Here are recent rough notes from the site.
One of the dynamics playing out as model improvement slows, and LLMs move toward being inference-centric, is that new silicon is emerging that runs cooler, with more caching, and much faster speeds.
For example, check Taalas's demo here to see 70x token production speedup, often approach 16,000 tok/s. https://chatjimmy.ai/
Current event remind me, again, of my friend Michael Cembalest's year-ago coinage of "the alchemists" to describe the madcap experimentation in global markets and geopolitics. Instead of transmuting lead into gold, you end up with cracked furnaces and a ruined crucible.
Active ETF launches helped power total ETFs launched in the US in 2025 to an all-time record.
Bloomberg reporting tonight that AI-centric code editor thingie Cursor is now on a $2b annual revenue run rate, which is impressive. Let’s see if this is the peak, like I have argued here—beset by agentic harnesses like Copilot, CCode, and Codex—or if I’m, you know, wrong.