Hacker News Reader

Comment by siliconc0w | Hacker News Reader

There is a lot of noise from VC-types about how AI copilots turn random outsourced devs into 10x engineers. My experience is that it's pretty good at writing boilerplate, that is - it saves me typing what I would expended little thought in typing. I suspect this is where most of the productivity statistics are coming from. Co-pilot especially seems prone to hallucinating APIs or suggesting entire functions that are wrong, repetitive, or even dangerous. This is the case even for side projects where I'm working with public python libraries which should be like the best case scenario. It's generally better than IDE autocomplete when I already am familiar with the API and so can catch the errors.

I will say GPT-4 is mostly better than referencing stack-overflow or library documentation. Maybe once that gets fast/cheap enough to use as co-pilot we'll see some of these mythical productivity gains.

View on HN

I don't think so. I think it will be more like a flare that burns out.

The reason being that the source of material for GPT-4 is rotting underneath. As the internet gets ruined by AI garbage, the quality of output of AI will get worse, not better.

edit: This is a modern day tower of Babel event. We had a magnificent resource and then polluted it with garbage and now it's quickly becoming useless.

This is wrong for at least two reasons:

1) The data scraped from the web is filtered, deduped, ranked, and cleaned in a variety of other ways before being used for training. While quantity is necessary for a model to learn the structure of language, quality is even more important once you have a model that can produce coherent output so a lot of work goes into grooming the data.

2) There have been a bunch of papers and training runs that show synthetic data created specifically for training a model is as good or better than scraped human produced data. The importance of scraped web content is quickly declining because you can now generate infinite higher quality examples using existing trained models. The only relevance it has now is for knowledge about new developments and that is a much easier stream to filter since most of the important stuff comes from official sources and you don't need as many variations since you can just generate your own using one or more examples.

How much time is going to be wasted cleaning the data? It seems like AI can generate junk much faster than we could possibly filter it.

There could be an infinite amount of junk, in fact the amount of junk available online before LLM output was added to the mix was already incredibly huge. At this point it no longer matters because you can take 1 or 2 examples of some knowledge you want to train a new model on and you can then feed it to an LLM with instructions saying to create 10,000 variations on this text while still retaining the factual details, you then filter this list and throw away any output with errors or other garbage and what you're left with are a good number of high quality examples of the information you want to add to your training set. Incredibly, there is new research that shows even this might not be needed at some point because there is evidence that transformers can learn from a single example.

Fortunately there are now AIs that can help with the data cleaning task.

Seems like this is part of the feedback loop that will render AI useless though.

No, it's an effective technique that has been used by Microsoft Research among many others.

But who monitors the monitor?

At a certain point you just need to take a statistically representative sample of the output for review.

you dont need to clean all the junk created. you just need to clean enough to feed the models.

Infinite?

A bit hyperbolic, I'm sure the different variations are finite but you can produce a very large number of examples with essentially the same information content.

A recurring type of problem I've seen goes like: there's an API function f(x) where x can be any of half a dozen different object types. You want to do something similar to f(x) on a different object, which actually requires a completely different multi-step approach, but the LLM will just try to use f(x).

It's really good at spitting out code in common, familiar patterns, but the failure rate is very high for anything a little unusual.

Overtime the products we build might become more and more generic so that they can be easily built by AI?

Reply on news.ycombinator.com