And now I'm picturing Harrison Ford going around asking people what time it is.

As usual, there are always caveats. Many youths these days can't tell time from an analog clock, or perhaps, they actually are replicants. Who is to say since we are in the age were many people think correlation is "factual" proof.

I find the various LLMs are helpful as a form of search engine, enabling me to drill down to potentially useful information more quickly. However at the same time they are far worse than a search engine because they aren't able to actually give you the sources to check. When ChatGPT generates a chunk of code, if you ask it where it got it from, it will say it didn't get it from a specific site, it just knows this stuff. Which of course ends up wrong half the time. So you end up with wrong stuff confident

So, basically if you have a nail, a hammer is still better than a screwdriver to push it in the wall.

That is to say, there are some things that LLM don't do well because, basically, they are just that: large language models. Returning information about curernt, timely events is not in their nature.

Not really. If you ask it what's going on in the news then you'll get an up-to-date response, because it knows in that instance to check and summarize the news rather than just generating something from its LLM.

And if you ask Gemini what time it is you'll get the right answer, for the same reason.

The fact that ChatGPT fails to do this is a problem with ChatGPT, not any inherent problem for AI. Probably in response to this embarrassing article it will be fixed within a couple weeks.

Any modern AI model can be provided "tools" that it can use to perform various tasks or retrieve various information. The current date and time is easy to do. I can't say why the author and/or ChatGPT seems to have trouble but you can easily set up a tool to return the current date and time, instruct the AI "this will return the current date and time" and then if the user asks for it the AI will automatically leverage the tool. It's possible ChatGPT just has a lot of tools at its disposal and is getting con

We aren't asking it to guess based on how tired it is. Just check its own system clock.

What do humans have to do with super human intelligences? Are you saying that if humans are bad at telling the time, then super human intelligences should be excused for being worse than humans at that task? Did you miss the bit about "super"? It's from Latin, it means "above", "over", "beyond", "better".

> automated image pattern matching has been around for decades

The problem is that the LLM only does one trick. When you start integrating other software with it, the other software's input has to be fed in the same way as your other tokens. As the last paragraph of TFS says, "every clock check consumes space in the model's context window" and that's because it's just more data being fed in. But the model doesn't actually ever know what time it is, even for a second; the current time is just mixed into the stew and cooked with everything else. It doesn't have a concept of the current time because it doesn't have a concept of anything.

You could have a traditional system interpreting the time, and checking the LLM's output to determine whether what it said made sense. But now that system has to be complicated enough to determine that, and since the LLM is capable of so much complexity of output it can never really be reliable either. You can check the LLM with another LLM, and that's better than not checking its output at all, but the output checking is subject to the same kinds of failures as the initial processing.

So yeah, we can do that, but it won't eliminate the [class of] problem.

For the record, maybe we shouldn’t call it a bug because they have rudimentary sentience and thinking ability. Maybe unproductive mutation would be more apt.

> Because AI cannot do anything it hasn't been programmed to do.

This is comically incorrect.

Yes it will keep being LLMs are LLMs week until the outside world moves on from hammering "LLMs are magic" week every week.

LLMs are of course LLMs.

> Same if you ask it about the next full moon - or really anything temporal subject about the present or future where there's a lot of training data from the past - as is to be expected.

I just asked Chat GPT "When is the next full moon?" and it did a web search to give me the correct answer for my location.

> The ChatGPT (or any AI product) people actually use is more than just an LLM. People get wound up over the inherent limitations of LLM's as if that's some brick wall preventing these things from ever becoming useful.

> It's ludicrously simple to program a chatbot script to return the time instead of diving into the LLM if the user asks for the time. Same with mathematical operations - yes, LLM's sometimes get basic math wrong because that's not what they're designed for. But again, super simple to channel math requests to an engine that's built for that.

> If they haven't fixed a particular shortcoming yet it's because they're seeking more complete solutions than picking off one tiny complaint at a time. But these are trivial to fix and smarmy posts like this will not age well.

The whole problem with this approach is a) we don’t have all the systems we need individually at a level of capability that meets the current level of hype and b) we don’t have a capable enough system management AI architecture, just using some weighted tokens is horribly deficient.

Yeah I have used Semantic Kernel to code AI in .NET and I did not give it the capability to tell the current date and time but it would be a 5 minute fix to do so since getting the current time is trivial. The bigger problem would be ensuring the offline server the AI runs on has its clock set correctly.

Asked the thing to tell me:

> I have heard that you sometimes give wrong time and sometimes even answer that you cannot tell the time, why does this happen to some people?

> ChatGPT said:

> You see different behavior because the system reads time from the userâ(TM)s environment. Some users block this data. Some browsers hide it. Some devices send inconsistent region info. This leads to gaps.

> You see steady answers because your browser shares stable time data.

> Other users trigger fallback behavior. In those cases I answer that I cannot read the time.

is this it?

Because ChatGPT and most AIs only work with digitally encoded data. The movement of an arm, like turning a dial, implies a translation from digital to real-world analog. Translating digital to interface with an analog world is really hard. Another example would be self-driving.

There has been [1]plenty of progress [interestin...eering.com] in using AI to control robotics; they use robotics-specific AIs for that, of course.

The fact that ChatGPT (or even LLMs in general) isn't particularly useful for robots shouldn't be a surprise, since robots (other than maybe C3PO) are about physical manipulation of objects, not about language generation.

[1] https://interestingengineering.com/ai-robotics/humanoids-complete-household-tasks-with-less-training

It all depends on what you mean by "think." For many purposes that I've used them, they "think" a lot better than employees I've had. These programs are more than just LLMs.