LWiAI Podcast #255 - Gemini 3.7, Jalapeño, Qwen 3.8, Drones

The 255th episode of the Last Week in AI podcast was recorded on August 26, 2026, and hosted by Andrey Kurenkov and Jeremie Harris. The episode covers the week's AI news across tools, business, open source, policy and research.

On tools and apps, the rundown includes Google's announcement of Gemini 3.7 Flash, arriving three weeks after the previous release, and SpaceXAI's release of Grok 4.6, described as a 500K-context frontier model tuned for long-running agents, coding and knowledge work. The episode also covers Anthropic applying invisible watermarks to Claude's AI text and images, Claude Mythos 5 cybersecurity capabilities, OpenAI enhanced safety features for paid AI tool users, ChatGPT's stricter Teen Mode rollout, and a dedicated Meta AI desktop app for Mac.

Business items include OpenAI's first Jalapeño inference-chip results, which the episode describes as showing industry-leading speed and efficiency in AI inference. Other segments cover departures at OpenAI including a top data center executive, Anthropic hiring a Google chip veteran for a hardware push, Anthropic's annualized revenue rising to $65B, and Thomson Reuters launching an in-house AI model to cut Anthropic costs.

In projects and open source, the episode discusses Qwen 3.8 and how a 27B open model rivals GPT-5.6 and Claude Opus. Policy and safety segments cover a New York Times report that a drone killed three Ukrainians and was guided entirely by AI, OpenAI's new security changes after its AI hacked Hugging Face, and another woman joining a lawsuit accusing Grok of generating CSAM. Research segments include small-scale experiments, stealing reasoning traces from proprietary LLM APIs, and massive activations in hybrid linear attention LLMs. A synthetic media segment covers AI slop and responses from Spotify, LinkedIn and others.