4 ms·

We’ll see - but given the gap between chatgpt 3 and 4, I think AIs will be competitive with mid level programmers by the end of the decade. I’d be surprised if

We’ll see - but given the gap between chatgpt 3 and 4, I think AIs will be competitive with mid level programmers by the end of the decade. I’d be surprised if they aren’t. The training systems we use for LLMs are still so crude. ChatGPT has never interacted with a compiler. Imagine learning to write code by only reading (quite small!) snippets on GitHub. That’s the state llms are in now. It’s only a matter of time before someone figures out how to put a compiler in a reinforcement learning loop while training an LLM. I think the outcome of that will be something that can program orders of magnitude better. I’ll do it eventually if nobody else does it first. We also need to solve the “context” problem - but that seems tractable to me too. For all the computational resources they use to do training and inference, our LLMs are still incredibly simple. The fact they can already code so well is a very strong hint for what is to come.

kenjackson 3y ago
Why do you say the snippets are small? They don’t get trained on the full source files?
- mike_hearn 3y ago
  Nope. LLMs have a limited context window partly because that's the chunk size with which they're presented with data to learn during training (and partly for computational complexity reasons). One of the reasons I'm feeling very bullish on LLMs is because if you look at the exact training process being used it's full of what feels like very obvious low hanging fruit. I suspect a part of the reason that training them is so expensive is that we do it in really dumb ways that would sound like a dystopian hell if you described it to any actual teacher. The fact that we can get such good results from such a terrible training procedure by just blasting through it with computational brute force, strongly suggests that much better results should be possible once some of that low hanging fruit starts being harvested.
DrSiemer 3y ago
With today's mid level programmers, yes. But by that time, many of today's mid level programmers will be able to do stuff high level programmers do today. Many people underestimate an LLM's most powerful feature when comparing it with something like Stackoverflow: the ability to ask followup questions and immediately get clarification on anything that is unclear. I wish I had had access to LLM's when I was younger. So much time wasted on repetitive, mundane in-between code...
- StableAlkyne 3y ago
  > the ability to ask followup questions and immediately get clarification on anything that is unclear. Not only that, but it has the patience of a saint. It never makes you beg for a solution because it thinks there's an XY problem. It never says "RTFM" before posting an irrelevant part of the documentation because it only skimmed your post. It never says "Why would you use X in 2023? Everyone is using framework Y, I would never hire anyone using X." The difference comes down to this: unlike a human, it doesn't have an ego or an unwarranted feeling of superiority because it learned an obscure technology. It just gives you an answer. It might tell you why what you're doing is suboptimal, it might hallucinate an answer that looks real but isn't, but at least you don't have to deal with the the worst parts of asking for help online.
  - josephg 3y ago
    Yeah. You also don't have to wait for an answer or interrupt someone to get that answer. But - in the history of AIs written for chess and go, there was a period for both games where a human playing with an AI could beat either a human playing alone or an AI playing alone. I suspect we're in that period for programming now, where a human writing code with an AI beats an AI writing code alone, and a human writing code alone. For chess and go, after a few short years passed, AIs gained nothing by having a human suggesting moves. And I think we'll see the same before long with AI programmers.
    - DrSiemer 3y ago
      Good riddance. I can finally get started on the massive stockpile of potential projects that I never had time for until now. It's a good time to be in the section of programmers that see writing code as a means to an end and not as the goal itself. It does surprise me that so many programmers, whose mantra usually is "automate all the things", are so upset now that all the tedious stuff can finally be automated in one big leap. Just imagine all the stuff we can do when we are not wasting our resources finding obscure solutions to deeply burried environment bugs or any of the other pointless wastes of time!
arthur_sav 3y ago
Imagine being able train a model that mimics a good programmer. It would talk and program in the principles of that programmer's philosophy.

kenjackson 3y ago

Why do you say the snippets are small? They don’t get trained on the full source files?

mike_hearn 3y ago
Nope. LLMs have a limited context window partly because that's the chunk size with which they're presented with data to learn during training (and partly for computational complexity reasons). One of the reasons I'm feeling very bullish on LLMs is because if you look at the exact training process being used it's full of what feels like very obvious low hanging fruit. I suspect a part of the reason that training them is so expensive is that we do it in really dumb ways that would sound like a dystopian hell if you described it to any actual teacher. The fact that we can get such good results from such a terrible training procedure by just blasting through it with computational brute force, strongly suggests that much better results should be possible once some of that low hanging fruit starts being harvested.

DrSiemer 3y ago

With today's mid level programmers, yes. But by that time, many of today's mid level programmers will be able to do stuff high level programmers do today. Many people underestimate an LLM's most powerful feature when comparing it with something like Stackoverflow: the ability to ask followup questions and immediately get clarification on anything that is unclear. I wish I had had access to LLM's when I was younger. So much time wasted on repetitive, mundane in-between code...

StableAlkyne 3y ago
> the ability to ask followup questions and immediately get clarification on anything that is unclear. Not only that, but it has the patience of a saint. It never makes you beg for a solution because it thinks there's an XY problem. It never says "RTFM" before posting an irrelevant part of the documentation because it only skimmed your post. It never says "Why would you use X in 2023? Everyone is using framework Y, I would never hire anyone using X." The difference comes down to this: unlike a human, it doesn't have an ego or an unwarranted feeling of superiority because it learned an obscure technology. It just gives you an answer. It might tell you why what you're doing is suboptimal, it might hallucinate an answer that looks real but isn't, but at least you don't have to deal with the the worst parts of asking for help online.
- josephg 3y ago
  Yeah. You also don't have to wait for an answer or interrupt someone to get that answer. But - in the history of AIs written for chess and go, there was a period for both games where a human playing with an AI could beat either a human playing alone or an AI playing alone. I suspect we're in that period for programming now, where a human writing code with an AI beats an AI writing code alone, and a human writing code alone. For chess and go, after a few short years passed, AIs gained nothing by having a human suggesting moves. And I think we'll see the same before long with AI programmers.
  - DrSiemer 3y ago
    Good riddance. I can finally get started on the massive stockpile of potential projects that I never had time for until now. It's a good time to be in the section of programmers that see writing code as a means to an end and not as the goal itself. It does surprise me that so many programmers, whose mantra usually is "automate all the things", are so upset now that all the tedious stuff can finally be automated in one big leap. Just imagine all the stuff we can do when we are not wasting our resources finding obscure solutions to deeply burried environment bugs or any of the other pointless wastes of time!

arthur_sav 3y ago

Imagine being able train a model that mimics a good programmer. It would talk and program in the principles of that programmer's philosophy.