Photo of DeepakNess DeepakNess

AI collapses

Unproofread notes

I had an interesting exchange of thoughts about AI or rather the future of AI with Ralf Christian on X. He made some great points that I thought should collect here:

I think the main problem is the tech itself. It doesn't 'know' anything, it 'simply' spits out content based on probabilities in the training data.

What's good in the training data is the spoken language. That's why it 'speaks' so well. But the training data is full of wrong info, that's why there's wrong output even after reasoning.

If people publish less human written content and more AI generated content, and we don't have a way to identify that with 100% accuracy, this will definitely not make those models better in the future

You might be able to still improve it here and there, like that it better keeps context, but don't expect any leap here. That's why there are no big improvements since they released chatgpt 3

I think the future if this might be niche LLMs, where you train them on a specific topic with almost hand picked training data and fine tune it for your specific use. For example, if you're Microsoft you could train it with all your company's code. I guess this gives output more close to what you want than training it with all of githubs code randomly

ChatGPT is really impressive, but it's far from making a real difference in real business (unless you are into spam 😇)

Yesterday I tried to generate a video with Sora. It failed so hard. I think what you are seeing on social media is 1000 people trying to do a video, 999 generating crap and not posting it and 1 got lucky and posts it. That's not value, that's luck.

I loved the simple explanation he made. Also, I loved this paper on "AI models collapse when trained on recursively generated data" that Ralf shared earlier in the same thread.

Comment via email