Reproducing Steering of Evaluation Awareness

Our 241st episode with a summary and discussion of last week's big AI news!

Recorded on 04/18/2026

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:

• Anthropic released Claude Opus 4.7 with improved benchmark performance, new reasoning controls, better vision and memory, and a detailed system card discussing deception risk, evaluation-awareness steering, and a training bug that accidentally supervised chain-of-thought in 7–8% of episodes.
• Meta unveiled its closed Muse Spark model and “contemplating mode,” highlighting test-time scaling, thought compression, large infrastructure plans like the Hyperion data center, and findings that it shows unusually high evaluation awareness.
• OpenAI introduced limited-access GPT 5.4 Cyber for defensive security teams and rolled major Codex updates including computer use, browser and plugins, image generation, and long-horizon task scheduling; competing agent products also launched from Anthropic, Canva, and Adobe.
• Business, policy, and safety news included continued government blacklisting litigation affecting Anthropic, CoreWeave compute deals, Perplexity revenue growth tied to agents, a potential Cohere–Aleph Alpha merger, attacks targeting Sam Altman and OpenAI, AI propaganda trends, and new alignment research on automated weak-to-strong supervision and steering evaluation awareness.

Timestamps:

• (00:00:10) Intro / Banter
• (00:03:43) News Preview
• (00:04:14) Response to listener comments

• Tools & Apps
• (00:05:30) Anthropic releases Claude Opus 4.7, narrowly retaking lead for most powerful generally available LLM | VentureBeat
• (00:24:15) Meta debuts the Muse Spark model in a 'ground-up overhaul' of its AI | TechCrunch
• (00:34:23) OpenAI Launches GPT-5.4-Cyber with Expanded Access for Security Teams
• (00:39:44) OpenAI’s big Codex update is a direct shot at Claude Code | The Verge
• (00:42:10) Anthropic launches Claude Design, a new product for creating quick visuals
• (00:42:30) Anthropic’s New Product Aims to Handle the Hard Part of Building AI Agents | WIRED
• (00:42:54) Canva’s AI 2.0 update goes all in on prompt-powered design tools | The Verge
• (00:43:06) Adobe’s new AI Assistant marks a ‘fundamental shift’ in creative work | The Verge
• (00:43:38) Gemini can now pull from Google Photos to generate personalized images | The Verge
• (00:43:52) Google rolls out a native Gemini app for Mac | TechCrunch
• (00:44:04) Chrome now lets you turn AI prompts into repeatable ‘Skills’ | The Verge

• Applications & Business
• (00:44:22) Anthropic loses appeals court bid to temporarily block Pentagon blacklisting
• (00:49:07) Jeff Bezos’ AI lab poaches xAI cofounder Kyle Kozic from OpenAI. | The Verge
• (00:51:39) Perplexity's Shift to AI Agents Boosts Revenue 50%
• (00:53:53) Anthropic Agrees to Rent CoreWeave AI Capacity to Power Claude
• (00:57:32) Canada’s Cohere, Germany’s Aleph Alpha reportedly in merger talks
• (01:04:23) ChatGPT has a new $100 per month Pro subscription | The Verge
• (01:05:10) OpenAI has bought AI personal finance startup Hiro | TechCrunch
• (01:07:03) Allbirds announced a switch from shoes to AI and its stock jumped 600 percent | The Verge

• Projects & Open Source
• (01:07:26) HY-World 2.0: A Multi-Modal World Model for Reconstructing, Generating, and Simulating 3D Worlds + Lyra 2.0: Explorable Generative 3D Worlds

• Policy & Safety
• (01:19:12) Daniel Moreno-Gama is facing federal charges for attacking Sam Altman’s home and OpenAI’s HQ | The Verge
• (01:20:15) Duo accused of shooting at Sam Altman’s house are freed; no charges filed 
• (01:24:50) The Iranian Lego AI video creators credit their virality to ‘heart’ | The Verge
• (01:27:19) Hundreds of Fake Pro-Trump Avatars Emerge on Social Media - The New York Times
• (01:27:31) The AI images Trump can’t get enough of | Donald Trump | The Guardian
• (01:29:25) Automated Weak-to-Strong Researcher
• (01:43:51) Reproducing steering against evaluation awareness in a large open-weight model
• (01:49:53) Iran threatens ‘complete and utter annihilation’ of OpenAI's $30B Stargate AI data center in Abu Dhabi — regime posts video with satellite imagery of ChatGPT-maker's premier 1GW data center
• (01:53:57) Wall Street Banks Try Out Anthropic’s Mythos as US Urges

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