In-training defenses against emergent misalignment

Our 238th episode with a summary and discussion of last week's big AI news!

Recorded on 03/18/2026

Hosted by Andrey Kurenkov and Jeremie Harris

Feel free to email us your questions and feedback at andreyvkurenkov@gmail.com and/or hello@gladstone.ai

Read out our text newsletter and comment on the podcast at https://lastweekin.ai/

In this episode:

* OpenAI released GPT-5.4 mini and nano with 400k-token context windows, higher per-token prices but claimed token-efficiency gains in Codex; nano is API-only and pitched for high-volume classification/data extraction despite a major price increase.

* Mistral open-sourced the Small 4 model family (MoE, 119B total/6B active) combining reasoning, multimodal, and coding-agent capabilities, and announced Forge to help businesses train or post-train custom models.

* Agent “operating system” competition intensified with Meta’s acquired Manus launching a local Mac agent, Nvidia announcing NeMo/“Open Shell” sandboxed agent runtime, and Nvidia also unveiling DLSS 5 plus major hardware forecasts including Groq LPU integration.

* Business and safety updates included OpenAI shifting focus toward productivity/enterprise amid competition, Microsoft reorganizing Copilot and frontier-model efforts, Meta delaying its next model, China-linked ByteDance deploying large Nvidia clusters abroad, and new safety work on steganography, chain-of-thought faithfulness, fine-tuning defenses, cyber-attack evals, and constitution/spec compliance.

A thank you to our current sponsors:

• Box - visit Box.com/AI to learn more
• ODSC AI - go to odsc.ai/east and use promo code LWAI for an additional 15% off your pass to ODSC AI East 2026.
• Factor - head to factormeals.com/lwai50off and use code lwai50off to get 50 percent off and free breakfast for a year

Timestamps:

• (00:00:10) Intro / Banter
• (00:01:56) News Preview
• Tools & Apps
• (00:02:39) OpenAI ships GPT-5.4 mini and nano, faster and more capable but up to 4x pricier
• (00:08:04) Mistral's new Small 4 model punches above its weight with 128 expert modules
• (00:14:03) Meta's Manus launches 'My Computer' to turn your Mac into an AI agent - 9to5Mac
• (00:17:57) NVIDIA Announces NemoClaw for the OpenClaw Community | NVIDIA Newsroom + Nvidia boosts knowledge work with Open Agent Development Platform
• (00:24:09) DLSS 5 looks like a real-time generative AI filter for video games | The Verge
• (00:26:36) OpenAI to Launch ChatGPT 'Adult Mode' Despite Warnings From Its Own Advisers - CNET
• Applications & Business
• (00:33:46) OpenAI Reportedly Pivoting to a Focus on Business and Productivity Only
• (00:41:25) Nvidia GTC 2026: CEO Jensen Huang sees $1 trillion in orders for Blackwell and Vera Rubin through ’27
• (00:45:44) Mistral launches Forge to help enterprises build their own AI models
• (00:54:17) China's ByteDance gets access to top Nvidia AI chips, WSJ reports
• (00:57:57) Meta Delays Rollout of New A.I. Model After Performance Concerns
• (01:02:50) Microsoft Shakes Up AI Division As Copilot Falls Behind Google and OpenAI
• Policy & Safety
• (01:07:26) A Decision-Theoretic Formalisation of Steganography With Applications to LLM Monitoring
• (01:13:09) Reasoning Theater: Disentangling Model Beliefs from Chain-of-Thought
• (01:18:29) In-Training Defenses against Emergent Misalignment in Language Models
• (01:23:07) How do frontier AI agents perform in multi-step cyber-attack scenarios?
• (01:25:20) Eval awareness in Claude Opus 4.6’s BrowseComp performance
• (01:29:49) Introducing Bloom: an open source tool for automated behavioral evaluations
• (01:32:26) How well do models follow their constitutions?
• (01:37:11) Nvidia’s H200 License Stirs Security Concern Among Top Democrats
• Research & Advancements
• (01:40:050) [2603.15031] Attention Residuals
• (01:47:11) Mamba-3: Improved Sequence Modeling using State Space Principles

See Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.