Post in: SiliconANGLE theCUBE
The gap between what an AI agent can almost do and what an enterprise will actually deploy is where most agentic projects stall. In this theCUBE / NYSE Wired conversation, Pokee AI founder Bill Zhu — who led applied reinforcement learning at Meta before striking out on his own — argues that closing the final 5–10% is a rubric problem, not a model-size problem: enterprises are punishing on the details of a slide deck or an Excel analysis, and RL tuned to each customer’s rubric is what gets an agent over that bar. Pokee‘s pitch is to own the last mile end to end: reinforcement learning trained on the enterprise’s own standards, an air-gapped sandbox with data-loss prevention and a credentials vault to keep agents from going rogue, and its own long-context model built to undercut the token costs that can hit $200 a day per user.
Watch the full interview: