5 ms·

AI can very efficiently apply common patterns to vast amounts of code, but it has no inherent "idea" of what it's doing. Here's a fresh example that I stumbled

by sysmax 1y ago

AI can very efficiently apply common patterns to vast amounts of code, but it has no inherent "idea" of what it's doing. Here's a fresh example that I stumbled upon just a few hours ago. I needed to refactor some code that first computes the size of a popup, and then separately, the top left corner. For brevity, one part used an "if", while the other one had a "switch": if (orientation == Dock.Left || orientation == Dock.Right) size = /* horizontal placement */ else size = /* vertical placement */ var point = orientation switch { Dock.Left => ... Dock.Right => ... Dock.Top => ... Dock.Bottom => ... }; I wanted the LLM to refactor it to store the position rather than applying it immediately. Turns out, it just could not handle different things (if vs. switch) doing a similar thing. I tried several variations of prompts, but it very strongly leaning to either have two ifs, or two switches, despite rather explicit instructions not to do so. It sort of makes sense: once the model has "completed" an if, and then encounters the need for a similar thing, it will pick an "if" again, because, well, it is completing the previous tokens. Harmless here, but in many slightly less trivial examples, it would just steamroll over nuance and produce code that appears good, but fails in weird ways. That said, splitting tasks into smaller parts devoid of such ambiguities works really well. Way easier to say "store size in m_StateStorage and apply on render" than manually editing 5 different points in the code. Especially with stuff like Cerebras, that can chew through complex code at several kilobytes per second, expanding simple thoughts faster than you could physically type them.

gametorch 1y ago
[flagged]
- npinsker 1y ago
  Sweeping generalizations about how LLMs will always (someday) be able to do arbitrary X, Y, and Z don't really capture me either
  - gametorch 1y ago
    [flagged]
    - agentultra 1y ago
      Until the day that thermodynamics kicks in. Or the current strategies to scale across boards instead of chips gets too expensive in terms of cost, capital, and externalities.
      - gametorch 1y ago
        I mean fair enough, I probably don't know as much about hardware and physics as you
        agentultra 1y ago
        Just pointing out that there are limits and there’s no reason to believe that models will improve indefinitely at the rates we’ve seen these last couple of years.
        soulofmischief 1y ago
        There is reason to believe that humans will keep trying to push the limitations of computation and computer science, and that recent advancements will greatly accelerate our ability to research and develop new paradigms. Look at how well Deepseek performed with the limited, outdated hardware available to its researchers. And look at what demoscene practitioners have accomplished on much older hardware. Even if physical breakthroughs ceased or slowed down considerably, there is still a ton left on the table in terms of software optimization and theory advancement. And remember just how young computer science is as a field, compared to other human practices that have been around for hundreds of thousands of years. We have so much to figure out, and as knowledge begets more knowledge, we will continue to figure out more things at an increasing pace, even if it requires increasingly large amounts of energy and human capital to make a discovery. I am confident that if it is at all possible to reach human-level intelligence at least in specific categories of tasks, we're gonna figure it out. The only real question is whether access to energy and resources becomes a bigger problem in the future, given humanity's currently extraordinarily unsustainable path and the risk of nuclear conflict or sustained supply chain disruption.
- DataDaoDe 1y ago
  The interesting questions happen when you define X, Y and Z and time. For example, will llms be able to solve the P=NP problem in two weeks, 6 months, 5 years, a century? And then exploring why or why not
- guappa 1y ago
  If you need a model per task, we're very far from AGI.
- sysmax 1y ago
  I am working on a GUI for delegating coding tasks to LLMs, so I routinely experiment with a bunch of models doing all kinds of things. In this case, Claude Sonnet 3.7 handled it just fine, while Llama-3.3-70B just couldn't get it. But that is literally the simplest example that illustrates the problem. When I tried giving top-notch LLMs harder tasks (scan an abstract syntax tree coming from a parser in a particular way, and generate nodes for particular things) they completely blew it. Didn't even compile, let alone logical errors and missed points. But once I broke down the problem to making lists of relevant parsing contexts, and generating one wrapper class at a time, it saved me a whole ton of work. It took me a day to accomplish what would normally take a week. Maybe they will figure it out eventually, maybe not. The point is, right now the technology has fundamental limitations, and you are better off knowing how to work around them, rather than blindly trusting the black box.
  - gametorch 1y ago
    Yeah exactly. I think it's a combination of 1) wrong level of granularity in prompting 2) lack of engineering experience 3) autistic rigidity regarding a single hallucination throwing the whole experience off 4) subconscious anxiety over the threat to their jerbs 5) unnecessary guilt over going against the tide; anything pro AI gets heavily downvoted on Reddit and is, at best, controversial as hell here I, for one, have shipped like literally a product per day for the last month and it's amazing. Literally 2,000,000+ impressions, paying users, almost 100 sign ups across the various products. I am fucking flying. Hit the front page of Reddit and HN countless times in the last month. Idk if I break down the prompts better or what. But this is production grade shit and I don't even remember the last time I wrote more than two consecutive lines of code.
    - sysmax 1y ago
      If you are launching one product per day, you are using LLMs to convert unrefined ideas into proof-of-concept prototypes. That works really well, that's the kind of work that nobody should be doing by hand anymore. Except, not all work is like that. Fast-forward to product version 2.34 where a particular customer needs a change that could break 5000 other customers because of non-trivial dependencies between different parts of the design, and you will be rewriting the entire thing by humans or having it collapse under its own weight. But out of 100 products launched on the market, only 1 or 2 will ever reach that stage, and having 100 LLM prototypes followed by 2 thoughtful redesigns is way better than seeing 98 human-made products die.
- realusername 1y ago
  Maybe it will improve or maybe not, I feel like we're at the same point as the first release of Cursor, in 2023
soulofmischief 1y ago
> AI can very efficiently apply common patterns to vast amounts of code, but it has no inherent "idea" of what it's doing. AI stands for Artificial Intelligence. There are no inherent limits around what AI can and can't do or comprehend. What you are specifically critiquing is the capability of today's popular models, specifically transformer models, and accompanying tooling. This is a rapidly evolving landscape, and your assertions might no longer be relevant in a month, much less a year or five years. In fact, your criticism might not even be relevant between current models. It's one thing to speak about idiosyncrasies between models, but any broad conclusions drawn outside of a comprehensive multi-model review with strict procedure and controls is to be taken with a massive grain of salt, and one should be careful to avoid authoritative language about capabilities. It would be useful to be precise in what you are critiquing, so that the critique actually has merit and applicability. Even saying "LLM" is a misnomer, as modern transformer models are multi-modal and trained on much more than just textual language.
- mattbee 1y ago
  What a ridiculous response, to scold the GP for criticising today's AI because tomorrow's might be better. Sure, it might! But it ain't here yet buddy. Lots of us are interested in technology that's actually available, and we can all read date stamps on comments.
  - soulofmischief 1y ago
    You're projecting that I am scolding OP, but I'm not. My language was neutral and precise. I presented no judgment, but gave OP the tools to better clarify their argument and express valid, actionable criticism instead of wholesale criticizing "AI" in a manner so imprecise as to reduce the relevance and effectiveness of their argument. > But it ain't here yet buddy . . . we can all read date stamps on comments. That has no bearing on the general trajectory that we are currently on in computer science and informatics. Additionally, your language is patronizing and dismissive, trading substance for insult. This is generally frowned upon in this community. You failed to actually address my comment, both by failing to recognize that it was mainly about using the correct terminology instead of criticizing an entire branch of research that extends far beyond transformers or LLMs, and by failing to establish why a rapidly evolving landscape does not mean that certain generalizations cannot yet be made, unless they are presented with several constraints and caveats, which includes not making temporally-invariant claims about capabilities. I would ask that you reconsider your approach to discourse here, so that we can avoid this thread degenerating into an emotional argument.

by sysmax 1y ago

gametorch 1y ago

[flagged]

npinsker 1y ago
Sweeping generalizations about how LLMs will always (someday) be able to do arbitrary X, Y, and Z don't really capture me either
- gametorch 1y ago
  [flagged]
  - agentultra 1y ago
    Until the day that thermodynamics kicks in. Or the current strategies to scale across boards instead of chips gets too expensive in terms of cost, capital, and externalities.
    - gametorch 1y ago
      I mean fair enough, I probably don't know as much about hardware and physics as you
      - agentultra 1y ago
        Just pointing out that there are limits and there’s no reason to believe that models will improve indefinitely at the rates we’ve seen these last couple of years.
        soulofmischief 1y ago
        There is reason to believe that humans will keep trying to push the limitations of computation and computer science, and that recent advancements will greatly accelerate our ability to research and develop new paradigms. Look at how well Deepseek performed with the limited, outdated hardware available to its researchers. And look at what demoscene practitioners have accomplished on much older hardware. Even if physical breakthroughs ceased or slowed down considerably, there is still a ton left on the table in terms of software optimization and theory advancement. And remember just how young computer science is as a field, compared to other human practices that have been around for hundreds of thousands of years. We have so much to figure out, and as knowledge begets more knowledge, we will continue to figure out more things at an increasing pace, even if it requires increasingly large amounts of energy and human capital to make a discovery. I am confident that if it is at all possible to reach human-level intelligence at least in specific categories of tasks, we're gonna figure it out. The only real question is whether access to energy and resources becomes a bigger problem in the future, given humanity's currently extraordinarily unsustainable path and the risk of nuclear conflict or sustained supply chain disruption.
DataDaoDe 1y ago
The interesting questions happen when you define X, Y and Z and time. For example, will llms be able to solve the P=NP problem in two weeks, 6 months, 5 years, a century? And then exploring why or why not
guappa 1y ago
If you need a model per task, we're very far from AGI.
sysmax 1y ago
I am working on a GUI for delegating coding tasks to LLMs, so I routinely experiment with a bunch of models doing all kinds of things. In this case, Claude Sonnet 3.7 handled it just fine, while Llama-3.3-70B just couldn't get it. But that is literally the simplest example that illustrates the problem. When I tried giving top-notch LLMs harder tasks (scan an abstract syntax tree coming from a parser in a particular way, and generate nodes for particular things) they completely blew it. Didn't even compile, let alone logical errors and missed points. But once I broke down the problem to making lists of relevant parsing contexts, and generating one wrapper class at a time, it saved me a whole ton of work. It took me a day to accomplish what would normally take a week. Maybe they will figure it out eventually, maybe not. The point is, right now the technology has fundamental limitations, and you are better off knowing how to work around them, rather than blindly trusting the black box.
- gametorch 1y ago
  Yeah exactly. I think it's a combination of 1) wrong level of granularity in prompting 2) lack of engineering experience 3) autistic rigidity regarding a single hallucination throwing the whole experience off 4) subconscious anxiety over the threat to their jerbs 5) unnecessary guilt over going against the tide; anything pro AI gets heavily downvoted on Reddit and is, at best, controversial as hell here I, for one, have shipped like literally a product per day for the last month and it's amazing. Literally 2,000,000+ impressions, paying users, almost 100 sign ups across the various products. I am fucking flying. Hit the front page of Reddit and HN countless times in the last month. Idk if I break down the prompts better or what. But this is production grade shit and I don't even remember the last time I wrote more than two consecutive lines of code.
  - sysmax 1y ago
    If you are launching one product per day, you are using LLMs to convert unrefined ideas into proof-of-concept prototypes. That works really well, that's the kind of work that nobody should be doing by hand anymore. Except, not all work is like that. Fast-forward to product version 2.34 where a particular customer needs a change that could break 5000 other customers because of non-trivial dependencies between different parts of the design, and you will be rewriting the entire thing by humans or having it collapse under its own weight. But out of 100 products launched on the market, only 1 or 2 will ever reach that stage, and having 100 LLM prototypes followed by 2 thoughtful redesigns is way better than seeing 98 human-made products die.
realusername 1y ago
Maybe it will improve or maybe not, I feel like we're at the same point as the first release of Cursor, in 2023

soulofmischief 1y ago

> AI can very efficiently apply common patterns to vast amounts of code, but it has no inherent "idea" of what it's doing. AI stands for Artificial Intelligence. There are no inherent limits around what AI can and can't do or comprehend. What you are specifically critiquing is the capability of today's popular models, specifically transformer models, and accompanying tooling. This is a rapidly evolving landscape, and your assertions might no longer be relevant in a month, much less a year or five years. In fact, your criticism might not even be relevant between current models. It's one thing to speak about idiosyncrasies between models, but any broad conclusions drawn outside of a comprehensive multi-model review with strict procedure and controls is to be taken with a massive grain of salt, and one should be careful to avoid authoritative language about capabilities. It would be useful to be precise in what you are critiquing, so that the critique actually has merit and applicability. Even saying "LLM" is a misnomer, as modern transformer models are multi-modal and trained on much more than just textual language.

mattbee 1y ago
What a ridiculous response, to scold the GP for criticising today's AI because tomorrow's might be better. Sure, it might! But it ain't here yet buddy. Lots of us are interested in technology that's actually available, and we can all read date stamps on comments.
- soulofmischief 1y ago
  You're projecting that I am scolding OP, but I'm not. My language was neutral and precise. I presented no judgment, but gave OP the tools to better clarify their argument and express valid, actionable criticism instead of wholesale criticizing "AI" in a manner so imprecise as to reduce the relevance and effectiveness of their argument. > But it ain't here yet buddy . . . we can all read date stamps on comments. That has no bearing on the general trajectory that we are currently on in computer science and informatics. Additionally, your language is patronizing and dismissive, trading substance for insult. This is generally frowned upon in this community. You failed to actually address my comment, both by failing to recognize that it was mainly about using the correct terminology instead of criticizing an entire branch of research that extends far beyond transformers or LLMs, and by failing to establish why a rapidly evolving landscape does not mean that certain generalizations cannot yet be made, unless they are presented with several constraints and caveats, which includes not making temporally-invariant claims about capabilities. I would ask that you reconsider your approach to discourse here, so that we can avoid this thread degenerating into an emotional argument.