I remember when the US captured Venezuelan president Maduro, and when I posed a prompt related to this, the model said that’s pure fiction. I told it to double check. Still didn’t want to entertain the idea. It only acquiesced when I specifically directed it to check Reuters. I haven’t noticed this problem in months. Model cutoff seems to be less of a problem these days.
It's a "problem" of compute, I think. If you query without an account on ChatGPT you will see the model look up less stuff and research less, than when you have a paid account and choose "medium" or "high" in the effort slider.
Which makes sense, because of you have looked into search and crawlers you notice that search is actual quite expensive (which is why e.g. Kagi charges a few bucks for search every month).
It's not strictly compute, because this has noticeably improved in open-weight models too, such as Gemma and Qwen. I suspect they noticed this issue and adjusted their training to be better about it over time.
Came here to say the same thing. Models used to rely heavily on world knowledge from their training data. They are now much better at tool use and deciding when to research a topic, rather than just answering from memory.
I wonder how much that extends to using LLMs for programming. I assume most knowledge of programming language syntax still comes from training data.
After Trump's last inauguration, ChatGPT would still tell me that Biden was President of the US. I understand that the training cutoff was before Biden dropped out. But it knew, or should have known, the current date and that there had been an election since its last update, but it didn't qualify the answer. When I asked it to search the web, it got it right. The moral I took away was to always ask for the search whenever I ask about current events. I do that so routinely that I wouldn't know if this problem has been fixed. I suppose that failing to update my priors per individual model release is a form of bigotry against a widely hated class.
ChatGPT recently started web searching for for basically every general knowledge question, which I found quite odd. Maybe an overcorrection to the issue you were having?
I remember when the US captured Venezuelan president Maduro, and when I posed a prompt related to this, the model said that’s pure fiction. I told it to double check. Still didn’t want to entertain the idea. It only acquiesced when I specifically directed it to check Reuters. I haven’t noticed this problem in months. Model cutoff seems to be less of a problem these days.
It's a "problem" of compute, I think. If you query without an account on ChatGPT you will see the model look up less stuff and research less, than when you have a paid account and choose "medium" or "high" in the effort slider.
Which makes sense, because of you have looked into search and crawlers you notice that search is actual quite expensive (which is why e.g. Kagi charges a few bucks for search every month).
It's not strictly compute, because this has noticeably improved in open-weight models too, such as Gemma and Qwen. I suspect they noticed this issue and adjusted their training to be better about it over time.
Came here to say the same thing. Models used to rely heavily on world knowledge from their training data. They are now much better at tool use and deciding when to research a topic, rather than just answering from memory.
I wonder how much that extends to using LLMs for programming. I assume most knowledge of programming language syntax still comes from training data.
It still matters, but in the age of good reasoning, tool use, and web search, this is much less of a problem than it used to be.
all the reasoning still comes from pretraining data
Says who? Models can also use results from tool calls in their reasoning loops.
some logic here
https://www.lesswrong.com/posts/HYCGA2p4bBG68Yufh/thinking-a...
I think they mean reasoning their way to the need for a web search.
Or other kind of search / knowledge acquisition / computer use etc to get the information needed
Do people prefer the new flat style LLMs are producing? I don’t mind it as much as the gradient theme they were pumping out previously.
Pre-AI internet data is like pre-war steel
The slop would multiply if we keep feeding it to new models in a loop
This is one of the problems that eventually solves itself, somehow
I remember running the docker container for ollama and its knowledge cutoff is somewhere in 2023 still. That's unacceptable.
ollama is just an inference engine - it just runs models.
it must ship with some default old model if you didn't need to explicitly download one
After Trump's last inauguration, ChatGPT would still tell me that Biden was President of the US. I understand that the training cutoff was before Biden dropped out. But it knew, or should have known, the current date and that there had been an election since its last update, but it didn't qualify the answer. When I asked it to search the web, it got it right. The moral I took away was to always ask for the search whenever I ask about current events. I do that so routinely that I wouldn't know if this problem has been fixed. I suppose that failing to update my priors per individual model release is a form of bigotry against a widely hated class.
ChatGPT recently started web searching for for basically every general knowledge question, which I found quite odd. Maybe an overcorrection to the issue you were having?
Depending on the use case certain models very well remain as or more reliable for certain tasks.