GPT-6 is impressive and underwhelming

Yes, current AI models are very impressive. But I was thinking about if Aayush from 2023 (GPT-4 era) came to the present (GPT-6 era) how he would mostly feel underwhelmed.

Sure, agents can do single-objective long-horizon tasks really, really well. But they tend to oscillate a lot when there are multiple objectives, and outside of the domain of single-objective benchmark tasks, they are not very good. The thing is that you don’t run into tasks like that unless you are deeper into a problem.

Some examples from my own experience:

• Agents still mostly struggle with good browser use.
• They struggle designing documentation with good UI and prose.
• The struggle to work with different tradeoffs associated with building distributed systems.
• They are not very good at verifying (or building systems that can verify) their work.

In 2023, Aayush would have expected GPT-6 to be almost able to bootstrap entire businesses from scratch. He would be shocked by how far we are from that.