LWiAI Podcast #246 - Gemini 3.5 + Omni, Musk Loses, OpenAI vs Erdős

Google I/O week always turns into a firehose, and this year's edition didn't disappoint. Gemini 3.5 landed with the Flash variant getting most of the attention for raw speed and benchmark wins, but the more interesting release might be Gemini Spark, an always-on agent running on Google Cloud with native MCP tool support. Pair that with Gemini Omni, which stitches together images, audio, and text into generated video, and Google is clearly betting that the next fight isn't about chatbots anymore. It's about agents that stay running and models that handle every modality at once. Throw in Antigravity 2.0's desktop and CLI refresh, new science-focused tooling, and a Genie world model that now simulates real streets using Street View and Waymo data, and you get a sense of just how wide Google is casting its net.

Coding agents had their own mini-arms race this week. Cursor pushed out Composer 2.5, fine-tuned on Moonshot's Kimi K2.5, claiming it matches Opus 4.7 and GPT-5.5 on benchmarks for a fraction of the cost. xAI answered with an early release of Grok Build, its own coding agent, though it arrives amid chatter about staff departures and questions over how efficiently xAI is actually using its compute. Whether Cursor and xAI end up in some kind of partnership is still speculation, but the fact that everyone's racing to ship a coding agent at once says something about where the money's flowing right now.

On the legal and money side, Elon Musk's lawsuit against OpenAI got knocked out on statute-of-limitations grounds, a fairly unceremonious end to a fight that generated a lot more headlines than courtroom drama. Meanwhile Anthropic locked in a $30 billion funding round at a $900 billion valuation and is telling investors it expects its first profitable quarter, a notable milestone for a company that's spent years burning cash to keep pace with OpenAI. Chip maker Cerebras had the opposite kind of good week, with its IPO jumping roughly 90 percent, the biggest debut of the year so far. And reports surfaced that the OpenAI-Apple partnership is fraying, with the two sides reportedly circling something closer to a legal fight than a smooth product integration.

The research corner had a genuinely striking result: OpenAI's models helped crack an 80-year-old Erdős problem in geometry, enough to leave professional mathematicians impressed rather than dismissive. Less flashy but arguably more important were findings on