As if they have fine controls over these models... Anyone who's been playing with language models have seen this false confidence problem, and the way they're "solving" it is by training larger models.

They don't need to enshittify their lower models when they're already bad. Not everything that doesn't work as you expect is enshittification. If we could stop overusing this word and focus it on the actual things that deteriorate over time, that'd be great.

Free access are probably the sub billion size, or very limited in CPU which amounts to the same.

And yes why give people the fancy model for free, they can use some random model from openai

Just because it's not human doesn't mean I can't show some basic manners. 😅

It seems more expressive than GPT-4 in general and better at itemizing information, but it's a bit worse at keeping track of context and seems to hallucinate more than GPT-4 (but less than GPT-3). I'm currently on the free trial period so I'll try it out for these two months and then see if it justifies switching over (the subscription is the same price so that won't be the deal-breaker). One thing that it's noticeably better at is researching online, probably because it uses Google as a backend instead of Bing lol.

It's supposed to integrate with other services, if you use the Android/iOS app, as it can replace Google Assistant entirely and get access to info from other Google services. However, the app is unavailable in the EU and a few other countries as of yet, along with image generation. Classic case of EU favoritism towards Microsoft and not sanctioning it like it does Apple and Google... It's also rumored that the ChatGPT Android app will get Assistant API integration "soon", but it's also not yet available (anywhere, AFAIK). Presumably, even if that is the case, Google can probably add in deeper system-level integrations, given how it's their OS and their ecossystem.

That explains so much.