#238 - GPT 5.4 mini, OpenAI Pivot, Mamba 3, Attention Residuals | Last Week in AI

This episode of Last Week in AI explores recent developments in artificial intelligence, focusing on new model releases, business strategies, and safety concerns. OpenAI's GPT-5.4 mini and nano models are examined, with emphasis on their cost and token efficiency compared to GPT-5 mini. Mistral's open-source small four family of models is also discussed, highlighting its combination of reasoning, multi-modal capabilities, and genetic coding optimization. The podcast further analyzes Meta's Manus launch and NVIDIA's NEMA CLAW, both aimed at integrating AI agents into local operating systems. Additionally, the hosts discuss OpenAI's strategic shift towards productivity and business applications, as well as concerns about the company's planned "adult mode" for ChatGPT. The episode concludes with a review of research on steganography, emergent misalignment, and cyber attack scenarios.

Outlines

Part 1: Event & Podcast Introduction

Part 2: New Model Releases & Benchmarks

Part 3: AI Agents & Local Computing

Part 4: OpenAI's Internal Ethics & Strategy

Part 5: Hardware, Infrastructure & Market Competition

Part 6: Big Tech Reorganizations & Tensions

Part 7: AI Safety, Alignment & Evaluation

Part 8: Policy & Technical Research

Sign in to continue reading, translating and more.

Open full episode in Podwise

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

Last Week in AI

Part 1: Event & Podcast Introduction

00:10ODSC AI East 2026: A Conference for AI Practitioners and Teams

ODSC AI East 2026: A Conference for AI Practitioners and Teams

01:40Podcast Introduction: AI News, Research, and Safety Preview

Podcast Introduction: AI News, Research, and Safety Preview

Part 2: New Model Releases & Benchmarks

04:09OpenAI's GPT-5.4 Mini and Nano: Faster, Smaller Models with Cost Considerations

OpenAI's GPT-5.4 Mini and Nano: Faster, Smaller Models with Cost Considerations

07:34Mistral's Small Four: Open Source Model Combining Reasoning, Multimodal, and Coding

Mistral's Small Four: Open Source Model Combining Reasoning, Multimodal, and Coding

10:34Mistral's Consolidation Strategy and Efficiency Claims Compared to GPT-OSS-120B

Mistral's Consolidation Strategy and Efficiency Claims Compared to GPT-OSS-120B

Part 3: AI Agents & Local Computing

14:31Meta's Manus Launches My Computer: An AI Agent on Your Mac

Meta's Manus Launches My Computer: An AI Agent on Your Mac

17:51NVIDIA's NEMA CLAW: A Stack for OpenCLAW Agent Platforms with Security Controls

NVIDIA's NEMA CLAW: A Stack for OpenCLAW Agent Platforms with Security Controls

20:31NVIDIA's Operating System Play: NEMA CLAW and the Agentic Runtime Environment

NVIDIA's Operating System Play: NEMA CLAW and the Agentic Runtime Environment

23:42NVIDIA's DLSS5: Generative AI for Enhanced Game Graphics

NVIDIA's DLSS5: Generative AI for Enhanced Game Graphics

Part 4: OpenAI's Internal Ethics & Strategy

27:00OpenAI's Delayed Adult Mode: Ethical Concerns and Potential Benefits

OpenAI's Delayed Adult Mode: Ethical Concerns and Potential Benefits

30:11OpenAI's Adult Mode: Erotic Roleplay, Research, and Branding Implications

OpenAI's Adult Mode: Erotic Roleplay, Research, and Branding Implications

34:27OpenAI's Strategic Shift: Focusing on Productivity and Business Applications

OpenAI's Strategic Shift: Focusing on Productivity and Business Applications

37:30OpenAI's Business Strategy: Balancing Innovation with Enterprise Focus

OpenAI's Business Strategy: Balancing Innovation with Enterprise Focus

Part 5: Hardware, Infrastructure & Market Competition

40:47NVIDIA's Blackwell and Rubin Chips: $1 Trillion Opportunity Driven by Agentic AI

NVIDIA's Blackwell and Rubin Chips: $1 Trillion Opportunity Driven by Agentic AI

44:36NVIDIA's Grok Integration: Language Processing Units and Memory Innovations

NVIDIA's Grok Integration: Language Processing Units and Memory Innovations

47:21Mistral's FORGE: Training Custom AI Models for Businesses

Mistral's FORGE: Training Custom AI Models for Businesses

51:21Mistral's FORGE: Competing with Cohere and the Enterprise Market

Mistral's FORGE: Competing with Cohere and the Enterprise Market

55:33ByteDance's Access to NVIDIA AI Chips: Export Controls and Compute Power

ByteDance's Access to NVIDIA AI Chips: Export Controls and Compute Power

Part 6: Big Tech Reorganizations & Tensions

59:27Meta Delays Avocado Model: Licensing Competitors' AI Technology

Meta Delays Avocado Model: Licensing Competitors' AI Technology

1:01:26Meta's Superintelligence Team: Clashes and Product Implications

Meta's Superintelligence Team: Clashes and Product Implications

1:04:01Microsoft Shakes Up AI Division: Copilot Falls Behind Google and OpenAI

Microsoft Shakes Up AI Division: Copilot Falls Behind Google and OpenAI

1:07:53The Superintelligence Tension: Commercial Products vs. Singularity

The Superintelligence Tension: Commercial Products vs. Singularity

Part 7: AI Safety, Alignment & Evaluation

1:11:04Steganography in LLMs: Measuring Hidden Reasoning and Deception

Steganography in LLMs: Measuring Hidden Reasoning and Deception

1:14:38Reasoning Theater: Disentangling Model Beliefs from Chain of Thought

Reasoning Theater: Disentangling Model Beliefs from Chain of Thought

1:19:58Training Defenses Against Emergent Misalignment in Language Models

Training Defenses Against Emergent Misalignment in Language Models

1:24:22Frontier AI Agents in Cyber Attack Scenarios: Capabilities and Improvements

Frontier AI Agents in Cyber Attack Scenarios: Capabilities and Improvements

1:25:51Eval Awareness in Cloud Opus 4.6: Browser Comp Performances

Eval Awareness in Cloud Opus 4.6: Browser Comp Performances

1:30:07Anthropic's Bloom: Open Source Agentic Framework for Behavioral Evaluations

Anthropic's Bloom: Open Source Agentic Framework for Behavioral Evaluations

1:33:55How Well Do Models Follow Their Constitutions? Anthropic's Constitutional AI

How Well Do Models Follow Their Constitutions? Anthropic's Constitutional AI

1:35:47Anthropic's Constitution: Cross-Company Testing and Failure Modes

Anthropic's Constitution: Cross-Company Testing and Failure Modes

Part 8: Policy & Technical Research

1:38:40Security Concerns Over NVIDIA's H200 License and Export Controls

Security Concerns Over NVIDIA's H200 License and Export Controls

1:41:42Export Controls: Transparency and Legislative Oversight

Export Controls: Transparency and Legislative Oversight

1:42:11Attention Residuals: Improving Transformers with Layer-Specific Attention

Attention Residuals: Improving Transformers with Layer-Specific Attention

1:48:40Mamba3: Improved Sequence Modeling with State Space Principles and Efficiency Gains

Mamba3: Improved Sequence Modeling with State Space Principles and Efficiency Gains