This baffles me honestly. ChatGPT, for some reason, does not know of time or the date. It just assumes everything happens at the same time or recognizes ‘day’ or ‘night’ when you tell it. But it never keeps track of anything.
LLMs are not smart. LLMs do not know anything. They are token predictors. They are autocomplete with extra steps. Nothing about what comes from them is tied to the real world, just what text has appeared with what other text most often within a context window.
It wouldn’t be any different if you did the same with a human brain (with the memory capacity of ChatGPT):
“Human, read and memorize this giant pile of written text, which is a mix of things on the Internet, almost every book ever written, and such things.”
“OK, have done that.”
“Now, based on what you’ve read, tell me: Who is Barack Obama?”
“Barack Obama is a man born in 1961 who was the US president from 2009 to 2017. [… other facts about Obama that were somewhere in the training data]”
“OK, and who was Konrad Adenauer?”
“Konrad Adenauer was the first federal Chancellor of the Federal Republic of Germany. [… other facts about Adenauer that were somewhere in the training data]”
“OK, what’s the current date and time?”
“No idea. How am I supposed to know THAT?”
It is not smart.
It never was, and the current technology never can be.
There’s a good reason they say that LLMs are basically fancy autocorrect. They use billions (or trillions for newer frontier models) of parameters and all kinds of neural network voodoo and have been trained on a ton of data to make it sound intelligent. Since it’s kinda like querying a static database (except not deterministic because reasons), there’s no way it would know the date and time. However, they can now use “tools”, which are pieces of software that do one thing or another. And they can “think”, which is just the AI re-prompting it more or less. The LLMs have been trained to output a sequence to call tools when available and applicable. The software running it (in the case of a local LLM, it’s often ollama or the much better llama.cpp) then calls the tool with the parameters specified and returns the data into the LLM’s context. At the end of the day, much of all the fancy agent magic is injecting new data into the context of an LLM.
So now that they have tools, this is Qwen 3.5 9B (9 billion parameters), a Chinese model, running on my ol’ RTX 3060ti.
I tried it twice, both times I got the date and time though the code it used was different (first time did not output date, time and then datetime, it only did datetime).
ChatGPT on the web interface did the same for me, but it’s possible that if you call it via an API or something then it won’t. I don’t pay for it so I can’t test that.
However, this is the critical bit: I made it fairly obvious that I want to know the date and time. If you don’t do this, it won’t have the date and time in context.
Also, I created a new chat and asked it when the next bus from Tallinn to Tartu is. It looked up a bunch of websites, mostly focusing on LuxExpress, and gave me 7 AM as a result. Thing is, LuxExpress isn’t the only bus company in Estonia. Their first bus might leave at 7, but there’s another one at 6 AM. This is a very small model and fancier ones will likely perform better, but I think this is a great example of how these things might make mistakes: They don’t actually think, they just fake it by re-prompting themselves and calling tools (including web searches, which is probably one of the most useful ones for chatbots in particular). An actual human would just go to tpilet.ee or peatus.ee and get the full schedule of all or most bus companies.
Current generative AI does not think. It is a word prediction engine with a huge database of websites and other data to pull from, and and natural language input and output.
Really being able to input and output fairly good language is the most impressive bit to me. But it doesn’t think or understand the words it outputs.
If most people on the internet kept posting Bugs Bunny was the first president, AI would say that too.
For the date and time problem, AI would be looking for keywords from millions of lines of text, and some of those might say something like “The date was September 11, 2001”, but there would be millions of other posts with dates and times. Generative AI can be programmed to recognize that someone is asking for the current time and date, but that is probably on their to-do list of thousands of things to fix and fake.
Despite how the information is presented to you, LLMs aren’t reasoning engines. They don’t come to conclusions based on the information fed to them; they don’t learn. They’re complex probability sieves that try to predict the expected output based on a set of input tokens.
So they have no concept of anything, including time. They look at a prompt, compare it to the prompts they were trained on, and with a bit of random path selection thrown in, try to find the best path to the goal.
My guess is that up until now, this hasn’t required taking time into consideration. It would definitely be possible to create a model that valued time context, but apparently that hasn’t been an important variable to this point in time, at least for frontier models.
It’s not “smart” it’s a “convincing text generator”. It’s very good at being convincing, like making you think it is smart. But it doesn’t know anything, it doesn’t understand anything, it is a parrot repeating what other people on the internet have said in similar situations.
(Current) AI is no more than a text generator.
When does it generate text?
Whenever it is asked to.The more technical explanation: each individual ‘Model’ (LLM) was trained for some time (maybe half a year), and then, afterwards, it works for some time (one or two years). It does not continually amend it’s training while doing it’s work.
A living being would do both at the same time, training and working, but the AI’s cannot.
Not quite right. An existing model can and is tweaked during its lifetime, but these are minor changes to the weights and not worthy of a new version, but they can significantly reduce (or increase) certain outputs.
There are many good answers here, but have you heard of the Chinese Room thought experiment?
https://en.wikipedia.org/wiki/Chinese_room
That’s basically what LLM interfaces are, albiet with a calculator + “fuzzy” dictionary inside. Your current chat is an LLM’s entire view of the world; otherwise it’s like a clone taken out of a vat into a black box, frozen in time forever.
Better LLM interfaces will inject relevant info (like the date) into the context, but ChatGPT is especially poorly made, and no one should be using it (even over alternatives) unless your work makes you do it… Much less pay for it.
If you’re still curious, I’d encourage you to try a “raw” LLM completion API in a unobfuscated interface, like mikupad:
https://github.com/lmg-anon/mikupad
It “pulls the curtain back,” so to speak. And you’ll be left wondering why everyone is acting like these tools are genuinely intelligent.
Because it’s not “smart”. It’s not intelligent at all. It’s not sentient, doesn’t think, it’s not even aware of what it’s generating. Llm stands for Large Language Model and it’s just a text prediction algorithm that decides what words should most likely come after other words. It does this just well enough to present to people who do not understand it as “intelligent”.
They’ve trained on static knowledge at a point in time. Their ability to ‘learn’ and adapt is mostly just fancy tricks with hidden context and the ability to Google.
If AI is so smart, how come it can’t figure out why kids like Cinnamon Toast Crunch. 😤
AI is the very definition of “fake it til you make it”. There’s no actual intelligence there
if llms are so smart why do I still have to go to work to live
Looking at the things llms can’t do is a good way to see their limitations. They only seem they seem to be doing is vague and undefined. code is about the only thing measurable and its considered by many to be crappy at it. Its basically a bit of an improvement to talking to the rubber duck as it combines that with web searching.
