MIT's self-improving AI, Mistral's trillion-parameter model, Google's Genie 4

In this Mixture of Experts podcast, panelists discuss MIT research showing AI can modify another model's weights to improve performance, raising questions about recursive self-improvement and reward hacking risks.

Mistral released a trillion-parameter open-weight model while NVIDIA expanded its open offerings, prompting debate on whether developers will continue relying on proprietary AI as open models approach frontier capabilities.

Google DeepMind introduced Genie 4, a new world model that simulates 3D environments, leading the panel to discuss how well these models understand physics and the risks of convincing simulations diverging from real-world behavior.