Technology

Source: Work in AI, sometimes on LLM's, mostly on the software engineering side rather than the science side.

I have a few theories on why ChatGPT was so successful, and why the hype is starting to crumble, but they all largely centre around a well-known problem that LLM's have had for years - they hallucinate, a lot.

When your product becomes popular, you deal with unique problems that don't seem to scale without insane amounts of money, or a literal army of people to plug gaps where your model is saying things it shouldn't - whether it's accusing high-level politician's of crimes they didn't commit, or telling people how to make chemical weapons in the style of your grandma. It's an expensive loop in compositional models, so I'd hate to know how much work it took ChatGPT to get to it's "best" version.

Over time, valuable data disappears, and your data naturally skews over time due to it being incorrect, invalid, or pushed into just being biased towards a given inaccuracy. Sometimes, you do everything right and you train on manual input that you've vetted as correct through expert analysis or user feedback - and it's still wrong. IMO, ChatGPT was always going to struggle to keep the hype, and it will eventually be seen as what it has always been: a concept that shows the utility of LLM's as a commercial product.

Make no mistake, the likes of Google, Amazon, Apple, and Meta will probably plug the gap, and will reach either parity with GPT4 or improve on it. However, the fundamental problem of hallucinations will not disappear, and we'll continue to see neutered experiences that make for great tools, but burn cash to provide these tools with minimal possibility of offending people/damaging the brand.

The main thing I hope to see from the rise and fall from ChatGPT is a rise in productivity tooling, but also people to finally see those that hype these technologies as what they are - grifters.