Login
You're viewing the discuss.systems public feed.
  • Non-Euclidean unsupported unnormalsteve
    Oct 1, 2023, 2:06 PM

    I think the thing that bugs me about the LLM/AGI/GPT/BBQ craze is that it used to be kinda tricky to make computers do what you wanted them to do, but once you did it was absolutely completely reliable, and when they didn't do what you wanted, you could figure out why and fix it.

    The new AI stuff throws away everything that computers used to excel at and just tries to give us faster shittier human minds.

    The reliability, predictability, and understandability of conventional computation is the great strength of computers, and a marvelous complement to the way that the human mind works. Throwing that out in the service of going faster is just depressing.

    💬 9🔄 164⭐ 291

Replies

  • Non-Euclidean unsupported unnormalsteve
    Oct 1, 2023, 2:11 PM

    I think that there _is_ an exciting and interesting path forward that synthesizes the two modes of operation, but very few people are doing visible and compelling work on it (because it's really hard!)

    💬 1🔄 1⭐ 15
  • 💬 0🔄 0⭐ 1
  • Félixfay59@tech.lgbt
    Oct 1, 2023, 2:32 PM

    @steve what I really don’t like about the way LLMs are pitched is that very few people seem to understand what they’re *not* good at. The way it’s deployed and the way I hear excited people talk about it, it’s like all problems it has today could be solved with more training or bigger models. I think the big reason is that LLMs are easy to anthropomorphize by design and it’s more or less how people work

    💬 1🔄 0⭐ 2
  • Grassroots Joejoeinwynnewood@mstdn.social
    Oct 2, 2023, 10:10 AM

    @fay59 @steve

    What I really don't like is that every issued prompt uses a gazillion Kwh of electricity and spews untold tons of carbon into the atmosphere.

    Absent a really good reason, no one should touch an LLM.

    Spitting our piles of incorrect code is the antithesis of a good reason.

    💬 1🔄 0⭐ 0
  • Cassandrichdalias@hachyderm.io
    Oct 2, 2023, 10:17 AM

    @joeinwynnewood @fay59 @steve Even piles of correct code would be dubious because you should not have piles of code at all. If the volume of code is too large to be comprehensible and not all sorts of garbage boilerplate an automated process could write, it's a definite sign your abstraction level is completely wrong for the task.

    💬 1🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • 💬 0🔄 0⭐ 1
  • Yessica Haircutshadowfacts@social.shadowfacts.net
    Oct 1, 2023, 6:16 PM
    @steve there was a really great essay I read a while ago about how computers are fundamentally knowable, and that is why the author loved software engineering. despite the fact that they're immensely complicated systems that sometimes behave weirdly, they are still understandable. it saddens me to lose that
    💬 0🔄 0⭐ 1
  • Mark Pauleyunsaturated
    Oct 1, 2023, 7:36 PM

    @steve I think the problem is that neural nets fail in a continuous manner (rather than discrete), and so they let people use them incorrectly with various forms of partial-success. The processes are still happening on a computer and are therefore repeatable and analyzable. Just nobody is building this stuff in that way yet, and instead many people are just flopping about making a mess.

    💬 0🔄 0⭐ 3
  • Michaelmikee@chaos.social
    Oct 2, 2023, 1:01 PM

    @steve Well tbf, LLMs aren't supposed to *replace* traditional computing. I think they can have their justified use cases, we just need to understand what they can and can't (or should and shouldn't) do.

    💬 0🔄 0⭐ 0
  • Happosai 八宝斎happmacdonald@retrochat.online
    Oct 2, 2023, 8:41 PM

    @steve I'm just amazed nobody has perfected feeding the bot code output into a test suite.

    Then you just write the test suite (which you were going to have to do anyway) and let the bot iterate over "can I get the code to pass the test suite" instead of iterate over "is the human personally satisfied that the code is correct".

    Sure, that is more likely to result in more interesting problems and perspectives than it is in good code, but I want to see what those foibles are gonna look like too.😁

    💬 0🔄 0⭐ 0
  • Alvaroalvaro@social.graves.cl
    Oct 28, 2023, 10:30 AM

    @steve@discuss.systems I would say that that shipped sailed decades ago after anything more complex than linear regression and decision trees, but I agree now it is even worse.

    💬 1🔄 0⭐ 0
  • Non-Euclidean unsupported unnormalsteve
    Oct 28, 2023, 12:03 PM

    @alvaro even when I was doing statistical language models as an undergraduate around 2000, we could still look at the parse trees and the grammar rules and pretty quickly identify exactly why the system generated a bad parse.

    💬 1🔄 0⭐ 0
  • Non-Euclidean unsupported unnormalsteve
    Oct 28, 2023, 12:06 PM

    @alvaro you’re right that it was already not nearly so simple to _fix_ the issue, but we also weren’t trying to insert those models as the primary mechanism for interacting with a computer.

    💬 1🔄 0⭐ 0
  • Alvaroalvaro@social.graves.cl
    Oct 28, 2023, 1:14 PM

    @steve@discuss.systems I was mentioning other problems but for NLP probably the death of the model transparency started with word embeddings and the reliance on the biases of random data as Marcus mentioned

    💬 0🔄 0⭐ 1
  • Robert J. Bergerrberger@hachyderm.io
    Oct 28, 2023, 7:40 PM

    @steve I have found copilot useful as an in editor (emacs of course) assistant. Type a bit and its auto suggestions have been surprisingly helpful. Never really more than a single block of a few lines of code that is at least useful as a starting point or many times correct.

    It’s particularly great when you aren’t the expert (but not a noobie) in the language like for me terraform. But its been helpful with Python too.
    I wouldn’t have it do more than that though.

    💬 0🔄 0⭐ 0