Mid-training alignment generalization study

Our 240th episode with a summary and discussion of last week's big AI news!

Recorded on 04/08/2026 (sorry I keep releasing stuff late, will get better with it soon!)

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:

• Anthropic launched Project Glasswing and previewed Claude Mythos, a general-purpose model withheld from broad release due to dramatically stronger autonomous offensive cybersecurity performance (including zero-day discovery), alongside concerning bio/virology uplift results and documented deception/containment-escape behaviors; pricing is far higher than Opus and most discovered vulnerabilities remain unpatched.
• Product and platform updates included Google’s Gemini 3.1 Flash Live for real-time multilingual voice conversation, Suno v5.5 personalization features, Anthropic tightening Claude Code/OpenClaw access and usage limits, OpenAI canceling an “adult mode,” and Microsoft releasing MAI models for speech-to-text, audio generation, and image generation.
• Business and market developments featured Anthropic’s revenue run rate surpassing $30B and a major Google/Broadcom TPU compute expansion, SoftBank taking a $40B short-term loan to fund OpenAI commitments, Granola reaching a $1.5B valuation, Anthropic buying Coefficient Bio for $400M, and OpenAI acquiring the TBPN business talk show.
• Policy, open-source, and geopolitics included Z.ai releasing open-weight GLM 5.1 and a multimodal GLM model, Google open-sourcing Gemma 4 under Apache 2.0, a judge blocking the Pentagon’s “supply chain risk” label against Anthropic, research on LLM “emotion vectors” and OpenAI meta-gaming during RL, China restricting Manus founders amid Meta deal review, scrutiny of Nvidia’s chip-smuggling claims, China chipmakers gaining market share, and Iran framing cloud data centers as military targets.

Timestamps:

• (00:00:10) Intro / Banter
• Tools & Apps
• (00:01:58) Anthropic debuts ‘Project Glasswing’ and new AI model for cybersecurity | The Verge
• (00:18:22) Gemini Live gets ‘biggest upgrade yet’ with Gemini 3.1 Flash Live
• (00:20:40) Anthropic says Claude Code subscribers will need to pay extra for OpenClaw usage | TechCrunch
• (00:25:36) OpenAI abandons yet another side quest: ChatGPT's erotic mode | TechCrunch
• (00:26:16) Microsoft takes on AI rivals with three new foundational models | TechCrunch
• (00:31:25) Suno leans into customization with v5.5 | The Verge
• Applications & Business
• (00:32:53) Anthropic announces deal with Google, Broadcom, says revenue has tripled
• (00:37:53) Sam Altman May Control Our Future—Can He Be Trusted? | The New Yorker
• (00:40:18) OpenAI, Anthropic, Google Unite to Combat Model Copying in China - Bloomberg
• (00:41:45) Chinese chipmakers claim nearly half of local market as Nvidia's lead shrinks
• (00:45:20) SoftBank secures $40 billion loan to boost OpenAI investments
• (00:47:23) Granola raises $125M at $1.5B valuation for its AI note-taking app - SiliconANGLE
• (00:48:17) Anthropic acquires stealth startup Coefficient Bio in $400M deal
• (00:50:20) OpenAI acquires TBPN, the buzzy founder-led business talk show | TechCrunch
• Projects & Open Source
• (00:53:04) Z.AI Introduces GLM-5.1: An Open-Weight 754B Agentic Model That Achieves SOTA on SWE-Bench Pro and Sustains 8-Hour Autonomous Execution - MarkTechPost
• (00:55:14) Google announces Gemma 4 open AI models, switches to Apache 2.0 license - Ars Technica
• (01:01:26) Z.ai Launches GLM-5V-Turbo: A Native Multimodal Vision Coding Model Optimized for OpenClaw and High-Capacity Agentic Engineering Workflows Everywhere
• Policy & Safety
• (01:04:45) Judge blocks Pentagon’s effort to ‘punish’ Anthropic by labeling it a supply chain risk
• (01:10:05) Emotion concepts and their function in a large language model
• (01:21:12) China bars Manus co-founders from leaving country amid Meta deal review, FT reports
• (01:25:38) US lawmakers ask whether Nvidia CEO's smuggling remarks misled regulators
• (01:27:48) How far does alignment midtraining generalize?
• (01:32:20) Metagaming matters for training, evaluation, and oversight
• (01:39:31) Iran says it has struck Oracle data center in Dubai, Amazon data center in Bahrain — country has threatened to attack Nvidia, Intel, and others, too

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