#232 - ChatGPT Ads, Thinking Machines Drama, STEM

Andrei Kerenkov and Jeremy Harris delve into the implications of AI in authoritarian regimes, with Jeremy, the "safety hawk," emphasizing the risks of surveillance and the power dynamics between governments and citizens. They discuss China's advancements in AI, particularly the use of facial recognition technology by the Chinese Communist Party to identify dissenters and the coercive tactics employed by authoritarian governments, such as social credit systems. The hosts stress the importance of governance in relation to superintelligence and the limitations of AI's physical presence.

The conversation shifts to OpenAI's testing of ads in ChatGPT, which will be clearly labeled and not interfere with model responses. Public reaction is mixed, with discussions about the implications of ads for user experience and OpenAI's need to generate revenue to scale operations. OpenAI emphasizes user privacy and control over data, allowing users to opt out of personalization, contrasting with typical ad-driven models.

The introduction of ChatGPT Go, a low-cost subscription service, aims to expand access to features while addressing safety concerns about minors interacting with AI. The podcast also highlights the implications of AI interactions for teenagers and the challenges posed by uniform communication styles of chatbots.

In the AI landscape, Google's Gemini is noted for offering free SAT practice exams, while Baidu's AI assistant, Ernie, has reached significant user numbers in China. The discussion transitions to developments at Thinking Machines, where co-founders have left to return to OpenAI, raising concerns about the company's stability and internal conflicts.

The podcast addresses advancements in AI and chip technology in China, particularly with JPU AI's new model trained on Huawei's AI compute stack. The evolving dynamics in the chip market are discussed, with Nvidia surpassing Apple as TSMC's largest customer. The conversation also touches on Musk's XAI, which has launched a gigawatt AI supercluster, Colossus, providing a competitive edge in securing power for data centers.

Innovative efforts at Thinking Machines aim to enhance collaboration through AI-driven software, while advancements in AI development, particularly in long horizon and multi-agent reinforcement learning, are highlighted. The release of Flux by Black Forest Labs showcases a compact image model optimized for modern GPUs, raising questions about market saturation in image models.

The podcast discusses the introduction of Hartmühle, a family of open-sourced music foundation models, and advancements in STEM scaling transformers, proposing a new approach for faster processing. The conversation also covers the mixture of experts (MOE) approach and the exploration of sparse auto encoders (SAEs) to enhance model activations.

A key paper reveals that reasoning models optimized through reinforcement learning demonstrate greater perspective diversity. The challenges faced by the LASFUNC group in creating a system to write AI papers are discussed, highlighting systemic issues in research processes.

In policy updates, the US Senate has passed the Defiance Act, allowing victims to sue over non-consensual AI-generated images. A new safety research paper focuses on building production-ready probes for Gemini, enhancing accuracy on complex prompts. Anthropic has published an updated constitution for Claude, emphasizing principles and values for model training.

The conversation highlights the importance of transparency in AI development and the cultural pushback against AI, with critiques from various political perspectives. The impact of technology on filmmaking is discussed, raising philosophical questions about prioritizing consumers or producers in the context of art.