Hacker News Reader

Comment by in_absentia | Hacker News Reader

So OpenAI is not a believer in recursive self-improvement after all?

Read the article, that's not what this is about. Title is clickbait, they fired contractors for using AI to label data when the whole point of labeling data is to distill human knowledge into weights, not distill weights into weights.

You're replying to a joke, and yes, that's all it's about. For all the tales of rapid self-improvement, all labs are real careful not to taint their precious training sets with anything that comes out of their systems. Website reputation modeling, fingerprints in generated text, firing contractors that rely on AI.

I've been told on HN about a year ago that the era of scraping is over and that it's all AI-training-AI now. My web server logs and stories like that disagree.

There's a difference between shoving LLM output back into the input and RSI more broadly. Using synthetic training data didn't work super well, that would have been one path to RSI, but it doesn't work. The main path towards RSI people are talking about today is using the models to do research and experimentation for training new models. Machine learning is a very empirical field, you need to run tons of experiments, tune hyperparameters, and try different architectures. It's something agents are extremely good at doing, just because one path to RSI doesn't work doesn't mean it won't work at all.

I wonder if that's why AI sounds so wooden - so that they can clearly distinguish AI generated text from manmade.

I've been told on HN that 1+1=3, so, yes, skepticism is required on broad claims.

There are more subtle truths on this. Things that can be proven algorithmically are much more apt to be in recursive AI loops now. Hence things like programming and hacking keep improving steadily over time with much less human training data being added.

That is exactly what "training AI" needs: labeled data.

What is the other way to interpret the title in your mind?

3rd party contractor hired to generate human labels uses AI generate labels, gets fired. "Training" is a lot different than "data labeling".

Why is it a lot different? Is the labeling less important or something?

Think of labeling as establishing a ground truth. It's more important than training, but it's far less complex than training at an individual level. A kid can tell you what an ice cream cone looks like, but they cannot tell you the best algorithm to use to get the best model with the least power usage.

Labeling is more like working a checkout at Walmart. Just about anyone can do it with the smallest amount of training, but you have to ensure your labelers are not just scanning one item multiple times and bagging up the rest as your dataset can skew from reality since AI cannot just capture this data fully reliably at this point (well in many fields it can or can do even better than humans, but it's still lumpy as to where and why).

Sounds a whole lot like "training" even if it isn't training. I still wonder why gp is so angry..

But isn't the point of recurive self-improvement that you can fire people training your AI?

Correct, hence we don't have RSI yet. RSI'ing before you have RSI doesn't make a workable RSI.

They’re a believer, but they won’t be paying third parties for that once it becomes feasible

I think people want the service they are paying for.

Reply on news.ycombinator.com