Last Week in AI #243 - GPT 5.5, DeepSeek V4, AI safety sabotage

Our 243rd episode with a summary and discussion of last week's big AI news! Recorded on 04/29/2026 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-5.5 with strong coding-oriented improvements, a system card discussing chain-of-thought monitorability and misalignment testing, higher pricing than GPT-5.4, and notable quirks like a system-prompt warning about “goblins.” * xAI launched Grok Voice Think Fast 1.0, claiming large benchmark leads for real-time voice agents and reporting major Starlink customer-support automation and sales conversion impact. * DeepSeek open-sourced DeepSeek V4 (Pro and Flash) featuring MoE scaling and 1M-token context via hybrid/compressed attention changes, while Tencent released Hunyuan 3 preview with weaker benchmark performance; a new long-horizon agent benchmark (Clawmark) shows low task success rates. * Major business, legal, and policy updates include Google’s planned up-to-$40B investment and 5GW compute commitment to Anthropic, Meta’s AWS Gravitron deal and China blocking Meta’s Manus acquisition, a revamped OpenAI–Microsoft agreement, ongoing Musk–OpenAI trial developments, and new safety/security research on sabotage, document degradation under delegation, and bit-flip attacks. Timestamps: (00:00:10) Intro / Banter (00:02:00) News Preview (00:02:26) Response to listener comments Tools & Apps (00:02:55) OpenAI Unveils Its New, More Powerful GPT-5.5 Model - The New York Times (00:20:33) xAI Launches grok-voice-think-fast-1.0: Topping τ-voice Bench at 67.3%, Outperforming Gemini, GPT Realtime, and More - MarkTechPost (00:26:00) Claude can now plug directly into Photoshop, Blender, and Ableton | The Verge Projects & Open Source (00:26:38) China’s DeepSeek releases preview of long-awaited V4 model as AI race intensifies (00:44:05) Tencent Unveils Hy3 preview; Model Enhances Agent Capabilities and Real-World Usability - Tencent 腾讯 (00:47:14) ClawMark: A Living-World Benchmark for Multi-Turn, Multi-Day, Multimodal Coworker Agents Applications & Business (00:50:03) Google Plans to Invest Up to $40 Billion in Anthropic (00:53:26) Meta will use hundreds of thousands of AWS Graviton chips (00:56:51) China blocks Meta’s $2 billion takeover of AI startup Manus (00:58:45) OpenAI shakes up partnership with Microsoft, capping revenue share payments (01:04:13) Elon Musk Testifies of AI Risk at Trial, Says OpenAI Tried to ‘Steal’ a Charity - WSJ (01:08:50) Judge rejects DOJ bid to delay Anthropic appeal in Pentagon dispute (01:11:42) Google’s Gemini can now run on a single air-gapped server — and vanish when you pull the plug (01:16:07) DeepMind’s David Silver just raised $1.1B to build an AI that learns without human data | TechCrunch Policy & Safety (01:19:47) Evaluating whether AI models would sabotage AI safety research (01:26:59) LLMs Corrupt Your Documents When You Delegate (01:29:50) Temporal Sparse Autoencoders: Leveraging the Sequential Nature of Language for Interpretability (01:36:53) Memorandum on Adversarial Distillation of American AI Models (01:38:41) Teen boys are dating their AI chatbots—and experts warn it could kill their careers | Fortune (01:40:57) Announcing the Anthropic Economic Index Survey (01:42:21) Scoop: CISA lacks access to Anthropic’s Mythos Synthetic Media & Art (01:45:03) Taylor Swift Files to Trademark Voice and Likeness to Protect Against AI Misuse Research & Advancements (01:46:15) Maximal Brain Damage Without Data or Optimization: Disrupting Neural Networks via Sign-Bit Flips