Kimi K3.1 Pro topped the Artificial Analysis Intelligence Index above GPT-6 Astra, but the launch was messy. Moonshot AI ran a paid promotion campaign that allegedly instructed participants not to disclose sponsorship, a practice Nathan Lambert and Florian Brand call out as industry-wide and indefensible. Model weights are withheld until November 4th, part of a broader shift toward exclusive access windows that now define frontier releases.
The more urgent half of this episode is cybersecurity. Nathan and Florian work through why open-weight models like Kimi K3.1 Pro scoring high on offensive cyber benchmarks landed differently with security professionals than with AI researchers. The conversation pulls in real incidents: LiteLLM's security breach, PHP2Shell exploits, and abliteration.ai serving safety-stripped DeepSeek models via public API. The gap between how cyber experts and AI experts frame risk is the sharpest thing in this episode and worth hearing in full.
The episode closes on Anthropic's published research showing the company is reducing junior technical hires as AI handles more entry-level work, and Florian's use of Muse as a productivity tool. The links section is unusually dense, including Joshua Saxe's Substack on AI and cybersecurity and Azeem Azhar's piece framing the current moment as the internet's Wuhan moment. Start at the 11:23 mark if you want the cyber reality check without the benchmark preamble.
[READ ORIGINAL →]