Hacker News Reader

Comment by YeGoblynQueenne | Hacker News Reader

>> GPT-Migrate is currently in development alpha and is not yet ready for production use. For instance, on the relatively simple benchmarks, it gets through "easy" languages like python or javascript without a hitch ~50% of the time, and cannot get through more complex languages like C++ or Rust without some human assistance.

In other words, it doesn't really work.

The current wave of LLM applications still seems to me like someone just invented homeopathy and a whole bunch of people are convinced it's real and are trying to use it to create a cure for cancer. It's just people waving their hands about and intoning magick formulae, that don't work and don't produce anything useful at all.

I am curious to see where all this is going to end up. Is someone going figure out a way to make LLMs work for real-world er work? Are we all waiting patiently the next big LLM version to see if it can do the things that the current best-of-the-best can't?

View on HN

I don’t understand why this talking point keeps surfacing. LLMs are in significant usage already in many production environments and are already improving peoples’ lives and productivity. I can have a dig for more examples but I’d urge you to do a bit of research and try out the gpt-4 API for yourself. Or make some more qualified statements with specific bugbears. Equating it all to homeopathy is so confusing.

- Image descriptions eg. Be My Eyes: https://www.bemyeyes.com/

- Summarisation and content distillation, question answering Etc: literally everywhere. This is now a solved problem (to the point of it being dull) thanks to LLMs.

- Customer service chat triage.

- helping students learn- eg khan academy, tutor Lily

- Helping create and debug software. If you don’t think is happening then you’re either living under a rock or just close to believe people are lying about how they’ve used it?

I'd push back on summarization being solved. I have extensively used GPT 3.5 and 4 for this in production, and it works very well but it still often just ignores critical aspects of the original text, and it will rarely even completely hallucinate details. And this is with a lot of iteration on the prompt.

We still have to throw out > 50% of its output because a human can summarize the text much better.

> - Summarisation and content distillation, question answering Etc: literally everywhere. This is now a solved problem (to the point of it being dull) thanks to LLMs.

Calling this problem domain "solved" only undermines whatever point you're trying to make, unless you really believe that current LLMs which fabricate information, flip flop between answers with follow up questions, and even gaslight the user, can be called "solutions".

LLMs have proven themselves to be such untrustworthy "sources" of information and knowledge that I'm struggling to understand why anyone would even try to make this particular claim. It's trivial to refute by anyone who's used them and has been thoroughly refuted by StackOverflow Developer survey which reported that only 2.8% of developers "highly trust" the output of AI tools.

https://survey.stackoverflow.co/2023/#section-developer-tool...

Summarization, language translation and tasks in that manner (morphing tasks?) get much lower hallucinations.

Not sure what that survey has to do with anything.

> Summarization, language translation and tasks in that manner (morphing tasks?) get much lower hallucinations.

Source?

There are no benchmarks that test that. It's my experience and the experience of others who use LLMs for such tasks.

Down below there's a comment pushing back on saying Summarization has been solved. Even he/she is saying hallucination is rare.

> and are already improving peoples’ lives and productivity.

Don't forget this is coming at the cost of jobs that the AI is replacing. I would say, for those people, it is hurting, if not destroying, their lives.

You would all do well to remember this.

So does all forms of automation. By your logic we should've never invented computers because they took jobs from human computers[0]. Improved technology and automation reduces jobs for people doing things the old way and makes the job easier for everyone else. No job is necessary forever, that's just how the world works.

[0]: https://en.m.wikipedia.org/wiki/Computer_(occupation)

And we've always had a shitty answer to dealing with the fallout.

Unskilled people are running out of things they can do or retrain to. Sending everyone to college has only destroyed the meaning of a college degree and created a surplus of postgrads who can't earn anywhere near their potential for lack of demand.

If we keep overfishing this lake, at some point (I am guessing in the near future) we will hit an inflection point where the number of people vastly outnumbered the availability of jobs, and social services will be strained to the point of collapse.

That is when the torches and pitchforks will come out, and these folks had better hope to whatever God they think exists that AI will save them. Because my bet is it won't.

What are YOU going to be doing to help these people? I advocate for UBI wherever feasible, I simply don't see a way to put these folks back to work without laying waste to whatever industry they decide to enter.

>> - Image descriptions eg. Be My Eyes: https://www.bemyeyes.com/

Not working yet. From the website:

Starting today, you can register for the waitlist in the Be My Eyes app.

>> - Customer service chat triage.

More specifically? Which company is currently using LLMs for this purpose?

>> - helping students learn- eg khan academy, tutor Lily

Helping them learn, but what? Per wikipedia:

Statements made in certain mathematics and physics videos have been questioned for their technical accuracy.[43]

Sounds like using LLMs will just generate more garbage teaching material.

>> - Helping create and debug software. If you don’t think is happening then you’re either living under a rock or just close to believe people are lying about how they’ve used it?

Oh, absolutely. People are absolutely lying to themselves. OpenAI's and DeepMind's systematic testing of their code-geneator LLMs make it very clear that those systems produce incorrect code the majority of the time. The best results reported are 28.8% for Codex (in perpetual preprint: https://arxiv.org/abs/2107.03374) and 29.6% for AlphaCode (preprint: https://arxiv.org/abs/2203.07814 Science, paywalled: https://www.science.org/doi/10.1126/science.abq1158) and the latter is with their special 10@k metric which basically means the LLM gets 10 guesses.

I have definitely observed people convince themselves online that CoPilot or ChatGPT even is helping them "improve their productivity" or some such. It is obvious that they are fooling themselves, badly. If you push them, they immediately say "oh yeah, it makes mistakes, but I can correct them" etc. So they just feel like it's useful, even when it's just making them do more work.

In fact, that's exactly like homeopathic self-delusions: people use it because it makes them feel better, not because it has any measurable benefit.

Btw:

>> ... you’re either living under a rock ...

Stop being a jerk.

>I have definitely observed people convince themselves online that CoPilot or ChatGPT even is helping them "improve their productivity" or some such. It is obvious that they are fooling themselves, badly. If you push them, they immediately say "oh yeah, it makes mistakes, but I can correct them" etc. So they just feel like it's useful, even when it's just making them do more work.

That mistakes can be made doesn't mean time isn't being saved. Everybody makes mistakes as is and first try code isn't typically being pushed human or not. The presence of mistakes means nothing.

Frankly, you're the one who comes off delusional here. "Most people are telling me one thing but they can't be right because i believe so so they must be lying to themselves" isn't normal behavior, especially when your biggest argument that they must all be wrong and you right is that the machine makes mistakes. That's very weak.

Take a step back and really look at what you're saying.

Dude I’m literally—right now— using an LLM to fuel and *build* new features in my production website (ablf.io). I’m just some random eng. I’m sure there are many like me. And I’m not talking about gpt just recommending the odd function. It builds entire modules and writes comprehensive test suites. I’d rate its competency as similar to a junior dev. I guess I’m lying to right? It helps me learn about and write ML and NLP stuff I’m very new to. It’s basically replaced stack overflow for me.

"Dude", there is no doubt that LLMs can generate code, and there is also no doubt that they can't generate correct code reliably and in fact they are extremely inaccurate. We know this because OpenAI, DeepMind, Salesforce (I kid you not) and others have tested their systems and though the measures of performance they use are arbitrary and designed to make them look better than they are (like the n@k measure which just lets the LLM guess any number of times) they still score very badly. See the results I quote in the above comment.

So I have no doubt you are literally, right now, using some LLM to do stuff, I just have no doubt that it is not doing what you think it does.

You say you're an engineer? I know that means you write code, but the first thing that's drilled into engineers in training and in work is that you don't just make a thing and call it a day, you make sure to understand the properties of the thing you built and what it can do, and what it can't. Like you don't just put some planks on stilts and say "here's a bridge, come and drive your cars over it". You sit down and do the maths and decide what loads the bridge can take (and you optimally do this before building the bridge). So have you done anything like that? Do you have any way whatsoever to tell how often your system works and how often it shits itself?

I came here looking for a rust->forth pipleline, are you telling me that isn't available? /s

I've long believed that interlingual glue should occur within llvm at the IR level. Reversing each layer of a compiler feels plausible if complex.

A prototype that works 50% of the time is not homeopathy. A new cure for some types of cancer that would succeed 50% of time would be an achievement.

Unless it straight up kills the other 50%.

Than it's certainly not homeopathy. It doesn't kill anyone.

The readme says "~50%" of the time, which seems to be a number pulled out of a hat (to be nice) rather than any serious attempt at quantifying performance (there are no metrics of any kind and the authors are asking for relevant benchmarks). It's more like the authors' feeling, than anything they have ascertained in any reliable way. That's pretty much the same way homeopaths test their "remedies" (i.e. water).

Reply on news.ycombinator.com