> Still, is there a world in which we look back 10 years from now and think that we were overvaluing the impact of AI?
AI? Probably no.
LLM-style AI? Yeah, I expect the tablet or blockchain story. A solution in search of a problem, with some very cool niche applications. In this case, Copilot.
Even with the models we have right now, if you built the right app on top of them, they'd be multi-billion dollar businesses that upset incumbents.
Imagine an IDE where you make assertions about the generated code, and it takes those and does a random walk through the latent space until it finds a point that satisfies those assertions. Instead of editing the modified code, you debug by making more assertions or describing the process more accurately.
Imagine art software where you describe what you want, then iteratively add refinements through more description and rough sketch-ups, and then get the final result neatly broken down into semantically consistent layers for a final pass in photoshop.
This stuff is all possible now, and if we see the same or better improvement in models in the next 10 years as we saw in the last 10 years the future versions will be amazing.
What you describe is it what the models can do today, but what they might be able to do with more work. It is possible conceptually now, but getting it to work on a product or MVP level - still risk there.
AI? Probably no.
LLM-style AI? Yeah, I expect the tablet or blockchain story. A solution in search of a problem, with some very cool niche applications. In this case, Copilot.