Why AI Models Aren’t the Product Any More | TWiAI Ep 18

TL;DR

• 1

  Cursor sold for $60 billion because compute plus product data beats model access alone: The panel says Cursor's coding usage, model ambitions, and SpaceX's Colossus compute created a strong fit, especially after Anthropic's Claude Code became a direct competitive threat.

• 2

  Platform dependence keeps ending the same way: Jason Calacanis compares Anthropic versus Cursor to Microsoft versus Lotus 1-2-3 and Facebook versus Zynga, arguing that when you build on someone else's platform, they can study your usage and eventually compete with you.

• 3

  The model is no longer the product: Ali Ansari argues the durable value is in the harness, agent evaluations, UI, memory, and workflow reliability on top of baseline models, which is why countless AI application companies will end up building their own intelligence layers.

• 4

  Services businesses may become the best AI data businesses: Crosby Legal and Micro1 both make the case that real customer work creates the highest-value training data because it is longer-horizon, economically meaningful, and grounded in real-world judgment rather than sandbox tasks.

• 5

  Most AI data spending will shift to agents, not base model labs: Ali predicts that close to 100% of future data spend will accrue at the application and agent layer because there will be orders of magnitude more agents than models, even if frontier labs keep increasing their own spending.

• 6

  Legal AI is still far from replacing lawyers on subjective work: Crosby and Micro1's new redlining benchmark suggests top models are only 10 to 20 percent of the way to a strong human lawyer on multi-turn contract negotiation, especially where style, strategy, and judgment matter.