Item — HN Reader

And that will kill AI itself, since much of what it knows is from what learned from StackExchange before this latest one demise.

And before the obvious comments on how GenAI is creative, then please do this OpenAI and Anthropic, for your next LLM. Just teach it Python, C and Rust and give it some good books. But dont give it access to Github...lets see what you can do then...

3 comments

LLM training will eventually transition from real data to synthetic data, same as alphago -> alphazero.

AI companies are also working to integrate training with real-world experience through sensors and robotics, to shrink the gap between human experience and hallucinated LLM experience.

They all have archives of pre-LLM content. There's also archive.org, google books, and pirate ebook archives. I don't know what they're doing to build video and audio archives, but judging from the cost of spinning rust, they're storing significant quantities of that, too.

Some parts of the internet are curated, and even with LLM influence they're still worth training on. I doubt wikipedia or stackexchange or rosettacode will ever cease to be useful at all.

Neither AlphaGo nor AlphaZero were transformers. Why would you expect the same results?

Going further, current LLMs have at least an order of magnitude more computing resources, but completely suck at go. Why would this suddenly change unless we dumped the countless games played by alphago for them to train on?

That (theoretically) solves training, but it doesn’t change the fact that even smart models can’t extract useful information from a dead internet, so you’ll always be stuck with a stale training cutoff. This is already a problem I run into a lot. I search something first. Top results are slop sites, so I switch to a chatbot. Its answers look suspiciously similar to the slop sites I just noped out of. Check the sources. It’s them.

And the training of future models will have to contend not only with slop, but also huge amounts of content specifically designed to "taint" future training data. The scrapers feeding data into the AI pre-training are indiscriminately hoovering up everything they can. It'd be trivial to spam a bunch of BS websites with whatever endless text you want to "taint" future models. Post tons of examples of insecure code, publish package.json files pointing to some malicious library, etc...

I don't see how AI needs StackOverflow anymore.

It can either examine the ground truth source code to answer your question "How to expire cookies using RoR Devise gem" or it can read docs for you or it can spin up local experiments to black box examine some software. If humans had done that before posting on StackOverflow, the question never would have made it there.

Its reasoning ability is long passed hoping an example exists online for it to copy.

The end goal though (in my understanding) has never been for an LLM to regurgitate knowledge it ingested during its training. The end goal is to use the patterns and correlations found in internet data to generate an emergent prediction and problem-solving machine.

Whether that's possible is something we'll discover, but no one is throwing billions on AI companies for the hope of them building a giant natural language queryable internet information repository.