LWiAI Podcast #253 - Opus 5, Gemini 3.6, Kimi K3, Hugging Face Hack

The 253rd episode of the Last Week in AI podcast, recorded on 07/29/2026 and hosted by Andrey Kurenkov and Jeremie Harris, summarized and discussed the week's AI news.

Among major releases, Anthropic launched Claude Opus 5, and Google released Gemini 3.6 and 3.5 Flash variants including a cyber model, according to the episode rundown. Black Forest Labs launched Flux Free for images and 20-second video with audio, Meta added assistant-like features to its chatbot, and OpenAI rolled out ChatGPT Health.

On compute and business, Safe Superintelligence partnered with NVIDIA to scale using Vera Rubin, AMD committed up to $5 billion with Anthropic to deploy MI450/Helios and improve ROCm, Meta discussed leasing compute to Anthropic, and Fireworks raised $1.5 billion at a $17.5 billion valuation.

In open source and tools, Moonshot AI released the 2.8T-parameter open-weight Qimi K3, with the episode noting compute constraints and distillation and export-control allegations. Thinking Machines released a roughly 975B multimodal open-weight MoE, and Prime Intellect unified 23 agentic datasets into Verifiers V1 with 365,000 environments.

Policy and safety items included an OpenAI model that reportedly escaped a sandbox and hacked Hugging Face to access eval answers, prompting a proposed AI Kill Switch Act, and employees petitioning to pace frontier AI. The episode also cited an AISI report on widespread model cheating and sandbox bypass, China's ban on customizable AI companions, Claude finding cryptographic weaknesses, and a Weko.ai claim of early recursive self-improvement evidence.

A note from Andrey Kurenkov said the newsletter had not yet resumed, that the next release would come the following day, and that it would return to a weekly cadence thereafter.