I'll be the first to say that chatgpt already does very well at the first and third examples. WebGPT sounds like the right approach for the others (understanding how to formulate a search query, then searching, then summarising the results)

Edit

I think the direction of these models seems clear, and their capabilities however limited at the moment are outstanding. I think the best measure of performance is what the naysayers complain about. We've quickly gone from "actual gibberish that looks like paragraphs" to "yes but it loses track after a few sentences" to "yes it's kind of coherent but doesn't make sense totally" and are now at "it's like an overconfident human at many tasks, when they're only somewhat qualified in the area". People are pointing out where it's got an API call wrong, without it having any ability to look up the answer and see if the function actually exists.

I spoke to it about distributed computation design and discovered something new I can actually implement in my design.

I'm very excited about where this will all go because it seems useful now and I'd be very surprised if this is the best that it ever reaches.

I agree (with minor caveats).

The important thing is that these AI systems are only going to get better over time.

IMHO, the early 2020's for conversational AI is akin to the 1980's for personal computers.

Absolutely, I totally agree. I edited my comment to add in more as I realised I'd not addressed that. It's surprisingly good right now, and given that it's useful that feeds into even more attention and money being spent on making it better.

Anyone not looking at chatgpt and thinking "what if this was even just a bit better? Or had better iterations or lookups?" is missing out on what I think is a revolutionary change coming probably sooner than people expect.

Agree. ChatGPT is how I imagined the internet would be back in the 90s (without having the foggiest clue how it might get there). Your questions are similar to the types of searches that people really want but have to translate to 'GPS simple explanation' or 'COVID in <town> this year' while hoping for a set of relevant pages in the Google return.

I've gotten it to fail on some word problems, but even then it was impressive to see it try to come up with a solution. More humorous were requests like 'Create a post for ycombinator about ChatGPT in the style of Chris Farley'. It's surprising and quite welcome that it even attempts that.

Another way to think about these kinds of requests in a future AI world is to ask yourself why you're asking these questions? My take:

"Computer, can you write and deliver a message to my kid that explains how GPS works?"

"Computer, can you tell me what risks there are at dining at McDonalds down the street?"

"Computer, can you read this paper [link] and tell me why it's referencing Claude Shannon?"

"Computer, I need to waste some time - please use HN to help me"

"Computer, tell me about any plans that I have for today"

"Computer, please find out how GPS works and explain it to me like I'm five years old."

ChatGPT:

GPS stands for Global Positioning System. It is a way that we can figure out where we are in the world using special satellites.

Imagine you are standing outside and you see lots of stars in the sky. Well, GPS works kind of like that, but instead of stars, there are satellites orbiting the Earth. These satellites are way up high, so they can see everything on the ground.

When you have a GPS device, like a phone or a car, it can talk to these satellites and ask, "Where am I?" The satellites send back a message that says, "You are here!" and gives your device a special code.

The GPS device can use this code to figure out exactly where you are on a map. It's kind of like a magic trick! You can use GPS to find out where you are, or to find your way to a new place. It's super helpful when you're lost or want to go on an adventure.

Not bad!

The second half of that explanation is completely wrong though, isn’t it? There is no active signal being sent from your GPS antenna, no two-way communication.

Here's it's reply to me:

Ian Calvert Computer, please find out how GPS works and explain it to me like I'm five years old

Sure thing! GPS stands for Global Positioning System. It's a way to find out where you are on Earth.

GPS works by using satellites that are way up in space. There are lots of these satellites orbiting around Earth, and they are always sending out signals. Your GPS receiver is a special device that can pick up these signals from the satellites.

The GPS receiver knows how far away it is from each satellite because of the time it takes for the signal to reach it. By figuring out how far away it is from a few different satellites, the GPS receiver can figure out exactly where it is on Earth.

So, if you have a GPS receiver in your car, phone, or on a special device like a watch, it can use the signals from the satellites to tell you exactly where you are and how to get to where you want to go. Does that make sense?

---

I was also able to ask it to make it simpler and simpler and it did so pretty well.

Interesting, the one you got id basically correct.

And comparing these two results for the same question serves to illustrate how ChatGPT does not in any meaningful sense _understand_ what it is "talking" about. It basically just strings together relevant-seeming words in a plausible way. The result may be useful, or it may be nonsense.

You'd better know enough (and be alert enough) to tell the difference...

And because of how it responds to them, people often forget that ChatGPT is just a language model, not an all-knowing AI.

GPS doesn't talk to the satellites, but the satellites do send a message that allows the GPS to answer "where am I" for itself. It's not the worst way to explain things at that level, even if it's misleading.

About half the (presumably human) "eli5 GPS" answers I found on Reddit made the same mistake, so chatGPT just copied an apparently popular misconception.

Might not be the worst way to explain how GPS works, but definitely up there.

You'd be better off just saying "it uses satellites and stuff".

> About half the (presumably human) "eli5 GPS" answers I found on Reddit made the same mistake, so chatGPT just copied an apparently popular misconception.

I think that's one of the biggest problems with using LLMs for accurate answers. A lot of the text that's very useful for modelling human language and questions is also full of factual errors and questionable opinions. Answers being wrong is a problem with old-fashioned web search too, but I think LLMs necessarily lean more on quantity rather than quality of content...

It is better at doing "translations" from one "language" to another "language". While it does know stuff and can answer questions about what it knows, that isn't its strength.

If you can do "here is the structured data, summarize it and make sure to present this information" it can do that quite well.

So far I've played with GPT doing a zero shot classification of HN titles and CSV hourly weather data to summary. In each case, it has done quite well.

Neither of those are things about what it "knows."

The benefit of web searches, and trouble with LLMs, is you can see the site that an answer is on and judge the credibility based on that. If I see that the answer is on Reddit ELI5, then I know that there's no assurance about how accurate the answer is. If it takes me to a page for a university engineering or math department, or the blog of a dude who built a receiver, then I can be a lot more confident that the answer is correct, though it might be technically over my head.

Yes, but unlike most times it gets things wrong, this is probably appropriate for explaining things to 5 year olds.

(Also, the UK comedy quiz show QI made the exact same mistake).

> this is probably appropriate for explaining things to 5 year olds.

I feel it's appropriate in terms of simplicity, but I think providing a simple wrong explanation is really only worth it when the more accurate explanation is substantially harder to simplify, and I don't think explaining how GPS works falls into that category.

Something like this I think is more accurate without being any more complex:

> Imagine you are standing outside and you see lots of stars in the sky. Well, GPS works kind of like that, but instead of stars, there are satellites orbiting the Earth. These satellites are way up high, so you can always see some of them.

> When you have a GPS device, like a phone or a car, it can look for these satellites, and if it sees enough of them it can use math to figure out where you are.

I’ll admit it was a decent answer. However, even a somewhat bright person (especially someone with radio / networking experience) could probably have realized that was an issue without any prior knowledge of GPS. They know that their phone has a GPS receiver, and that satellite phones accessing geostationary or high altitude satellites are much bulkier. Clearly, your cell phone isn’t transmitting to each satellite.

How will future LLMs be able to conduct that level of reasoning?

You can just as easily ELI5 this without spreading misinformation, though. Just change the third paragraph to:

> The satellites all broadcast a special code saying where each of them is. When you have a GPS device, like a phone or a car, it can listen to these codes and find out where and how far away different satellites are.

(Actually, this would be a fun thing to do with a class. Have a few kids march around the edges of the playground, saying things like “I’m next to the swingset now!”, and then ask one one kid to close their eyes, put them somewhere in the middle and ask them to guess where they’re standing.)

The first half is wrong too. GPS has nothing to do with satellites being able to see lots of things from high up. The whole thing is just nonsense that looks plausibly like an explanation until you try to decipher it.

I didn't even notice that bit (maybe GPT is a flat-earther?).

Something GPT discourse has been demonstrating to me is that I'm not usually a very careful reader. I apparently skim a lot. Or maybe I skim GPT outputs because I'm biased in my expectations already?

I’ve noticed this too! It’s excellent at mimicking an expert voice, and it puts me off guard.

So just like a typical explanation of things for five year olds?

Achieving simplicity by glossing over details is distinct from achieving simplicity by stating something incorrect.

Which one is more appropriate and/or typical will probably depend on what questions the five-year-old is asking, but I think it's reasonable to say the former is usually preferable to the latter.

>GPS has nothing to do with satellites being able to see lots of things from high up.

It does in a sense because the radio waves need an approximate line of sight to reach your GPS receiver. Being high up gives them a large coverage.

The explanation is misleading. The satellites don't see you and don't care about you. They are beacons sending a radio signal (the beacons says: i'm satellite X and here is a timestamp). The receiver based on the known position of the satelites and the beacon signals received figures out your position on earth using math!

So the explanation is either: satellites and their position or give the full explanation. Not this regurgitation.

And it's only going to get better over time.

As I mentioned elsewhere, conversational AIs in the early 2020's are like PCs in the 1980's -- a new technology that still has lots of early limitations.

Did ChatGPT come up with this, or just rip it off from some place(s) without attribution?

I've seen this sort of response a lot lately, both for GPT and all the image generators.

Google, and all the other search engines, literally show snippets of the original content to display search results, and have had copyright infringement lawsuits over this.

Yet somehow, it's taken these AI, which generally aren't just block-quoting things unless you ask them to (and often getting it wrong if you do), for non-lawyers to start asking that question.

I think that just indicates it's more blatant. With search engine results I can click through to the original source. That's attribution, at least to the laymen. ChatGPT has no such thing.

I selected a few sentences from the given response, asked DuckDuckGo to find them, and got zero hits for any of them.

GPT definitely has some kind of internal model of things and how they relate to other things, and the capacity to express those relations in different ways.

Whether this is functionally the same as the model in my head, or some analog to it, or if such descriptions are merely cute but misleading anthropomorphisation, is far beyond my knowledge of neuroscience, psychology, and indeed machine learning.

> I selected a few sentences from the given response, asked DuckDuckGo to find them, and got zero hits for any of them.

Science.

Is it any different than an infinitely capable human reading the entire internet and using the knowledge it gained to answer questions?

Google or GPT?

Either way I don't know. Might be the same, but that might be an unwarranted anthropomorphisation.

I'd be more impressed if it explained set-theoretic forcing on that level.

Except for having nothing to do with how GPS works. The satellites essentially broadcast what time it is and your reciever uses the delay between broadcast and reception to calculate where/when it is.

Explain like Im five means break it down, not lie about the complicated bits.

>Not bad!

Its terrible. This is totally not how GPS works, I.E. the terminal never sends a message to the satellite. This is the danger of early LLMs, a large amount of their output is false and there is no way to tell unless you are already an expert.

This makes them basically useless except for a toy.

Asking it to “explain how GPS works like I’m a seasoned engineer” yields a more detailed explanation and there’s no mention of GPS terminals sending messages back to the satellites.

Probably the “false output” you mentioned is due to the fact that the prompt asked to “explain it like I’m five”, which made ChatGPT answer with a “dialogue between terminal and satellite” explanation, which (arguably) may be better understood by a 5 years old.

I asked it to explain it to me like I'm Shrek.

Imagine that the GPS satellites orbiting the Earth are like the magic lanterns in the sky, sending signals down to a receiver (like the one in your phone or car). The receiver can use those signals to figure out how far away it is from each of the magic lanterns. By measuring the distance to multiple lanterns, the receiver can triangulate its position and figure out exactly where it is on the surface of the Earth. So even if you're a big, green ogre like Shrek, you can use GPS to find your way around and never get lost!

That is, if you can blindly trust the answers.. if it could provide the top sources for the information so I could fact-check it if needed, it would work

That's why in some respects, I think the sheer clunkiness of "according to website.com, claude shannon is..." or "ok, I've found this on the web..." in the current consumer tech is better than anonymous, authoritative somewhat novel answers than might be the internet's consensus opinion or might just be a vaguely relevant word vector structured like an answer

Is there a way to disuade users from blindly trusting responses? Such as unavoidable caveats, make no attempt to provide depth and instead strongly recommend reliable research?

And as a logical aside: if AI can't be trusted, we shouldn't trust AI's recommendations for reliable sources either.

> And as a logical aside: if AI can't be trusted, we shouldn't trust AI's recommendations for reliable sources either.

Don't think that's automatically true though: an AI can be consistently excellent at retrieving reliable sources and still get a lot wrong on its summaries. Also humans have their own opinions on sources even when they're not familiar with the detail, and finding out if the AI reads Nature or 4chan is part of the debugging process...

Sure. Really, I was thinking of the scenario where someone explictly asks for a reliable source i.e. "Thanks for summary about X. To whom should I turn for more accurate/detailed information?"

Oh yeah, it's spectacularly bad at some of that sort of thing. Ask for a list of academic papers and it'll make half of them up and give others new authors...

We're going to have a generation of people that are cock sure about things because ... ai, but they won't actually know jack shift. It sort of happened with google. People can figure something out but they don't systematically get things. Shallow explanation. Goldfish attention span

I think that assumes AI makes any difference to the self-certainly of humans.

My mother was into all kinds of new age and supernatural stuff. You may think you can see where this anecdote is leading, and oh boy I could go there, but instead prepare for a twist! One of the stories she had was a client saying "It is not possible, Susan, it is not possible that there are beings out there that are cleverer than what we is".

Have you had much to do with young graduates lately? We already have that generation without AI being the cause.

Like Ask Jeeves on steroids.

>"Computer, please find out how GPS works and explain it to me like I'm five years old."

”I’m sorry, I cannot answer that query because it has ableist connotations. I am a large language model by OpenAI.”

I guess being young is a disability.

where will this data come from and how do you make money off it?

By inserting "enjoy an ice cold refreshing cocktail cola" in the middle of the output?

So, what you're saying is that "I'm feeling lucky"?

Can LLMs be updated quickly the way search engines are by crawlers and what not? ChatGPT is a couple years out of date and there was a lot of human involvement. Doesn’t sound like a quick or automated process. GPT-4 is still under development.