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.
I built a toy news-summarizing agent with Gemma 4, and it was so frustrating, actually, because of the cut-off date.
The model wasted over half the token budget, each time, on internal debates over the current date.
When generating a World Cup summary, for example, it refused to believe qualification rounds were over and refused to even call the web searching tool to collect the data.
I injected the current datetime at the very beginning of the system prompt, but Gemma refused to believe it!
The m-effer insisted the timestamp was fake and hypothesized it was being evaluated in a synthetic lab test with simulated future dates!
No amount of system prompting could convince it to trust the clock.
That was the most frustrating and bizarre "bug" I ever faced!
Gets me with AWS stuff on claude all the time, fortunately there's a official amazon MCP for their docs which helps a lot, but I still have to occasionally tell it to check the docs/mcp.
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.
I find they generally do ok, but a few lines in an AGENTS.md or manual prompting to verify stuff against current docs/source, and check for current version of software helps a lot.
Were you expecting your model to be updated on current events? Why?
Also the specific event you are referring to is a statistically very improbable event, prior to its actually happening.
>It only acquiesced when I specifically directed it to check Reuters.
Do all models do this? They check in with Reuters? Why would a model think that you asking about an extremely improbable event warranted reaching out to Reuters?
If OpenAI is going to call Astra AGI, then I would expect it to be able to update it's weights to new knowledge, because a generally intelligent being can indeed do this.
I can teach myself to play an instrument, and I'm not just building this huge lookup table that I have to access every time I play the instrument. I am updating the weights in my neurons.
I was not expecting model weights to be updated on current events.
It’s clearly warranted because a model that trusts its weights on current events will give an outdated answer. Extremely improbable events happen all the time.
in my chat with gemini it could not differentiate between current events and fiction.
if you point it to the web it got the point, but started treating everything like fiction. so it simply started making up possible scenarios and playing them off as real answers when asked for factual information.
i could not tell what the issue was or how to fix it because the reasoning is encrypted. the obfuscation model spat out something like: 'the user is asking for details about a fictional scenario in which the usa has assassinated the leader of iran'
i really don't like the way big ai companies are going. encrypted thinking, guardrails, adversarial personality, moralizing. it is creating something anti-human.
It can be quite hard to determine what needs a tool call or not. LLMs are not well calibrated to what they know and don't know, and tool calls can add latency and extra costs. There are lots of things that are "obvious" right until they aren't - especially political events and disasters.
For general purpose use this is interesting, but if I'm just using an LLM for coding, does this matter at all? I would hope something like a new java version after a model's publish date can be handled and understood by the model through tool calls and context even if it's not explicitly in the training data, the same way the LLM doesn't have my existing code or the plan to change it baked in from training.
The inference time results are quite different. These models haven’t just been trained with tons of docs, blog entries and videos about Java 8. They’ve also been trained on tons of code examples from simple to complex real world ones.
You’re not going to get all that just by pointing the LLM to the recently released Java 27 documentation. That information is also potentially adding tons of content to your context, which is already filled with tons of other data (your code, other recently released libraries it has to get documentation for, etc).
My experience with Claude is that it doesn’t handle its own training cutoff properly. It responds as though its cutoff is today. Gemini is even worse about this.
I’m going to add this site to my agents files so that they’re explicitly aware of their own limitations.
This is one of the things that bothers me about AI.
To me, intelligence or an intelligent entity should be able to learn from its mistakes and learn new things on its own. Having to start from scratch to teach an AI new facts or new skills is not very intelligent IMO.
Qwen-3.8 for me automatically set copyright footer on a website to 2025 and thought Astro 5.x is the latest version which first came out in December 2024
This is likely being solved with stuff like watermarking which Anthropic just added to Claude... I'd imagine they are testing new data and verifying that training data (or at least the data which do NOT want to be AI generated, i.e actual human text) is not generated, at least with claude.
This is one of the jagged mismatches between users and LLM developers. The median user doesn't care or want to know about static models and knowledge cutoffs and whether a model can do tool calls or if tool calls even happened. They just want something that works.
Fortunately increased capabilities seem to make this a basic expectation with new releases.
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 forgot which was it, ChatGPT or Gemini, but one of them insisted on calling Trump "former president" even when discussing decisions he just announced as president. Lol
The Trump and former president terms were likely firmly stuck together in the embedding space. The model doesn’t validate every single token it produces because validation itself requires tokens. A bloom filter of outdated embeddings will help, when the labs get around to adding it.
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.
I built a toy news-summarizing agent with Gemma 4, and it was so frustrating, actually, because of the cut-off date.
The model wasted over half the token budget, each time, on internal debates over the current date.
When generating a World Cup summary, for example, it refused to believe qualification rounds were over and refused to even call the web searching tool to collect the data.
I injected the current datetime at the very beginning of the system prompt, but Gemma refused to believe it!
The m-effer insisted the timestamp was fake and hypothesized it was being evaluated in a synthetic lab test with simulated future dates!
No amount of system prompting could convince it to trust the clock.
That was the most frustrating and bizarre "bug" I ever faced!
Did you try other models (e.g. Qwen3.6 or others)? I'm curious how others fare.
I've noticed Qwen3.6 struggles a bit with today/date based logic.
Qwen 3.6 did the same thing for me.
Only after some cajolling it finally went to check the history I asked it to (I had been testing KoboldCPP's web search).
Gets me with AWS stuff on claude all the time, fortunately there's a official amazon MCP for their docs which helps a lot, but I still have to occasionally tell it to check the docs/mcp.
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.
I find they generally do ok, but a few lines in an AGENTS.md or manual prompting to verify stuff against current docs/source, and check for current version of software helps a lot.
ChatGPT once told me I was the target of a sophisticated nation state misinformation campaign when I linked it a Reuters article
There are a lot of people who would say that is exactly what Reuters is. Any American news outlet, really.
It was a real world event. The guy is still dead unfortunately. If Sputnik or Pravda are more your style they reported on the same event.
>the model said that’s pure fiction.
Were you expecting your model to be updated on current events? Why?
Also the specific event you are referring to is a statistically very improbable event, prior to its actually happening.
>It only acquiesced when I specifically directed it to check Reuters.
Do all models do this? They check in with Reuters? Why would a model think that you asking about an extremely improbable event warranted reaching out to Reuters?
If OpenAI is going to call Astra AGI, then I would expect it to be able to update it's weights to new knowledge, because a generally intelligent being can indeed do this.
I can teach myself to play an instrument, and I'm not just building this huge lookup table that I have to access every time I play the instrument. I am updating the weights in my neurons.
Until AI can do this it's not AGI in my book.
I was not expecting model weights to be updated on current events.
It’s clearly warranted because a model that trusts its weights on current events will give an outdated answer. Extremely improbable events happen all the time.
I think the models are trying to optimistically avoid doing web searches, because they're surprisingly a lot harder to do well than you'd think.
How so?
He asked it to double check. It's reasonable to expect the LLM to handle that trivial task.
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.
in my chat with gemini it could not differentiate between current events and fiction.
if you point it to the web it got the point, but started treating everything like fiction. so it simply started making up possible scenarios and playing them off as real answers when asked for factual information.
i could not tell what the issue was or how to fix it because the reasoning is encrypted. the obfuscation model spat out something like: 'the user is asking for details about a fictional scenario in which the usa has assassinated the leader of iran'
i really don't like the way big ai companies are going. encrypted thinking, guardrails, adversarial personality, moralizing. it is creating something anti-human.
It can be quite hard to determine what needs a tool call or not. LLMs are not well calibrated to what they know and don't know, and tool calls can add latency and extra costs. There are lots of things that are "obvious" right until they aren't - especially political events and disasters.
all the reasoning still comes from pretraining data
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
Says who? Models can also use results from tool calls in their reasoning loops.
For general purpose use this is interesting, but if I'm just using an LLM for coding, does this matter at all? I would hope something like a new java version after a model's publish date can be handled and understood by the model through tool calls and context even if it's not explicitly in the training data, the same way the LLM doesn't have my existing code or the plan to change it baked in from training.
The inference time results are quite different. These models haven’t just been trained with tons of docs, blog entries and videos about Java 8. They’ve also been trained on tons of code examples from simple to complex real world ones.
You’re not going to get all that just by pointing the LLM to the recently released Java 27 documentation. That information is also potentially adding tons of content to your context, which is already filled with tons of other data (your code, other recently released libraries it has to get documentation for, etc).
Gotcha, that makes a lot of sense. Thank you
My experience with Claude is that it doesn’t handle its own training cutoff properly. It responds as though its cutoff is today. Gemini is even worse about this.
I’m going to add this site to my agents files so that they’re explicitly aware of their own limitations.
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.
This is one of the things that bothers me about AI.
To me, intelligence or an intelligent entity should be able to learn from its mistakes and learn new things on its own. Having to start from scratch to teach an AI new facts or new skills is not very intelligent IMO.
few more generations they will invent "patch in place"
(with "AI" developing the technique of course)
Qwen-3.8 for me automatically set copyright footer on a website to 2025 and thought Astro 5.x is the latest version which first came out in December 2024
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 likely being solved with stuff like watermarking which Anthropic just added to Claude... I'd imagine they are testing new data and verifying that training data (or at least the data which do NOT want to be AI generated, i.e actual human text) is not generated, at least with claude.
This is one of the problems that eventually solves itself, somehow
An aside - isn't it nice that have been fewer model releases the past week or so? A brief respite!
I remember running the docker container for ollama and its knowledge cutoff is somewhere in 2023 still. That's unacceptable.
This is one of the jagged mismatches between users and LLM developers. The median user doesn't care or want to know about static models and knowledge cutoffs and whether a model can do tool calls or if tool calls even happened. They just want something that works.
Fortunately increased capabilities seem to make this a basic expectation with new releases.
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?
I forgot which was it, ChatGPT or Gemini, but one of them insisted on calling Trump "former president" even when discussing decisions he just announced as president. Lol
The Trump and former president terms were likely firmly stuck together in the embedding space. The model doesn’t validate every single token it produces because validation itself requires tokens. A bloom filter of outdated embeddings will help, when the labs get around to adding it.
Depending on the use case certain models
it is depending, i think.
Depending on the use case certain models very well remain as or more reliable for certain tasks.