Learning on the Job: The Future of Post-Training — Raymond Feng, Applied Compute

Learning from real customer interactions creates a problem that controlled AI training can avoid: the same situation cannot always be tried again. In Applied Compute's presentation, training moves into a business's existing agent workflow, with model requests and responses recorded for learning. A simulated task can be reset to compare several attempted solutions. A completed customer-support conversation supplies only the response that actually happened; it cannot reveal how that customer would have reacted to every alternative. This makes extracting useful feedback from ordinary work a central research problem. The presentation describes automated selection of training examples and learning from written feedback as directions toward that goal.

“if the agent learns the exact environment distribution, why don't we just use that for our training?”

Raymond Feng

“it makes a lot of sense that because there's so many potholes, the model doesn't want to run for that long uh because it might fall in a pothole and then get a zero reward for the rollout”

Raymond Feng

“it will actually just be incentivized to like abuse the tool calls and just like call a lot of them in quick succession and try to time out the sandbox”

Raymond Feng

“you're kind of playing a game of Whac-A-Mole where as soon as a new thing pops up, you need to scramble and like create new data or new environments”

Raymond Feng

“AI is at the cusp of a new period in which experience will become the dominant medium of improvement and ultimately dwarf the scale of human data used in today's systems”

Raymond Feng