Episode 31:21:14

If you ask potential hires to steal IP ...

Moonshot AI's Kimi K3 is hitting frontier-level benchmarks at a fraction of the price, and Will, Jim and Alex reckon the real moat for labs like Anthropic is now the harness, not the model. They also pick apart Thinking Machines Lab's Inkling, a 975 billion parameter fine-tuneable model they argue is five years ahead of what businesses can actually use, and WeCo's claimed recursive self-improvement, which managed improvements in just 7 of 100 iterations. From there it's cost per task versus cost per token, with Databricks finding Opus worked out cheaper than Sonnet on a per-task basis, plus AI washing, Demis Hassabis backing a FINRA-style regulator, Australia's proposed national AI framework, and Apple suing OpenAI. The thread running through it all: if you can't measure the task, you can't pick the model, price the work, or claim the ROI.

Will Turner

Founder of Beyond Data and moderator of Business Idiots.

Jim Lovell

Founder and Head of Solutions at QuantalAI.

Alex Stenlake

Co-host of Business Idiots.

Subscribe

New essays and episodes, sent when there is something worth reading. No noise, unsubscribe anytime.