Hacker News Reader

Comment by it_citizen | Hacker News Reader

> How much can AI help?

  In 2019, this is sci-fi
  In 2023, it can help engineers understand the complexity of taxes, translate some tax rules into code, help refine edge cases and fix mistakes.
  In 2030, ???

Things move insanely fast lately. I would not bet on anything.

View on HN

It honestly seems sort of like a straight path toward an LLM that can hold a whole codebase in context and can produce new functionality on command. Fundamentally, software is just taking inputs and producing outputs; we will almost certainly be able to define those and let the LLM do everything in between within a year or two.

And brain surgery is just cutting someone's head open with a scalpel. Programming languages are stupidly information dense. I'm highly skeptical an AI will be useful for non-trivial boiler plate any time soon, they lack the reasoning powers to get the details right even occasionally for common tasks and basically never for uncommon ones.

While I take your meaning here, I would suggest also that natural languages are even more stupidly information-dense. And while the rules of their logic are not as hard and immutable as the semantics of programming languages, they do have plenty.

I don't agree that natural language is more information dense than most programming languages, there's a bunch of shared context, work, and assumptions that must be understood before effective use of both, but natural language is typically pretty loose and relies heavily on context for real meaning, which is still subjective. Conversely, PLs have no subjectivity once executed/compiled. Natural language is very verbose when trying to express the concepts of a PL with the same specificity, just look at how long entry-level tutorials for writing hello world in any PL are. Or try explaining all the details of even a basic function call in C to a CS101 class. You need to say a lot of words to describe the same exact thing.

PLs benefit from being targeted at a specific domain dramatically that NL just can't.

Programming languages are stupidly information dense indeed, but if that's the most major roadblock, I don't see why throwing more compute at the problem won't eventually solve it. GPT-4 is already handling a very stupid amount of information density.

My point isn't the information density, it's the fact that every detail matters. LLMs aren't great at details or facts and therefore aren't really suited for writing software right now. It's not a computational limit, it's intrinsic to how LLMs work.

Consistency is still a huge problem. I tried to build a D&D character with Chatgpt the other day. Absolute nightmare. For example, it said the character gets a racial bonus to two attributes. I told it to give me my attributes taking into account racial bonuses. It applied the bonus to one attribute but not the other.

Today I got the first case of something I could not find in Google, that chatGPT found.

It was more an indexing question and a factor of me remembering the concept of a theater play, but nothing else.

I was really impressed when the LLM found the answer right away. Until I realized it was a lie.

I had to made it print a list of 50 plays, and I found what I was looking for.

I realize it’s not a good usage for a LLM but I think it’s how it’s gonna be with those

Something a lot of people aren't considering is the added power the LLM will have once we are building systems from the ground up specifically for the AI to use. Right now it's doing a decent job with languages and tools that are built for human usability. I imagine soon we'll have languages, systems, and tools specifically built to be extremely usable by AI.

I seen that said a few time and I struggle to see why a language would be more AI friendly than à other ?

> I imagine soon we'll have languages, systems, and tools specifically built to be extremely usable by AI.

Those are called programming languages, we have quite a lot of them.

Maybe at some stage software will also have limits to, I know that is hard to imagine but maybe why software has stagnated a little is because we’re running out of actual things it can do for us from a societal and economic standpoint?

I don't think your second list item here is true, or uniquely true of GPT/LLMs anyway, so I'm not sure we're any further along than we were in 2019 for this specific problem.

What part you don't think is true?

Help engineers understand taxes: it is pretty much part of the GPT-4 demo. You can copy the tax code in it and ask questions about it. As a matter of facts, it helped me with some taxes questions I had this year, although it was ChatGPT and I didn't paste the tax document in it.

Translate some tax rules into code? I did something similar last week for a complex billing project I am working on. It didn't give me the best variable names and made up a few some functions but the formulas were correct and I reused a good chunk of it.

Fixing the logic of the code it just wrote? I think there are enough examples of that on Twitter.

> You can copy the tax code in it and ask questions about it.

Doesn't it still have problems with outright hallucinating a bunch of crap?

If so, the dangers of using something flaky/unreliable for potentially life changing decisions...

I use it a fairly amount and it does indeed hallucinate things here and there. I know not to take things at face value.

But whenever I have to deal with a complex question such as taxes, I often find it very useful to break down a problem, rephrase and provide context around my question and offer different leads to follow or cross reference.

I see it as GPT spiting out a skeleton of an answer onto which I can attach some meat. It is very useful when I don't really know the domain and therefore how to attack a problem.

What part of what you just wrote couldn’t we do before, with human labor and/or existing tools?

People seem to assume tech progress is always some linear or exponential thing where things just keep getting better.

They can’t imagine we could also just hit a plateau and have no progress for a decade until some other major breakthrough happens.

AI art is hitting a plateau or has hit a plateau. Progress came quickly for months but nowadays each update is just a bit better details with more resolution and better hands. That’s it.

If you have AI anxiety, just read my comment history.

Is there a name for this fallacy? Because I see it a lot.

Maybe whatever you call “past performance doesn’t imply future results” or something.

(Roughly) Amara's Law: “We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run.”

False uniqueness effect? https://en.m.wikipedia.org/wiki/False-uniqueness_effect

Is there a name for the technology advancement fallacy? I see all the time that people think with enough time technology will simply make every problem surmountable. It doesn't matter if it's limited by physics, conservation of energy, or reality

I am not saying that some linear or exponential curve is certain but so far I don't see any sign of slowing down. The research papers seem actually full of new avenues and very promising low hanging fruits.

> AI art is hitting a plateau or has hit a plateau.

I am not sure I am following. Are you saying AI art has hit a plateau since stable diffusion and dall-e 2 got released? This is less than a year ago. Look for plateau at the scale of a 5 to 10 years, not 6 months.

In 2040 we are its support staff. By the end of the century our great-grandkids are basically its pets.

And I'm not even sure any of that is a bad thing. How do we deal with the idea that the AI industry might create a benevolent God?

Reply on news.ycombinator.com