[softirq]

No-code isn't a new concept, and there's a reason why all past attempts have failed, or why people still pay web developers despite the existence of tools like square space. Nothing about the LLMs of today suggests they have solved the no-code problems or will radically displace coding. They generate bad, oftentimes incorrect code for well trodden paths, while struggling to solve novel problems or work in private or unique code bases. They do not easily keep up with new trends or tools. They do not offer the type of semantic understanding that is necessary to work in a logic based field.

LLMs are nothing more than an alternative take on auto-complete, a feature that has been around forever and doesn't radically change programming. It will speed up good programmers to some degree and probably lead to bugs and more bad code from everyone else.

This is yet another hype cycle overselling a modest advancement in technology.

[cft]

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]

You 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]

I 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.

[softirq]

The post is about the demise of coding. I'm only responding to the topic of the discussion.

[cft]

My previous response was refuting your statement: "The CEO of Nvidia is a smart guy, he's pushing the hype train because his business is riding the wave"

[nyrikki]

Chess is a Perfect information game with a finite tree.

Very different problem than programming.

[itishappy]

Do 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]

Chess 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]

Knowing that we don't approach either by attempting to fully solve them, does that change the architecture, or just the difficulty?

[softirq]

Chess 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]

I 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.