Lecture on Llama 3 with a Meta Researcher

Hello, fellow human! Sophia here. Check out the virtual talks we’ve lined up for the next few weeks that cover state-of-the-art AI research.

This Thursday, June 20th, we will explore architectural alternatives competing with the Transformer models.

Since its introduction in 2018, the Transformer architecture has completely transformed the AI landscape. While there are plenty of ways to optimize Transformers, is there any room for other architectural decisions that may be, dare we say, more efficient than Transformers?

After years of development, alternatives such as Mamba, Griffin, SSMs, and others are gaining traction. Join us as our guest, Stanford University's Dan Fu, gets into the evolution of these alternatives and what they bring to the table. Dan will also share his latest research, presenting new approaches that rival Transformers.

Learn the details and register here

On June 27th, we are hosting a virtual talk with a researcher from Cohere. Reinforcement Learning from Human Feedback (RLHF) is widely recognized as a key factor in the high performance of large language models.

Traditionally, the AI community has adopted the Proximal Policy Optimization (PPO) algorithm as the standard for the Reinforcement Learning component of RLHF. But the problem with PPO – it’s computationally expensive and sensitive to hyperparameter tuning.

In response, Cohere's researchers are pioneering an alternative to PPO. They propose a new REINFORCE-style optimization method that not only outperforms PPO but also avoids the drawbacks associated with 'Reinforcement learning-free' methods like DPO and RAFT.

Learn the details and register here.

Virtual Lecture Recap: AMA with Thomas Scialom

We recently hosted an AMA (ask me anything) session with Thomas Scialom, an AGI researcher at Meta. Having led the training of Llama 2 and post-training of Llama 3, Thomas is currently developing the next generation of Llama models.

Highlights from our discussion:

• By the end of this year, open-access models are expected to match the performance of the latest private models like GPT-4.

• Meta's primary focus is enabling intelligent AI assistants, which is a step towards AGI. Once achieved, this milestone will guide the next phases of AGI development.

• Current AI research into building intelligent AI assistants relies on three core components, aptly named the "triangle" by Thomas:  (1) more computing power as researchers have yet to hit a plateau, (2) creating better human-generated and synthetic data, and (3) excellent people.

• Open source is the main driver for AI development. If Google hadn't publicly released the Transformer architecture, which the AI community adopted and advanced further, progress would have been slower. Even internally, Google would still be using LSTMs.

• The importance of continued open-sourcing of AI work and developing a solid framework for testing large-scale models to minimize harmful risks.

Catch the recording of the full AMA session on our YouTube channel

Towards a New Paradigm of Distributed AI Training by Google DeepMind

Modern AI depends on large-scale models boasting trillions of parameters that could theoretically scale indefinitely. How is it possible to achieve?

Google DeepMind researchers have recently released a prototype for a novel distributed training method called Distributed Paths Composition (DiPaCo), leveraging the Mixture of Experts approach—specifically, a Mixture of Paths—to enhance the efficiency and scalability of AI system computations.

Arthur Douillard, a Senior Research Scientist at Google DeepMind, has detailed the technical aspects of DiPaCo during a presentation to the BuzzRobot community.

Watch the presentation on our YouTube channel.