The other day I asked it about the place I live and it made up nonsense, I was trying to get it to help me with an essay and it was just wrong, it was telling me things about this region that weren't real.
Do we just drive through a town, ask for a made up history about it and just be satisfied with whatever is provided?
I have tried to use it many times to learn a topic, and my experience has been that it is either frustratingly vague or incorrect.
It's not a tool that I can completely add to my workflow until it is reliable, but I seem to be the odd one out.
I find this highly concerning but I feel similar.
Even "smart people" I work with seem to have gulped down the LLM cool aid because it's convenient and it's "cool".
Sometimes I honestly think: "just surrender to it all, believe in all the machine tells you unquestionably, forget the fact checking, it feels good to be ignorant... it will be fine...".
I just can't do it though.
It's the same issue with Google Search, any web page, or, heck, any book. Fact checking gets you only so far. You need critical thinking. It's okay to "learn" wrong facts from time to time as long as you are willing to be critical and throw the ideas away if they turn out to be wrong. I think this Popperian view is much more useful than living with the idea that you can only accept information that is provably true. Life is too short to verify every fact. Most things outside programming are not even verifiable anyway. By the time that Steve Jobs would have "verified" that the iPhone was certainly a good idea to pursue, Apple might have been bankrupt. Or in the old days, by the time you have verified that there is a tiger in the bush, it has already eaten you.
When I spend time on something that turns out to be incorrect, I would prefer it to be because of choice I made instead of some random choice made by an LLM. Maybe the author is someone I'm interested in, maybe there's value in understanding other sides of the issue, etc. When I learn something erroneous from an LLM, all I know is that the LLM told me.
People should be able "throw the ideas away if they turn out to be wrong" but the problem is these ideas unconsciously or not help build your model of the world. Once you find out something isn't true it's hard to unpick your mental model of the world.
Intuitively, I would think the same, but a book about education research that I read and my own experience taught me that new information is surprisingly easy to unlearn. It’s probably because new information sits at the edges of your neural networks and do not yet provide a foundation for other knowledge. This will only happen if the knowledge stands the test of time (which is exactly how it should be according to Popper). If a counterexample is found, then the information can easily be discarded since it’s not foundational anyway and the brain learns the counterexample too (the brain is very good in remembering surprising things).
The benefit is that I got a quick look at various solutions and quickly satisfied a curiosity, and decided if I’m interested in the concept or not. Without AI, I might just leave the idea alone or spend too much time figuring it out. Or perhaps never quite figure out the terms of what I’m trying to discover, as it’s good at connecting dots when you have an idea with some missing pieces.
I wouldn’t use it for a conversation about things as others are describing. I need a way to verify its output at any time. I find that idea bizarre. Just chatting with a hallucinating machine. Yet I still find it useful as a sort of “idea machine”.
I think even if an AGI was created, and humans survived this event. I'd still have trouble trusting it.
The quote "trust but verify" is everything to me.
I don't like being told lies in the first place and having to unlearn it.
It doesn't help that I might as well have just gone straight to the "verification" instead.
It's smart but can also be very dumb.
I think the current state of AI trustworthiness (“very impressive and often accurate but occasionally extremely wrong”) triggers similar mental pathways to interacting with a true sociopath or pathological liar for the first time in real life, which can be intensely disorienting and cause one to question their trust in everyone else, as they try to comprehend this type of person.
I think it's because Americans, more than nearly all other cultures, love convenience. It's why the love for driving is so strong in the US. Don't walk or ride, drive.
Once I was walking back from the grocer in Florida with 4 shopping bags, and people pulled over and asked if my car had broken down and if I needed a ride, people were stunned...I was walking for exercise and for the environment...and I was stunned.
More evidence of this trend can be seen in the products and marketing being produced:
Do you need to write a wedding speech? Click here.
Do you need to go get something from the store? get your fat ass in the car and drive, better yet, get a car that drives for you? Better than this, we'll deliver it with a drone...don't move a muscle.
Don't want to do your homework? Here...
Want to produce art? Please enter your prompt...
Want to lose weight? We have a drug for that...
Want to be the authority on some topic? We'll generate the facts you need.
As convenience in a domain becomes ubiquitous or at least expected among consumers, they quickly readjust their evaluation of "having time for X" around the new expectation of the convenient service, treating all alternatives as positive opportunity cost. This would explain a lot of those folks who are upset when it's suggested that they don't need Amazon, Instacart, etc. in their lives if they are to do something about their contributions to mass labor exploitation.
Of course these conveniences quickly become ubiquitous in large economies with a glut of disposable income, which encourages VCs to dump money into these enterprises so they're first to market, and also to encourage the public to believe that the future is already here and there's no reason to worry about backsliding or sustainability of the business model. Yet in every single case we see prices eventually rise, laborers squeezed, etc. A critical mass of people haven't yet acknowledged this inevitability, in no small part due to this fixation on convenience at the expense of more objective, reasoned understandings (read: post-truth mindset).
This. I hate being told the wrong information because I will have to unlearn the wrong information. I would rather have been told nothing.
They're only good on universal truths. An amalgam of laws from around the globe doesn't tell me what the law is in my country, for example.
I feel like using LLM today is like using search 15 years ago - you get a feel for getting results you want.
I'd never use chatGPT for anything that's even remotely obscure, controversial, or niche.
But through all my double-checking, I've had phenomenal success rate in getting useful, readable, valid responses to well-covered / documented topics such as introductory french, introductory music theory, well-covered & non-controversial history and science.
I'd love to see the example you experienced; if I ask chatGPT "tell me about Toronto, Canada", my expectation would be to get high accuracy. If I asked it "Was Hum, Croatia, part of the Istrian liberation movement in the seventies", I'd have far less confidence - it's a leading question, on a less covered topic, introducing inaccuracies in the prompt.
My point is - for a 3 hour drive to cottage, I'm OK with something that's only 95% accurate on easy topics! I'd get no better from my spouse or best friend if they made it on the same drive :). My life will not depend on it, I'll have an educationally good time and miles will pass faster :).
(also, these conversations always seem to end in suffocatingly self-righteous "I don't know how others can live in this post-fact free world of ignorance", but that has a LOT of assumptions and, ironically, non-factual bias in it as well)
I don't think it's quite the same.
With search results, aka web sites, you can compare between them and get a "majority opinion" if you have doubts - it doesn't guarantee correctness but it does improve the odds.
Some sites are also more reputable and reliable than others - e.g. if the information is from Reuters, a university's courseware, official government agencies, ... etc. it's probably correct.
With LLMs you get one answer and that's it - although some like Bard provide alternate drafts but they are all from the same source and can all be hallucinations ...
Yes and no. If the LLM is repeating the same thing on multiple drafts then it's very unlikely to be a hallucination.
It's when multiple generations are all saying different things that you need to take notice.
LLMs hallucinate yes but getting the same hallucination multiple times is incredibly rare.
and that wouldn't eliminate hallucinations just tell you if large details have likely been hallucinated.
But it's a method some research has used.
P.S. Also aren’t LLMs deterministic if you set their “temperature” to zero? Are there drafts if the temperature is zero? If not, then that’s the same as removing the randomness no?
>Just do the comparison on the user’s machine if the LLM provider is that cheap.
This is not possible. Users don't have the resources to run these gigantic models. LLM inference is not cheap. Open ai, Google aren't running profit on free cGPT or Bard.
>P.S. Also aren’t LLMs deterministic if you set their “temperature” to zero? Are there drafts if the temperature is zero? If not, then that’s the same as removing the randomness no?
It's not a problem of randomness. a temp of 0 doesn't reduce hallucinations. LLMs internally know when they are hallucinating/taking a wild guess. randomness influences how that guess manifests each time but the decision to guess was already made.
I never said it did.
> LLMs internally know when they are hallucinating/taking a wild guess.
No they don’t. If they did we would be able to program them to not do so.
I would argue that wild guesses are all LLMs are doing. They practically statistically guess their way to an answer. It works surprisingly well a lot of the time but they don’t really understand why they are right/wrong.
P.S. LLMs are kind of like students who didn’t study for the test so they use “heuristics” to guess the answer. If the test setter is predictable enough, the student might actually get a few right.
"In particular, we find that LMs often hallucinate differing authors of hallucinated references when queried in independent sessions, while consistently identify authors of real references. This suggests that the hallucination may be more a generation issue than inherent to current training techniques or representation."
https://arxiv.org/abs/2303.08896
"SelfCheckGPT leverages the simple idea that if a LLM has knowledge of a given concept, sampled responses are likely to be similar and contain consistent facts. However, for hallucinated facts, stochastically sampled responses are likely to diverge and contradict one another."
I've seen the hallucination rate of LLMs improve significantly, if you stick to well covered topics they probably do quite well. The issue is they often have no tells when making things up.
A person who uses ChatGPT must have the understanding that it's not like Google search. The layman, however, has no idea that ChatGPT can give coherent incorrect information and treats the information as true.
Most people won't use it for infotainment and OpenAI will try its best to downplay the hallucination as fine print if it goes fully mainstream like google search.
no I will not give the public credit, most people have no grounding to discern wtf a language model is and what it's doing, all they know is computers didn't use to talk and now they do
It seems to be able to speak on history, sometimes it's even right, so there's a use case that people expect from it.
FYI I've used GPT4 and Claude 2 for hundreds of conversations, I understand what its good and bad at; I don't trust that the general public is being given a realistic view.
Rather than asking it about facts, I find it useful to derive new insights.
For example: "Tell me 5 topics about databases that might make it to the front page of hacker news." It can generate an interesting list. That is much more like the example they provided in the article, synthesizing a bed time story is not factual.
Also, "write me some python code to do x" where x is based on libraries that were well documented before 2022 also has similarly creative results in my experience.
Like talking to most people you mean?
You will now be able to feed it images and responses of the customers. Give it a function to call complementaryDrink(customerId) Combine it with a simple vending machine style robot or something more complex that can mix drinks.
I'm not actually in a hurry to try to replace bartenders. Just saying these types of things immediately become more feasible.
You can also see the possibilities of the speech input and output for "virtual girlfriends". I assume someone at OpenAI must have been tempted to train a model on Scarlett Johansson's voice.
If people are treating LLMs like a random stranger and only making small talk, fair enough, but more often they're treating it like an inerrable font of knowledge, and that's concerning.
That's on them. I mean, people need to figure out that LLMs aren't random strangers, they're unfiltered inner voices of random strangers, spouting the first reaction they have to what you say to them.
Anyway, there is a middle ground. I like to ask GPT-4 questions within my area of expertise, because I'm able to instantly and instinctively - read: effortlessly - judge how much to trust any given reply. It's very useful this way, because rating an answer in your own field takes much less work than coming up with it on your own.
All human interactions from all of history called and they …
I verify just about everything that I ask it, so it isn’t just a general sense of improvement.
Take a look at https://chat.openai.com/share/41bdb053-facd-448b-b446-1ba1f1... for example.
Context: had a bunch of photos and videos I wanted to share with a colleague, without uploading them to any cloud. I asked GPT-4 to write me a trivial single-page gallery that doesn't look like crap, feeding it the output of `ls -l` on the media directory, got it on first shot, copy-pasted and uploaded the whole bundle to a personal server - all in few minutes. It took maybe 15 minutes from the idea of doing it first occurring to me, to a private link I could share.
I have plenty more of those touching C++, Emacs Lisp, Python, generating vCARD and iCalendar files out of blobs of hastily-retyped or copy-pasted text, etc. The common thread here is: one-off, ad-hoc requests, usually underspecified. GPT-4 is quite good at being a fully generic tool for one-off jobs. This is something that never existed before, except in form of delegating a task to another human.
https://chat.openai.com/share/338e7397-0201-44f4-a2c3-75b733...
I use ChatGPT for all sorts of things - looking into visas for countries, coding, reverse engineering companies from job descriptions, brainstorming etc etc.
It saves a lot of time and gives way more value than what you pay for it.
It used to be more reliable when web browsing worked, but it's still pretty reliable.
Ah yes, I dont understand how to talk to people either!
Comments like yours make me think that no one cares about this...and judging by a lot of the other comments, I guess they don't.
Probably going to be people, wading through a sea of AI generated shit, and the individual is supposed to just forever "apply critical thinking" to it all. Even a call from ones spouse could be fake, and you'll just have to apply critical thinking or whatever to workout if you were scammed or not.
Then it makes stuff up far less frequently.
If the next version has the same step up in performance, I will no longer consider inaccuracy an issue - even the best books have mistakes in them, they just need to be infrequent enough.
> Then it makes stuff up far less frequently.
Now there's a business model for a ChatGPT-like service.
$1/month: Almost always wrong
$10/month: 50/50 chance of being right or wrong
$100/month: right 95% of the time