4 ms·
Computer chess has been failing for 30 years, until it didn't. Try winning a Go or a chess game against the computer now. There easily might be another architec
by cft 3y ago
Computer chess has been failing for 30 years, until it didn't. Try winning a Go or a chess game against the computer now. There easily might be another architectural find lateral to LLMs that will 10x the code generation quality.
- softirq 3y agoYou can say that about any field. We could invent the elixir of immortality tomorrow, but is that a realistic expectation? The CEO of Nvidia is a smart guy, he's pushing the hype train because his business is riding the wave. But you have to separate hype from an empirical view of what we can actually do today with these tools, versus what hasn't been delivered and is being oversold.
- cft 3y agoI think you're viewing it from the programming bubble. He's not that vested in AI success for programming. Even if AI code generation completely failed, NVIDIA's business is still more than OK because LLMs have a lot of other uses, killing Google search for example. That's not a small niche.
- nyrikki 3y agoChess is a Perfect information game with a finite tree. Very different problem than programming.
- itishappy 3y agoDo chess-bots rely on this? I was under the impression that a full search of the space was infeasible, so our current state-of-the-art approaches use heuristics, bounded search, and learned strategies. In other words, I suspect our current models apply to programming better than we might expect.
- nyrikki 3y agoChess is decidable, it may be PSPACE-hard or EXPTIME-hard, but there are reductions. Entscheidungsproblem and Halt are not decidable in the general case. While you have to find reductions, decidable problems having access to both yes-instances and no-instances makes it easier to find them.
- itishappy 3y agoKnowing that we don't approach either by attempting to fully solve them, does that change the architecture, or just the difficulty?
- softirq 3y agoChess is a bounded, non-moving target. Think about the difference between chess in the 1970s and today, and compare that to the same time period with programming. Chess is a single game whereas programming is a federation of tools, protocols, and standards that are ever evolving. They're not comparable in any sense.
- itishappy 3y agoI don't think that's a particularly relevant metric, as we can easily restrict programming to languages like Lisp/Pascal from the 70s, and the landscape doesn't change much. I'd also suggest that our chess bots have evolved dramatically in that time. Deep Blue works very differently than AlphaZero, for example. Deep Blue might not be suited to code generation, but AlphaCode spawned from AlphaZero.