Video / Short
AI Agents: Where They Actually Work (And Where They're Pure Hype)
Mykhailo Kushnir shares a builder's view of where AI agents work in practice and where business expectations are still mostly hype.
Video summary
Mykhailo Kushnir offers a builder’s view of AI agents. He expects more use cases in which agents drive conversations and predict possible next decisions, rather than waiting for a person to initiate every request.
The clip also raises the problem of hallucinations and argues that human expertise remains important for describing failures and rules precisely. In that account, a model can respond quickly once a failure is properly defined, but reliable work still depends on people specifying what good and bad behaviour look like.
Video transcript
Read the spoken content without loading the YouTube player.
I think we'll have a lot more use cases where agents are driving the conversations. Right now, it feels like we are the one triggering it, we're the one asking the questions. I think in 5 years we'll have more agents predicting our next decisions.
What is it that makes the LLM's reward system geared to hallucinating rather than giving you the correct answer? How do they motivate that?
If you can properly describe to AI model what the failure is, it can then really quickly try to admit that failure and not only fake admitting it, but but actually do the the proper fix. I think human expertise these days is required in describing rules.
AI, is it hype or help?