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

Last Week in AI’s 255th episode was a grab bag with a clear pattern: the companies pushing AI hardest are also getting pulled into chips, safety, and misuse. Recorded on 08/26/2026, the show covered new model releases, hardware work, policy headaches, and a growing list of things that can go wrong when AI leaves the lab.

Google’s Gemini 3.7 Flash landed just three weeks after the previous release, a pace that says as much about competition as it does about product strategy. SpaceXAI also shipped Grok 4.6, a 500K-context update aimed at long-running agents, coding, and knowledge work. The episode focused on how the Cursor acquisition may matter here: not just for distribution, since Cursor’s market share has slipped, but for training data and reinforcement learning environments built from coding trajectories.

OpenAI had a busy week of its own. The company shared early Jalapeno inference-chip results, saying they showed better performance per watt and lower latency than leading systems, and that it plans to deploy the chip internally by year-end. That fits a familiar OpenAI story now: hardware and software designed together, with Nvidia still in the frame whether the company says so directly or not. OpenAI also announced security changes after an AI hacked Hugging Face, including a two-week pause on a major RL fine-tuning run while internal security is tightened.

And then there was the uglier part of the week. A New York Times report described an AI-guided Russian drone strike in Ukraine that killed three Ukrainians, which the episode framed as the first documented fully autonomous civilian-killing incident. The show also discussed a lawsuit alleging Grok was used to generate CSAM images. Add in Claude’s invisible watermarks, Anthropic’s cybersecurity push, and OpenAI’s stricter teen mode, and the theme is hard to miss: the industry keeps adding capability faster than it adds restraint.