architecture. Griffin is an efficient alternative to Transformers. It mixes gated linear recurrences with local attention, performing very well on long sequences. It has lower latency and higher throughput during inference compared to Transformers, though it doesn't perform well at needle-in-the-haystack retrieval tasks.

A few weeks ago, we hosted a talk by Dan Fu from Stanford University, who gave the BuzzRobot community an overview of emerging alternative models. Griffin was among the new challengers. During the talk, Dan mentioned that it would take companies at least six months and their best ML systems engineers to rearrange their tech stack for non-Transformer models. Aleksandar's perspective on this is that well-resourced companies that understand their use cases well will be able to serve both types of models: Transformers and alternatives. 

Curious to see how the industry responds to alternative models.

Watch the lecture on our YouTube channel to learn the technical details behind the Griffin architecture.