Software as the Shortcut to AGI: Eiso Kant (poolside) at CVAI London

Eiso Kant, cofounder of coding model-maker poolside, said his company makes heavy use of synthetic data for model training — essentially creating problems and workflows and training the model to solve them. The challenge, he said, is to develop “a set of tasks that requires extra information, but are not so difficult that they’re impossible to solve.” More compute will go to reinforcement learning than any other phase of model training as the use of synthetic data ramps up, he predicted. AI agents promise to supercharge that process, he added, with RL advancing from “a single-shot approach to an agent approach.” He envisions a software development environment where there is first a thin agent layer sitting on top of an editor, but there will also be other “surface areas” that are “headless,” with agents running the show.