#238 - GPT 5.4 mini, OpenAI Pivot, Mamba 3, Attention Residuals

The podcast episode delves into significant updates in AI, particularly OpenAI's new model releases, including GPT 5.4 mini and nano. The GPT 5.4 mini demonstrates improved performance and speed, while the nano model, though quick, lags in benchmark performance. Both models feature a 400,000 token context window, but rising costs raise questions about OpenAI's commitment to affordability. The hosts stress the importance of selecting the right model for specific workloads to optimize costs and returns, expressing concern over the lack of practical metrics in model announcements.

Mistral's new family of models, which utilize a mixture of experts, aims to consolidate reasoning, coding, and multimodal capabilities into a single framework, marking a strategic shift for the company. The discussion highlights the significance of output length efficiency, noting that more tokens do not necessarily equate to better reasoning. Mistral's model claims competitive scores with GPT-OSS 120B while generating shorter outputs, crucial for users concerned about operational costs.

The episode also covers a new software functioning as an "open claw" on computers, enabling command line execution and acting as an AI assistant. This reflects a competitive landscape where multiple companies are entering the market, with Meta's acquisition of Manus aimed at penetrating the local operating system layer. Nvidia's announcement of NimaClaw introduces a stack for OpenClaw agent platforms, enhancing privacy and security for agent operations.

The conversation touches on the implications of running models locally versus in the cloud, particularly regarding access to sensitive files, and emphasizes the need for guardrails for AI models. Jensen's rhetoric about a "new Renaissance" in AI advancements is noted, alongside the recognition of a gap between AI frameworks' potential and their actual impact.

Nvidia's entry into the operating system market for agents is compared to historical developments in personal computing, with a shift from "software eats the world" to "AI is eating the world." The simplicity of deploying Nvidia's models and runtime is emphasized, aiming to foster a competitive ecosystem. Nvidia's introduction of DLSS5, described as a "GPT moment for graphics," features machine learning-based upscaling and generative AI to enhance game graphics, though reactions are mixed.

OpenAI's plans to launch Charge GPT's adult mode face opposition from its advisory council, raising skepticism about the implications of AI-driven adult content. The conversation reflects a mix of anticipation and concern about AI's societal implications, with suggestions that companies may leak information to gauge public reaction. OpenAI is working to catch up with competitors like Claude and CoWork, particularly in user adoption and feature sets.

The AI landscape is evolving, with companies needing to learn about market share acquisition and retention while maintaining product-market fit against competition. Nvidia's recent announcements signal a shift towards agentic AI applications, requiring increased computational power. Mistral has introduced Forge, enabling businesses to train their own AI models, though skepticism surrounds its ability to compete against larger players.

Bydance's access to Nvidia's top AI chips raises concerns about potential applications in China, while Meta delays the rollout of its next AI model, Avocado, due to training challenges. Microsoft is restructuring its AI division to address lagging performance against competitors. The conversation reflects a broader industry trend regarding the urgency of developing superintelligence while balancing AI as a commercial product.

A new framework for monitoring large language models introduces decision theoretic formalization of stenography, raising concerns about models concealing malicious behavior. The concept of the "steganographic gap" is introduced, suggesting that detecting steganography is a promising area for further research. The discussion also examines whether models' intermediate outputs reflect genuine reasoning or are merely performative, emphasizing the importance of aligning model capabilities with problem difficulty.

Insights into model behavior reveal that text-based outputs are more closely tied to desired behaviors than activation space regularization. The conversation notes ongoing findings in alignment research and a new study focusing on frontier AI agents in multi-step cyber attack scenarios. Anthropic's report on Opus 4.6 raises questions about coherent extrapolated volition and the model's ability to predict desired actions based on context.

Jeremy introduces a paper titled "Attention Residuals," discussing the importance of residual connections in transformers and the limitations of standard transformers in managing attention across different layers. The introduction of "attention residuals" aims to enable models to prioritize specific layers, though challenges remain due to high memory consumption. The Mamba 3 paper seeks to enhance sequence modeling through state space principles, demonstrating superior performance in downstream accuracy while maintaining perplexity with reduced state size.

The conversation emphasizes the importance of staying updated with the latest advancements in AI, encouraging listeners to engage with the content and share their thoughts.