to be fair, i can barely fucking believe it myself... this fucking timeline

So why don't they update their knowledge more often? Because that's actually harder than the above learning approach. By a lot. It can be done, but there is currently no real incentives to do the doing, when the actually paying cusotmers already have other ways to achieve it (as mentioned above)

in particular - imagine if every chatgpt request also did a bunch of google, a bucnh of wikipedia. this would be.... not a popular approach. grok-4 can do it, but it can do it at _very limited quota_ for exactly taht reason - web search is MORE EXPENSIVE than LLM tokens, and it's not even close.

b) so-called "tools" for research - this is actually how we approach it at work, each LLM has access to a deep research tool that lets it verify facts from places like wikis and other trusted sources

So why does ChatGPT not do this out of the box? Becuase it's not controlled, and that's liability.

You can augment this in a couple of ways:
a) memories - you can actually, with modern LLMs, add knowledge on top that is "your" knowledge. There are a bunch of ways to achieve this - either more pre-trained (RAG style), or "remembering" - but that's _per user_.

(more)