This is terrible. The idea of ChatGPT being able to fill in gaps of not knowing what you don’t know is appealing, but you need to actually know something about the topic to evaluate ChatGPTs ability.
This is just spewing pure nonsense, and the author thinks he has learned something from it.
The worst part is the "p-values" it completely made up.
Verification that the numbers were legitimate -- "What's a p-value."
At least it got the values right .9912 (twice), .8813 (twice). It estimated 88-99% chance of bullshit... To 4 significant digits... Repeating the exact values.
Turns out, this is an excellent example of how poor these models are with accurate knowledge vs confident bullshitting.
This. If you've been following GPT at all you'd be crazy to try to use it to learn new things. It's very clear the output is not accurate, and prone to hallucination, and requires an already knowledgable user to judge its veracity and validity.
I was worried something like this would happen. AI will usher in a dark age thanks to overfitting and convenience and naivety. It's going to be like the current info literacy crisis raised to the nth degree.
It's also highly suggestible. If the input was interpreted wrongly, it can deviate from a set goal. Garbage in => garbage out.
Example:
>> If I'm feeling a little horse, where am I?
>> If you're feeling a little horse, you might be experiencing a sore throat or hoarseness in your voice. This is a common symptom of many respiratory illnesses, such as the common cold, flu, or bronchitis, as well as other conditions that affect the throat and vocal cords, such as acid reflux, allergies, or overuse of the voice. It's important to stay hydrated and rest your voice if you're experiencing hoarseness, and to seek medical attention if your symptoms persist or worsen over time.
Treated as an opportunity for criticism, it can be an effective learning tool, especially for developing your critical literacy. Think of it as a game where the point is to prove it wrong.
People already don't do that with our current tools, what makes anyone believe it will happen with ChatGPT? Especially when they now have to go back to their favorite search engine to find those sources and validate them.
It's as if everyone has forgotten how humans behave, they don't even read or go to actual wikipedia sources that are properly laid out! Nevermind having to go find trustworthy sources.
I don't really care about how others use it. I myself find it enough useful and productive to use it, and what I wrote before is how I "protect" myself. As you say, people who don't check sources won't do so with or without GPT4.
Maybe you'll eventually care about how others use it when your trustworthy sources are flooded with irrelevant, outdated, and downright false information due to the constant generation of "kind of true" content.
Do you think that other sources are infallible? Almost every source is wrong sometimes. ChatGPT is as good as my professors were in college. They would regularly say things that were just completely incorrect but were knowledgeable and correct often enough to be helpful.
Your professors generally don’t make stuff up if they don’t know something. Your professors are also more likely to catch on if something doesn’t make sense - unlike ChatGPT which can’t seem to reason effectively because most of the time it doesn’t know what’s it’s talking about.*
* Maybe it has the ability to reason, maybe not. Maybe it just needs more information - after all it’s trying to model reality with just the words used to describe it. Or maybe it really can’t reason at all, a limitation of its algorithm, and has to rely on having arguments/explanations already memorised - which seems to be mostly what it does now. Either way so far it hasn’t demonstrated that it’s all that good at reasoning.
Some of them most certainly did make stuff up. Ego prevents many people from admitting that they don't know something.
I also interacted with some who wouldn't even admit to being wrong when they were confronted with facts. So I don't know that I would say they catch on either.
Pilots make mistakes sometimes. I, the bike rider also make mistake sometimes. Therefore I can fly an airplane and I will be as good.
Jokes aside, making mistakes is not a good common factor in order to evaluate if two different systems can be practically interchanged. You need more. What type of mistakes? In what situations, and how to mitigate them. This is a surprisingly common misunderstanding running wild here. You cannot even switch the places of two humans, even two professors in real life using this line of reasoning.
How do you know that ChatGPT is as good as your professors in college? How do you know that they aren't just as good as a cable news pundit? Or a relatively uninformed Reddit commenter?
All I'm saying is that ChatGPT is knowledgeable. It has absorbed a ton of information and can distribute that information fairly well. I don't think that can be argued with.
Yes, it's completely wrong sometimes but the exact same can be said for any human (my professors, the cable news pundit, average redditor).
To say I can't learn from it doesn't make any sense. I have learned everything from flawed sources. We all have.
> I have learned everything from flawed sources. We all have.
Sure, some historical books in the US South are wildly inaccurate!
This is why we deal with multiple sources, as well as defining, validating, and exploring our sources. ChatGPT does not do this. It does not pull from definable sources, it is an LLM that outputs text based on statistical word sequences from an almost bottomless source.
You can learn from it, but the sources may be sketchy, and it will mix source biases. I tried to ask it for sources on the fall of ancient Rome:
>> As an AI language model, I don't have the ability to consult sources or remember my previous responses. However, the information I provided about the fall of Rome is widely accepted in historical scholarship and can be found in many reputable sources such as encyclopedias, textbooks, and academic articles.
Personally, I think we need to tune the LLMs according to use. Creative? Then it one way. More factual, the other way. This does mean we would need to know sources for LLMs.
I was thinking maybe you had some specific metrics or examples in mind since it's your own comparison...
The rest of your post reads like the old "there are two possible outcomes therefore the odds are 50/50" probability fallacy.
Do I believe it's more useful than chatroullette? Yes. Does that mean it's as useful as a college professor on a given subject? That's a very different claim, and would need some serious evaluation and demonstration.
Other sources are fallible as well, but also less likely to make big mistakes and more likely to cover the major pitfalls of a hobby/area of expertise by virtue of actual experience and understanding of the subject matter. (versus statistical probabilities in word choice)
Also ChatGPT has no consistent internal sense of self and is highly sensitive to input. You can convince it that just about anything is true if you aren't careful about how you phrase your input.
Wikipedia works since the world is full of pedantic sticklers. If you remove the pedantic sticklers then wikipedia would no longer work, it would just be full of bias or bullshit, so we should be thankful for all the pedantic sticklers out there.
> In contrast to AI’s many benefits is the fact that it will displace millions of people around the world from their current workplace roles, especially those in white-collar jobs such as customer service, copywriting, and computer programming. It has already started to do so.
The claim might be true but this evidence does not establish it. A company having planned to replace a kind of job with AI is quite distinct from the number of that kind of job falling as a result of AI.
I agree, millions is entirely unknowable right now, but there were two distinct statements and you didn’t say which you wanted a citation for, so I supplied a citation for the verifiable statement and asked which you were challenging.
...or more things will be made, more demand equals more people copy and pasting from ChatGPT-4 just like you?
The only way what you're saying is true if there is a limit to the amount of new opportunities everyone being so productive will create. There's nothing to suggest there is.
I often feel depressed about how hard it is to make products better, if this tool enables that to happen, guess what I'll be doing? Building more products.
Also, many developed countries such as Japan, South Korea etc were almost facing economic collapse due to not having enough workers, so they basically "need" these tools to survive in the future.
Maybe we're going to see a world with less people, and an abundance of awesome shit?
For getting shipping minimum viable product and say hello to shipping maximum awesome product?
Tired of Crypto hype? Well, it passed. Now they hype is big on AI. Useful things will surely come out of it but before lots of money will be made on hype alone.
There was hype around Crypto because people were making large sums of money from it, despite unclear "real" uses beyond easily moving large sums of money around.
There is hype around AI because hundreds of millions of people (including me) are using AI tools every day to help them with their work. In contrast to crypto, random people are not making enormous amounts of money; there are no 18 year olds in AI stumbling into hundreds of millions of dollars. Even OpenAI is not particularly profitable and is mainly sustained through its close partnership with Microsoft.
I think this reversal of the money <-> utility relationship makes these technologies fundamentally different, and lumping them together into the category of "just another hype cycle" is unfair.
There's money to back the hype alright, but it resides with businesses/academia/government who are itching to find a way to buy their way into an AI-enabled modernization.
Just a quibble, but "hundreds of millions of people" seems high by an order of magnitude or two. That would be around 1 in every 20 people in the world.
But this is known. Has OpenAI released usage figures?
I'm on a one-person crusade against that 100m user number. It comes from a report released by a company that gathers analytics from people who were convinced to install dodgy browser extensions!
I made some notes here: https://simonwillison.net/2023/Feb/19/how-chatgpt-kicked-off... - I think ChatGPT had 30m users by Feb 3rd (a date later than that 100m number), because "two people with knowledge of the figures" leaked that to Kevin Roose at the NY Times.
I'm sure it's gone up a bunch since then. Maybe it has hit 100m even - but I don't think that original 100m number should be trusted by anyone.
Wow. 13 million per day = 1 in every 400 people on Earth with internet access, if Google is correct that world population is now 8e9 and 64.4% of the world has internet access.
It's hype for sure but for AI I can think of a ton of real-world use cases whereas all crypto use cases sounded pretty contrived. I think AI will be a revolution of the scale of the internet or even bigger. Maybe the current version of ChatGPT is flawed but the possibilities are there. I am thinking analyzing long contracts, writing letters, verifying system requirements and tests and so on. The potential seems endless.
Most of the times I have found chatgpt useful is to explain something I am not fully clear on, but have some idea about. Random example:
I read from this[1] source that CLO transactions(and hence cashflows) are "arbitrage". And I wanted to understand what is the exact arbitrage.
------------------------------
Prompt: Do you know what is Arbitrage CLOs in this context?
ChatGPT: Yes, I can explain what Arbitrage CLOs are in the context of Collateralized Loan Obligations (CLOs).
Arbitrage CLOs are a type of CLO that is structured to take advantage of differences in the yields of the loans included in the CLO's underlying asset pool. The goal of an Arbitrage CLO is to generate a higher yield than the cost of financing the CLO by purchasing the underlying loans at a lower cost and then securitizing them into higher-yielding CLO securities.
In an Arbitrage CLO, the issuer will purchase a pool of loans at a discount and then securitize them into several tranches of securities with different levels of risk and return. The securities are then sold to investors at a higher price than the cost of the underlying loans, generating a profit for the issuer.
The difference between the yield on the underlying loans and the yield on the CLO securities is known as the "spread," and it represents the profit that the issuer earns from the CLO. In an Arbitrage CLO, the spread is typically larger for the more junior tranches of securities, which offer higher yields but also carry higher risk.
Arbitrage CLOs can be an attractive investment for investors who are looking for high-yield, high-risk investments. However, they can also be volatile and carry significant risk, particularly in periods of economic stress or market volatility. As with any investment, it is important to carefully consider the risks and potential rewards before investing in an Arbitrage CLO.
-----------------------------------------
So, I learn't(or confirmed my intuition) that the arbitrage occurs because of discount on loans.
One thing ChatGPT probably won't tell you about macroeconomic theory unless specifically asked about it is that a lot of the foundational assumptions are highly idealized and aren't that useful for making real-world predictions:
> "But what is not widely known is that these now legendary economists—William Stanley Jevons, Léon Walras, Maria Edgeworth and Vilfredo Pareto—developed their theories by adapting equations from 19th-century physics that eventually became obsolete."
I'd probably trust ChatGPT more on elucidating the history of the development of macroeconomic theory then on applications of the theory to real-world problems, but that goes for the working professionals in the field as well, with all their rosy predictions of the benefits of NAFTA to the working public in the 1990s, or their predictions of the benefits of financial deregulation in the early 2000s prior to the 2008 Recession, and so on.
E.g. ChatGPT claims American academic institutions adopted macroeconomic theory with Milton Friedman at Chicago and Alvin Hansen at Harvard, and this was preceded by development in the 1920s at the National Bureau of Economic Research and in the 1930s with the Cowles Commission. If you ask it what role the above four individuals in the quote played in the development of macroeconomic theory, it gives illuminating answers on utility, general equilibrium, indifference curves, and Pareto efficiency (which, ChatGPT claims strongly, is not zero-sum game thinking). This all seems fairly accurate based on other sources, except that the relationship between game theory and Pareto efficiency seems fairly contentious.
Also, here's a good question to ask it in the context of the desirability of optimizing Pareto efficiency:
"Has macroeconomic theory ever been criticized on teleological grounds?"
Generally, this is probably not the best subject to give to ChatGPT as a quantitative test. Maybe try it out with something more concrete, like how to use gravitational assist (aka 'slingshot') to send a spacecraft to Jupiter or Saturn, and see how explicit it can get in terms of a launch trajectory?
Great effort article. But I just don't think re-skilling is a good solution to massive layoffs and social unrest. It's a pipe dream. How do you ask questions for macroeconomic topics you didn't already know? - by reading books which everyone could already do. chatGPT has become something you have to do after you've read the book. People will need more time reskilling now. Classic Jacques Ellul argument again. I'm sorry for negative comment but nobody should take this seriously. The point of the article is showcasing how chatGPT can be very useful study buddy.
I don't quite agree with the exact content of the article, but the headline, distilled into a general concept, has a lot more weight in my opinion.
Can AI help accelerate someone's learning or cause significant improvement? Short answer: Yes, specially with the right mindset. Chess AI managed to keep the game sharp and create new ways to enjoy chess, and I think it'll be the same when the dust settles.
Long answer, following my latest involvement with AI stuff:
I haven't really played that much with text models, ChatGPT and such, but I can kind of relate to this when Stable Diffusion is involved.
I'm a fairly competent artist, and I enjoy drawing both elaborate art pieces and doodling random stuff when bored or when someone says something funny in my socials, and despite the hate of many of my peers, I picked it up and got my own local instance to have fun with.
This thing has fully revolutionized my world as an artist. It's only been like half a year and I don't think I have drawn so much since high school, specially during tough work seasons where I'd normally be busy and too tired to draw in my free time. You can say that every year my quality goes up by a roughly equivalent amount, and this year I'm already outpacing myself.
I don't know, I can't quite explain it in simple terms, but I'd wager it's both the stimulus of seeing art of topics and characters that would normally take months or years between a single new image, and the AI failing at certain elements often (hands, fine details and decorations, eyes, etc...) making you more aware of those elements and thus becoming more mindful of them when drawing, even at a somewhat subconscious level.
Not to mention when the AI does something hilarious and you kinda make it a thing or draw it "for real" from scratch.
And, thinking about it, there's also some ancillary art exercises the AI has gotten me into.
Editing images (for training or "this generated image will be 100% perfect if I redraw this part myself" cases, which makes you learn a lot because of the huge amount of different styles this can generate. Gotta learn to recreate that shading style on the spot!), stuff like ControlNet accepting quick sketches as input, and wanting to duplicate some of the more fascinating things the AI generates. Training models can become a full "art project" given enough love and care and manual art supply. And this also incentivizes to be very clear and tidy for the training to go well. And all of this becomes practice and experience.
Being able to generate really decent images in a few seconds is never going to take away my fun or motivations when drawing, and it's opening new paths to have fun with art, too. In that regard, and finally tying with the thread's content, if the AI has managed to not only help me improve but make me even more active in art, I'm sure it can have similar effects in other fields as well. A writer can easily improve, learning from the terrible goofs, hilarious twists and occasional "getting it just right" moments the text models provide. I can see the same thing happening in areas like music or whatever comes next. An observant and motivated use can turn a ML system into a goldmine of custom examples, and at a ridiculous fast pace, without tiring and on demand!.
This is just spewing pure nonsense, and the author thinks he has learned something from it.