LWiAI Podcast #243 - GPT 5.5, DeepSeek V4, AI safety sabotage

The 243rd episode of the Last Week in AI podcast, recorded on 04/29/2026 and hosted by Andrey Kurenkov and Jeremie Harris, summarizes and discusses recent AI news.

OpenAI released GPT-5.5 with what the episode describes as strong coding-oriented improvements. It includes a system card discussing chain-of-thought monitorability and misalignment testing, carries higher pricing than GPT-5.4, and has notable quirks such as 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 in Pro and Flash versions, featuring MoE scaling and 1M-token context via hybrid/compressed attention changes. Tencent released a Hunyuan 3 preview with weaker benchmark performance, and a new long-horizon agent benchmark, Clawmark, shows low task success rates.

Business, legal, and policy items include Google's planned up-to-$40B investment and 5GW compute commitment to Anthropic, Meta's AWS Gravitron deal, China blocking Meta's Manus acquisition, a revamped OpenAI–Microsoft agreement, and ongoing Musk–OpenAI trial developments. New safety and security research covered includes work on sabotage, document degradation under delegation, and bit-flip attacks.

The episode also lists timestamped segments on tools and apps, projects and open source, applications and business, policy and safety, synthetic media and art, and research and advancements.