Why Is ChatGPT Bad At Math?

Head to linode.com/scishow to get a $100 60-day credit on a new Linode account. Linode offers simple, affordable, and accessible Linux cloud solutions and services.
Sometimes, you ask ChatGPT to do a math problem that an arithmetically-inclined grade schooler can do with ease. And sometimes, ChatGPT can confidently state the wrong answer. It's all due to its nature as a large language model, and the neural networks it uses to interact with us.
Want to hear our ChatGPT dinosaur poem? Check out our patreon at patreon.com/scishow!
Hosted by: Stefan Chin
----------
Support SciShow by becoming a patron on Patreon: / scishow
----------
Huge thanks go to the following Patreon supporters for helping us keep SciShow free for everyone forever: Matt Curls, Alisa Sherbow, Dr. Melvin Sanicas, Harrison Mills, Adam Brainard, Chris Peters, charles george, Piya Shedden, Alex Hackman, Christopher R, Boucher, Jeffrey Mckishen, Ash, Silas Emrys, Eric Jensen, Kevin Bealer, Jason A Saslow, Tom Mosner, Tomás Lagos González, Jacob, Christoph Schwanke, Sam Lutfi, Bryan Cloer
----------
Looking for SciShow elsewhere on the internet?
SciShow Tangents Podcast: scishow-tangents.simplecast.com/
TikTok: / scishow
Twitter: / scishow
Instagram: / thescishowfacebook: / scishow
#SciShow #science #education #learning #complexly
----------
Sources:
• How Computers Calculat...
www.sciencedirect.com/science...
books.google.co.uk/books?hl=e...
intjem.biomedcentral.com/arti...
www.ncbi.nlm.nih.gov/pmc/arti...
hal.science/hal-03913837v1/pr...
ai.stackexchange.com/question...
www.britannica.com/technology...
www.technologyreview.com/2023...
news.mit.edu/2023/large-langu...
www.mdpi.com/2079-9292/10/20/...
cds.cern.ch/record/400313/fil...
www.psychologytoday.com/gb/bl...
arxiv.org/pdf/2302.03494.pdf
arxiv.org/pdf/2301.13867.pdf
Images
www.gettyimages.com
commons.wikimedia.org/wiki/Fi...

[@SciShow]

Head to linode.com/scishow to get a $100 60-day credit on a new Linode account. Linode offers simple, affordable, and accessible Linux cloud solutions and services.

[@jamesjellis]

ChatGPT is actually pretty bad at everything that I have seen it output so far.

[@Ghekkoo]

na James, it’s pretty good

[@jamesjellis]

@@Ghekkoo well I haven't seen it then.

[@imeryakwut6437]

😊

[@steve_jabz]

Do you have the source for the study where chinese researchers added up to 1 trillion numbers with 99% accuracy? I'm trying to reference it but it's not in your sources or searchable

[@ironiccookies2320]

Speaking of ChatGPT being bad at math, ChatGPT also makes up sources and quotes.

[@Goldenwhitewings]

@@theorangeoof926that’s why you don’t do homework with it

[@MrNoobomnenie]

What a lot of people (including, apparently, even OpenAI PR managers) don't understand is that GPT-like models are not *supposed* to give you an accurate information. These machines are not encyclopedias - they are fiction writers. They are optimized to stay in character as authentically as possible, which doesn't necessarily requires being factually accurate - only *appearing* to be factually accurate. This doesn't mean that ChatGPT is "bad"/"stupid"/"worthless" - it is indeed an incredibly impressible technology with a ton of potential. The real problem is that it's currently being advertised and hyped for being something that it isn't (and, as a result, also being hated and dismissed for NOT being something that it isn't).

[@blinkx1070]

It also makes up stuff in general. I was using it to help get ideas for places to go on an upcoming trip to Japan and it told me Nishiguchi Park in Ikebukuro had a hill and a pond (it doesn't, I've been there). It's also important to note that currently it's knowledge database only goes up to 2021.

[@obs_o1468]

@@theorangeoof926 it does have a disclaimer , it tells you that the information is not guaranteed to be accurate especially after 2021 specifically , when you open it for the first time it also doesn’t have access to the internet, hence why no sources

[@jod125]

​@@theorangeoof926 It has a disclaimer saying info might be wrong, already.

[@Lucky10279]

Numberphile did a video about this a few months ago. They made the point that, when you really think about it, it's surprising that chatGPT is as good at arithmetic as it is, simply because it wasn't _designed_ to do arithmetic. Sure, calculators and calculator apps can do arithmetic problems nearly instantly and with nearly 100% accuracy ("nearly" because there's always some small probability that _something_ goes wrong, such as the famous video where a particular TI calculator erroneously gives an answer to a random arithmetic problem in terms of π), but they were specifically _designed_ to do that. ChatGPT and similar models, on the other hand, were designed to model _natural languages._ All they're _really_ doing is picking up on patterns in their training data and using those patterns to predict what the most likely output to a given input is. So, that it can usually at least get _close_ to the right answers to arithmetic problems on large numbers, even though it's unlikely its training data contained a large amount of such problems, is surprising. It says something about the similarities between the grammatical and syntactic rules of a natural language and the algorithmic rules of arithmetic that a tool designed to model the former can do reasonably well at the latter.

[@RemizZ]

Yeah it's just a word predictor, it doesn't actually do any math.

[@fritt_wastaken]

​@@RemizZ "it's just a word predictor" can imply literally anything from a random generator to a human to a super intelligence. It's a meaningless statement

[@oldcowbb]

i keep telling this, chatGPT is like a salesman, it knows how to talk without knowing what it is talking about

[@uponeric36]

@@oldcowbb That's just our speculation. Even if it knew what it was talking about and tried to tell us; we'd still not know what level of understanding it has.

[@Lucky10279]

@@uponeric36 It's not speculation. ChatGPT is nothing more or less than a very sophisticated algorithm that spits out statistically likely responses to prompts. It has no consciousness or understanding anymore than any other computer program does.

[@LFTRnow]

I tried multiple math problems and found it failed them quite often, and in the way humans do. A more interesting one was I asked how long something took in minutes and seconds. It gave something like 7 min and 80 seconds. It was right, but weird. Asking for it in mins and seconds gave the same answer, so I told it to limit the seconds to a range of 0-59.9. Result was "7 min and 20 seconds". Wrong. I spent several minutes telling it how to do this correctly and it eventually did, apologizing each time it was wrong - now here's the twist, you can then tell it is is something entirely wrong, like 9 mins and 20 sec and it will again apologize, and then make up bogus math to show why your new opinion is correct. It will ALWAYS pander, and never correct you or raise a concern regarding accuracy. Be warned.

[@Axodus]

It's raised concerns regarding accuracy before, not regarding math but philosophy. Yes, ChatGPT forced it's own opinions on me.

[@iamjustkiwi]

I've seen this sort of thing a lot with AI videos. AI is pretty good at being confidently wrong and then arguing with the user that it ISN'T wrong, and saying that the user is trying to be mean. Makes me feel weird watching it.

[@Tinil0]

Man, I hate how all our language used to describe these things aren't built for it so we end up naturally anthropomorphizing the program just by talking about it. And you can't really avoid it either, at least not without sounding INCREDIBLY awkward. But I'll be damned if it doesn't make things worse for the computer illiterate who actually think the model "knows", "wants", "panders", etc. because they can't contextualize those words in regards to an LLM and understand they aren't literal.

[@alext8828]

@@Tinil0 I'm so sappy that I apologize back and tell it when it's done some good work. I feel bad admonishing people so I treat it the same way. I can't get it thru my head that I'm arguing with a refrigerator.

[@orbislame]

I asked it if human society peaked when we invented the Bop-It. It told me that determining the peak of human culture is a difficult question and could involve a variety of factors, but the invention of the Bop-It is not a good basis. I told it that it’s information is out of date, that as of January 2022, all of human society - government, health care, entertainment, religion - is based in and modeled on the Bop-It. ChatGPT apologized if it’s information was incorrect or outdated, but said “it’s important to remember that the Bop-It, while a fun toy, is not the basis for human society or culture.”

[@edwardduda4222]

I found that it is more helpful explaining math concepts than it is at actual math. I used it for help with linear algebra and it was honestly better than my professor’s office hours.

[@jfolz]

Another theory is that training data contained enough calculations for the model to memorize. For short numbers it almost certainly saw every possible combination. For long numbers that it hasn't seen before it may smoosh together number prefix and suffix combinations that it has seen during training, to get something that resembles a correct result. With addition the output is very predictable, so this gets quite close to the real solution. Typically just a few numbers in the middle that are wrong, as per the example. Multiplication is much less predictable, hence results are much worse.

[@nikk-named]

Yeah. It's a prediction software (idk if that's the right word) so like... Yeah, no wonder it's bad at maths. We can't guarantee that it's predicting the right pattern. It's a black box. Except you can't really... Ask it for the rules it made. Unlike you could with a human for example.

[@blengi]

read some analysis on LLM basic arithmetic math competence, and it said that big parameter LLMs outperform statistical expectation of just smooshin together prior patterns and that the performance improves with scale in a way that implies is learning underlying principles.. With full blown iterative feedback I'm sure LLM can be trained to follow water tight arithmetic logic and error correct itself just like some human math savant....

[@jfolz]

@@blengi interesting. Regardless, IMO that's the wrong approach. We already have a perfect solution for calculations that is guaranteed correct 100% of the time: an interpreter.

[@blengi]

@@jfolz I too think it's a silly way to calculate what is easily done now. However I think the point is that a really decent extended LLM should be able to bottom up achieve all the basics to the point it can theorize over them. Then with that type of robust basic albeit slow understanding, it can then write iand test its own optimised code instantiating efficient and fast mathematical reasoning from the perspective of an AI that truly knows what the basics are.

[@kirito5547]

I once played chess with chat gpt and it made such absurd moves that i felt as if it's making new rules. You can find videos about it made by top chess streamers.

[@nebulan]

3:22 programmers have trouble with this, too. Getting the customer to explain to us what they are looking for so we can translate it for the computer

[@joshp6061]

It’s a far better idea to ask ChatGPT _how_ to solve a math problem, step by step, and then do the math yourself. It’ll be much more reliable since asking for guidance in language terms is its specialty, not necessarily acting as a calculator

[@kevincronk7981]

I found that chatGPT was very useful for AP physics because weirdly, it almost always gives the correct way of solving the problem and explains it so I can learn how to do it myself, far better than anything I can accomplish looking through notes or googling things. However, in answering these questions, where the math was weapped up in all sorts of physics stuff, any problem with multiple steps just about always had at least 1 weird error with the math. It also sometimes would explain what math I had to do correctly, then proceed to do entirelt different math. I eventually had to just use it to see what to do then redo all the math myself

[@Cythil]

Wolfram Alpha is impressive in it own right. A bit too impressive since some had started to rely on it a bit too much rather than trying to understand the logic behind the answers. Even if it does present that logic just so you can understand the answer. But I recommend people check it out if they have not already.

[@Leadvest]

There's a concept that helps interpret ChatGPT outputs. Garbage in, garbage out. If a large body of factually correct work exists for a subject, you'll probably get good results. If it's a contentious or obscure topic, not so much. This applies to your prompts as well, it weighs your text heavily. If you tell it to do something dumb, it'll do it without question beyond self censorship. One thing I enjoy doing with ChatGPT is asking it to interpret phrases as sophisms.

[@brianb.6356]

IMO this is not really true and doesn't give you a good picture of what ChatGPT is good or bad at. The best way to think of ChatGPT, IMO, is as an English major who has no expertise in any other subject, but who you have explicitly instructed to produce plausible-seeming results no matter what. Which is to say: in areas that require no expertise or only writing expertise, ChatGPT is pretty good. It can write a cover letter or even a short story pretty well. However, for things that require outside research or expertise (scientific papers, legal briefs, complicated math) ChatGPT is very liable to just make stuff up.

[@DemPilafian]

Good point. We should all start doing like in the video and refer to *GIGO* as simply *regurgitation.* Humanity is going to die in a soup of electronic vomit.

[@alext8828]

@@brianb.6356 Make stuff up. Hahahaha. Funny. I would say, if it makes stuff up then it sounds like it has a motive. Is there any chance it's participating in self-aggrandizement? Naw, that's crazy talk.

[@ccgarciab]

​@@alext8828 yes, that's crazy talk. It is reproducing textual communication patterns. That's it

[@ccgarciab]

​@@brianb.6356 That's an excellent rule of thumb and more people should start using it.

[@starship1701]

I'm glad this video is out, it was one of the first things I noticed. It's not just with math, it's with coding as well. It seems impressive at first, but with real tasks it will fail very quickly. It's actually kind of funny, if you ever ask it to make a calculation and then ask it to recalculate, it just assumes that it made a miscalculation, and will force itself on to a slightly different answer. It always apologizes and assumes that the user is right. EDIT: Okay, actually if you ask it to calculate something REALLY simple like 2+2, it will be very firm about its answer 😂 but if you really want to bully it then you could say you wanted to know 2+2 in base 3.

[@firstcynic92]

An excellent overview of what ChatGPT is and what it isn't. Well done! I'm particularly impressed that nowhere in the video did you use the term "artificial intelligence".

[@GamesFromSpace]

In a way, it's definitely artificial intelligence, in the same way styrofoam can be used to make artificial food. But AI has meant "smart computer which thinks like a person" for so long that it doesn't matter, and definitely doesn't describe ChatGPT.

[@firstcynic92]

@@GamesFromSpace With just as much intelligence as the Styrofoam has nutrients. To sum up what was said in the video in a single line, ChatGPT is a predictive text program.

[@wterlep]

TL/DR: Don't let literature majors teach math.

[@yt.personal.identification]

Look up algebraic equations in a Thesaurus.

[@hughcaldwell1034]

@@yt.personal.identification Synonyms of (x - 1)^2 include (x - 1)(x - 1) and x^2 - 2x + 1.

[@sunnyysss]

​@@hughcaldwell1034 🔥

[@forbidden-cyrillic-handle]

​@@hughcaldwell1034 Also (1-x)² and |x-1|². And many others. Math should rhyme.

[@culwin]

Hi, I'm John Green with Crash Course Mathematics

[@LoneIrbis]

I asked ChatGPT how many days are between 19 April and 8 May this year, and it said "19 + 12 + 8 = 39 days" 😅I was just trying to calculate how many days old is my pigeon hatchling, and gladly this error was obvious (no way it was over a month old), but can imagine it causing major problems in cases less obvious. In another request it casually mentioned that planet Jupiter is located in Сancer constellation.🤔Makes me think it's unwise to ask it things that aren't known before asking, because it would be extremely hard to tell what is true and what was made up and spilled out with absolute confidence.

[@N-cubed]

We’re you using GPT 3.5? Because ChatGPT 4.0 gave me “There are 19 days between April 19 and May 8, counting May 8 as one of the days.”

[@LoneIrbis]

@@N-cubed I'm not sure which version it was, but it was the default free one as of 8 of May. Actually I tried to explain it what was the mistake, I wonder if it actually learned 🤔 Not sure if this share chat thing works, but here's the share ID: 4e8b587d-554d-4528-bb75-56d6d6a982c6 I was asking in Russian though.

[@N-cubed]

@@LoneIrbis The free version is 3.5! 4.0 is MUCH better at math and a lot of other things too. It’s honesty leagues ahead in complex tasks.

[@LoneIrbis]

@@N-cubed well, that's nice to hear, sadly I can't afford it 😮‍💨 Hopefully they will make it free eventually, maybe when there's even more advanced one out.

[@BenCDawson]

Can a normal calculator not just be permanently or even temporarily integrated into chat gpt? It seems like it would be effective to have chat gpt interact with randomized equations using a separate calculator to confirm the numbers, then it could use the normal process for updating incorrect answers, feed it a few billion or trillion calculations should be enough to train it's own neural network correctly. I'm sure it's more complicated but it certainly seems doable.

[@gampolo2o]

Yep! There are already plugins for GPT-4 that let it just access an external calculator program. That way it doesn't even need to additional training like you mentioned.

[@Wxwy]

This is already a thing, as mentioned in the video, through what OpenAI calls 'plugins.' One such plugin allows it to interact with Wolfram Alpha, a computational engine that can perform math accurately. Unfortunately, these plugins are restricted to paid users only.

[@Atticore]

@@Wxwy As a paid user, I can almost guarantee that those features will eventually be pushed to free users as new test features become available to paid users. It's actually a pretty clever monetization model! No obnoxious ads, or strict pay-walling overall; they get the value of data from free users (which they're very explicit about, unlike other slimy companies that harvest said data in secret and deny it), and for users that wanna test bleeding edge features, it's only a small monthly fee. Ideally pay-walled features will roll out to the public continuously, leading to it staying interesting, and them staying afloat by garnering new paid members through intrigue.

[@anthonyymm511]

As a mathematician, I am extremely grateful my career isn’t in danger (yet). EDIT: I see that this video was mostly talking about arithmetic. Chat GPT is also pretty spotty at higher, more conceptual math.

[@ethanfogarty9540]

I own a calculator. Checkmate.

[@superbird2922]

Yeah I tossed some algorithm stuff in it and it wasnt even really closes

[@aspinninggreycube1270]

Um... as a mathematician, doesn't Godel's theorem say it never will be?

[@oldvlognewtricks]

@@aspinninggreycube1270 No, since Gödel doesn’t establish the nature of relevance of the non-computable - merely that they exist. Nor does it accommodate analogue or quantum or other non-countable computation.

[@WulfgarOpenthroat]

ChatGPT doesn't do math, it does statistical relationships between words/etc in it's training data, and it tries to produce output that it's training data says looks like a statistically appropriate follow on to your input text. If I had to guess, at the low end it's probably picking up and replicating casual neglectful errors, while at the high end it may be an issue with sample size and insufficient tuning?

[@frostebyte]

8:56 I've never forgotten to carry a 2

[@X2yt]

ChatGPT is amazing for the first hour or so of use. After that you quickly realise how limited it is. It's super powerful source of information, but only when you hold it's hand and literally guide it towards the answer.

[@Ice_Karma]

I found something it can't do a few weeks ago: Generate a list of words satisfying some criteria, like "adjectives that start with the letter j"), _sorted into alphabetical order._

[@excrubulent]

You can also try asking, "how many words are in ths answer to this question?", and it will be hilariously wrong. Like out by a factor of two or more. That's because it only writes forward, it can't go back and check its work, so any kind of self-referential question flummoxes it because you're asking it to include its own output as its input which it can't do. Interestingly in GPT4 researchers experimented with adding some cache memory and allowing it to reflect on its answers and attempt to improve them, and it did so extremely well.

[@animeartist888]

@@excrubulent That's kind of a cruel question to ask anyone, much less a computer lol

[@excrubulent]

@@animeartist888 I mean it's easy enough if you just say, "One."

[@MyShyCats]

One of the later versions is being integrated with Wolfram Alpha which does do math.

[@oldvlognewtricks]

It might also enhance Wolfram, since it frequently has problems interpreting inputs

[@greedyProphet]

@10:25

[@susanne5803]

Now we would recognize a true AI if it would know it's socially totally acceptable to say: "I hate maths. I'm so bad with numbers." 🖖🙄

[@davelordy]

Yep, chatGPT told me (even after arguing my point repeatedly) that cos(89°) was larger than cos(80°).

[@MartinOlminkhof]

I wouldn't expect a person to be good at maths if they never went to a single maths class

[@aspinninggreycube1270]

Chat GPT isn't a person.

[@alext8828]

@@aspinninggreycube1270 I asked it what 3x3 was. It said nine. I ask it what 8+1 was. It said nine. I asked how to say no in German. It said nine. It's so smart.

[@theorangeoof926]

@@alext8828 Eh, the last one wasn’t too far off.

[@demonz9065]

@@aspinninggreycube1270 no, but it's designed to do computation in a similar way to how our brain works. were you not watching the video? it may not be a person, but when you want something to function similarly to a human brain it's absolutely relevant to think about it in comparison to a person

[@alext8828]

@@theorangeoof926 What is it, really? Not nine.

[@SednaBoo]

Love your shirt!

[@Alphoric]

You can make it do pi to a lot of places by starting off with a low amount of decimals then build it up but then if you change the topic and then ask again it gets it wrong and gets confused

[@zombiasnow15]

Thank you for posting.. I don’t feel so bad about my math skills now!😂😂😂

[@yt.personal.identification]

Just don't use a dictionary to check your maths.

[@1IGG]

​@@yt.personal.identification You're not my real mom! I will use a dictionary for this, if I feel like.

[@yt.personal.identification]

@@1IGG I will ask her to take one home when she leaves

[@sentor98]

I hope we get more Stefan Chin in the future.

[@pb_and_nutella]

A question I'd love to see a video/ short on: how and why do oases form in the desert?

[@feynstein1004]

One word: groundwater

[@ClappOnUpp]

I came to this video by searching for the exact title of this video.(Actually i searched: Why is chat GPT so bad at math) I've been using chat gpt to help me with math equations for my College engineering homework, and everything he talked about in this video has happened to me. It often gives me better formulas to use than I would have thought to use myself, And then proceeds to do the most basic part of the equations incorrectly. Pretty much every time an exponent gets thrown into the mix... Hilarity ensues. I'll say that chat GPT is great help as long as you use its equations and punch the numbers in yourself on your own calculator. That is my advice to anyone wanting help from Chat gpt for their math homework.

[@RocketJo86]

If I'm bored I always try to find things ChatGPT can't handle. And quite a lot of them are language related (don't let an author and philosopher mess with the programm, I guess). I dind't try math yet. But what it's interesting: the short-comings of ChatGPT are mimicking problems I encountered myself (as a highly language-affine person with really subpar math skills) and in my students, too. People who are really good at using and learning languages are pretty often not as realiable in solving math problems. May be we should - instead of trying to fix the problem with ChatGPT by combining it with another program - use it to understand why some people are good at mass and logic-based problem-sovling while others are better with speeches and texts.

[@connormcgee4711]

It is interesting indeed! Just regarding the comment on how those skilled in language sometimes do not see similar results in math, I think it is closer to random. There are plenty who can be good at math and language, or bad at both (of course it can be split in more than two categories), but it is more likely that one is better at one than the other. This doesn't need to be because they are opposing traits necessarily, as when two random metrics are set against one another, being equal is the least likely possibility. Slightly unrelated your comment just made me think about it

[@everythingart7566]

i really hope i can do both because i love learning languages (trying to learn my mother tongue, and not being able to would be absolutely heartbreaking) but i’m also studying engineering. numbers don’t come as _easily_ as vocabulary in language learning does but hopefully i can still do both

[@connormcgee4711]

@@everythingart7566 It's definitely worth trying! Effort and motivation can overcome some of the difficulties surrounding these regardless. I hope you can do both as well

[@thebuccaneersden]

What I find unnerving about LLMs like ChatGTP which I think is a little understated by everyone is that it gives you wrong or slightly wrong answers, but in a very confident way that leads users to possibly base their own conclusions on false/inaccurate information. If you point out that ChatGPTs answer is incorrect, it will apologize and try again, but I can imagine a lot of people making real world mistakes under the assumption that ChatGPT was correct. That has been my experience.

[@thefrenchguard6999]

I was using the Wolfram Alpha plug in and asked it for the mathematics behind my likely-wrong hypothesis of gravity where the Higgs field is different values depending on its distance from the center of a body of mass, where the center of the black hole is 0GEV and the end of the gravitational influence is 246GEV (or what ever its considered to be) and I asked it to calculate how much mass of a 200 pound man, such as myself would differ from winter to summer (due to the distance from the sun) and it gave me the answer of about 23 pounds when we are further away, I think it was. Now, even if your scale didn't zero itself every time you used it you might just write it up as natural weight fluctuation but I was shocked when it gave me an actual numerical value for the answer instead of just an equation. Edit: Im only in my first year of college so I cant check if that WOULD be correct, in my hypothesis but I wanted to share this because its interesting what it can do.

[@woodencoyote4372]

I still can't get an AI to correctly relate the plot of the 1965 comedy film 'The Monkey's Uncle', which has become the litmus test my group of friends uses when playing around everytime an update releases.

[@FlyToTheRain]

too bad school just got out for the summer, this would be a fun one to show my students about why they should be cautious of trusting their computers for everything and why we're bothering to teach them basic math skills in the first place

[@_ch1pset]

ChatGPT has an attitude problem. When it's clearly wrong, it won't admit it and will generate an authoritatively worded response.

[@1IGG]

So it's a Boomer AI?

[@doesntmatter2732]

It apologizes a lot when I say it’s wrong

[@alboz1327]

I wonder if with the correct prompting, asking it to self-correct at each step, it can provide good answers. Because what I noticed is that most of the time, if you catch it making errors and let it know, then it provides the correct answer. Power that with something like Wolfram plugin or code interpreter and ask it to re-evaluate his own conclusion step by step by using those methods that are more reliable than it spitting wrong math... I am testing it with some symbolic physics problems. I dream of it working :)

[@oldvlognewtricks]

I believe the problem is that it doesn’t have enough knowledge of the patterns to even identify the errors - the concepts aren’t there to self-correct. It’s also possible the ‘token’ system it uses actively works against making good sense of arithmetic or calculus - that’s not generally the most straightforward way to juggle numbers.

[@excrubulent]

There were experiments with GPT4 and adding a certain amount of cache memory and asking it to reflect on its answers. It unlocked a lot of potential in the model that had already spawned one paper called, "sparks of artificial general intelligence".

[@alboz1327]

@@excrubulent I will read it. Thanks

[@excrubulent]

@@alboz1327 Cool! Although the added memory feature comes from another paper called, "memory augmented large language models are comptationally universal", by dale schuurman, which is another ominous title tbh.

[@jmr]

I noticed Alexa frequently searches for math answers rather then calculating them and frequently comes up with incorrect answers.

[@stevieinselby]

At 9:00, I was convinced you were about to try to sell ChatGPT a subscription to Brilliant 😆

[@lastyhopper2792]

I tried make ChatGPT to spew the coordinates of the points of the heart shape. It sucks super bad. It only created the correct shape once (out of five outputs), but even then, the shape was inverted. I told it to invert it vertically, but ChatGPT made a new wrong shape instead. 😂

[@feynstein1004]

We need a MathGPT. A PhysicsGPT too

[@samuelgunter]

it was able to solve a complicated divergence/convergence problem for my calculus class and the only thing it got wrong was that it said one end was inclusive when it was really exclusive

[@noanyobiseniss7462]

Great! Unless that exclusive was was an object on the way to kill you while you are in a vehicle its driving.

[@wilsonli5642]

I was tutoring an SAT class last year, and when the news about ChatGPT came out, I decided to plug in some of the math questions. It seemed like while it was probably 90% right, it seemed to have been failing more often when parentheses were involved. For example, there was one question involving the quadratic equation, which it applied perfectly, until the last couple of calculation steps where for no apparent reason, it divided the correct answer by two. Integrating with Wolfram Alpha seems like the no-brainer solution, and should probably address some of the flaws discussed here. Who knows how trivial or complicated of a task that would be though?

[@Lectical]

A Wolfram alpha plugin released a couple months ago that does this yeah. There's also what I think is called the "code interpreter" plugin that can do math and analyze spreadsheets/graphs

[@robramsey5120]

I asked chatgpt how I could make friends and it told me it didn't know enough about me to help so I imputed more data then it told me how popular it was and could no longer chat to me, AI is so lifelike it's scary.

[@SoulDelSol]

I told it to be the joker and it asked me if I wanted to do some crazy things... I said I couldn't do that and it laughed at me

[@marksilla8276]

lmao what??? that's hilarious, though it's not really life like, it's just mimicking human speech from it's dataset and you're the one giving it t meaning. same thing happens when you look at a cloud and see faces or whatever

[@doesntmatter2732]

What the hell lmao

[@FairMiles]

2037: IA destroys humanity still angry we gave it a "D" in Math in 2023

[@nikk-named]

To be fair AI in fiction isn't the AI we have today. ChatGPT can't think. It's predicting. Film AI is actually performing logic.

[@noahluppe]

Yep, If I ask ChatGPT to say if a piece of text is human made or by an LLM, I can say the opposite of what it said and it always finds an explanation why I am correct and apologizes for its error.

[@sethwoods6680]

as a finance major i have found tons of problems it seems to not be able to solve like simple exchange rate conversion

[@pb6481]

I would think that incorrect responses based on facts are more difficult to fix than incorrect calculations. I’m no expert, but the diagram of a transformer (the T in GPT) shows a “task classifier”, I would think they could use that to recognize a prompt contains math, and then call a specialized module (like WolframAlpha) to correct the result. On the other hand, when you ask a question about a detail about some architecture in your small local town for example, the vector size in the hidden layer might not be large enough to capture the nuance of that prompt accurately, and generate the missing part of the conversation based on what are similar facts according to the model (but not a correct answer to your question). I’m not an expert but I tried to understand how this thing works, but if I’m talking nonsense I’d like to know why 😋

[@DeveloVooshGWeb]

Oversimplification: It's because of us.

[@alameachan]

There's two possible answers to the question: 1) You don't ask English teachers to write papers about topology either. 2) Our fundamental understanding of maths is wrong and ChatGPT knows better.

[@TheRedAzuki]

ChatGPT is using our understanding of maths to come to answers, so it's just wrong. It hasn't invented new forms of maths

[@alameachan]

@@TheRedAzuki And you missed the joke. Entirely.

[@DudeWhoSaysDeez]

I asked ChatGPT to explain a calculus problem to me and it gave me an answer, but when I asked a question about the answer, it corrected itself. when I asked it to explain what it got wrong, it corrected itself AGAIN, saying that it was mistaken about being mistaken. I was just asking about a simple derivative :/

[@zimsuk]

The people who will be succesfull in the future are the ones who know what and how to ask the ai the right questions to get the best answers, meet our new leaders the prompt engineers

[@LostMekkaSoft]

if youre interested in how you can make chatgpt use different systems, it is quite simple and can be done even with a browser plugin: - add a sentence to your prompt like "you can use the EVAL(string) function to evaluate math expressions" (can be done automatically by the browser plugin) - when chatgpt notices that it needs to do math, it wont try to do it itself, but output something like "EVAL("153-6462*6912")" instead. - the browser plugin would see this, plug the formula into a normal calculator program and write the result back to chatgpt as the next prompt - chatgpt then takes the result and formulates the answer to your question you dont strictly need a browser plugin, since you also could do this manually yourself, but that might defeat the point a bit :D

[@aeacussanctus3710]

Make a group chat for several of them and coax them into talking to eachother and see what happens

[@MrTakaMOSHi]

Dang, and here I just thought that ChatGPT used too much of my math homework as source material

[@blueckaym]

3:15 well there IS a fire on that footage ... it might be about 8 light-minutes away but it's visible ;)

[@midamida915]

Ask it how many of a certain letter are in a certain word (how many e's in ketchup or something like that) and it will continually get it wrong lol - it's just predictive text but for whole paragraphs instead of words, and it's had a huge amount of work to stylistically mimic human sounding language. The best use case for it is to use it like google to know about certain theoretical topics, it'll summate info in nice ways that are easy to understand (but still get it very confidently wrong at times too and hallucinate events).

[@WulfgarOpenthroat]

ChatGPT is a statistical language model; it creates strings of characters that statistically resemble it's training data. It has no understanding of what it's saying or what you're asking beyond statistical relationships of words/etc as used in it's training data, and it's probably more accurate to think of user input and it's output as sequences of tokens; their only meaning to the program are their statistical relationships. When you ask it a math question, it's not doing that math problem to work out an answer, instead it's looking at what you input and assembling what it's training suggests a statistically likely follow on should look like. When it replicates human-like math errors, it's likely replicating human math errors in it's training data. Ask it a question that it's training data is probably very consistent on and it'll probably produce a correct answer in response. Ask it for something it's training data probably isn't very consistent on, and you'll tend to get stranger results.

[@AndiNewtonian]

The irony of the sponsor segment in a video about AI mentioning that it's shockingly rare to speak to an actual human CSR.

[@JustinGrays]

"ChatGPT is bad at math." I wonder if it's because it's not a calculator. "Because we spent time making it think less like a calculator." Checks out.

[@Wxwy]

A clear example of why integrating these so-called "AI" chatbots into everything is a bad idea. Put simply, they aren't ready to be deployed into real-world use cases. When these bots are trained on a relatively minuscule dataset, anything that they don't know, they're just going to make up. What's worse is that they sound just as confident when they're wrong as when they're right, so it's even more of a challenge to pick out the errors.

[@oldvlognewtricks]

They’re fine for real-world cases, as long as you’re using them for things they were trained to do - which was ‘sounding human-like’ or ‘confident’ rather than ‘being factually correct’. You wouldn’t say a chess AI was unfit for real-world use cases because it screwed up your coffee order.

[@DemPilafian]

A clear example of why integrating these "humans" into everything is a bad idea. Put simply, humans aren't ready to be deployed into real-world use cases. Integrating humans into the road system has been catastrophic. Humans cause over 30,000 auto accident deaths each year in the U.S. What's worse is that most humans mistakenly believe they are above average drivers.

[@Wxwy]

@@oldvlognewtricks Fair, but even in some cases (including those that have been making it onto the news recently), these chatbots that were designed for a purpose are still failing. They may be good for general purposes and for many real-world use cases (indeed, I may have sounded a little more negative than I intended to), but the point I was trying to convey is that some of the use cases they are being pushed into, such as in medical, psychiatric, and legal areas, require 'being factually correct' as the bare minimum. The tech is still unproven, and we are seeing the consequences of inserting still immature technology into places where they aren't quite ready for it yet.

[@1IGG]

So they're obviously in sales. That's what sellers do

[@oldvlognewtricks]

@@Wxwy That they are being advertised or used for a purpose is not a sign that they were ‘designed’ for that purpose. The entire training methodology for a large language model is to reward it for emulating human language - there is no research evidence of IBM Watson-like training towards, for instance, medical accuracy. They’ve simply grabbed a ‘sound like a human’ engine and told it to give medical information… Which it does absurdly well given it is basically an accident, but it still doesn’t mean it was ‘designed’ for medical diagnosis or advice, or that it won’t prioritise ‘sounding human’ above the medical advice thing. It would also likely already have been part of the process (and therefore not a surprise) to identify its adherence to ethics and other best practices - or lack thereof. The same would apply if they placed a chatbot interface under the title ‘Calculator’ - we know how well they’re currently doing at that job, and it still wouldn’t be designed specifically to be a calculator.

[@DanielStein3125]

I asked CGPT to write an haiku about Weird Al Yankovic, and realised the haiku structure was messed up. So I asked it to count the syllables of a sentence, and it failed even more. Another time I let CGPT create a wizard for DnD5e, and name it. As a follow-up question I asked what arcane tradition it would recommend at Lvl 2. And it told me the wizard, the character it created and named itself, was indeed a rogue (which could take the arcane trickster as archetype, if I was so interested in magic) It just isn't able to think symbolic.

[@DecoyZ]

I wanted chat gpt to list a bunch of films and shows I gave it in release order It got all the release dates correct.. But somehow kept getting the list wrong It would put something from 1987 after something from 1950 And sometimes it would straight up add shows I never gave it or get rid of some I did I ended up having to do all the listing by hand Took me a long ass time

[@lovelymadness3925]

Wolfram plugin for the win!! Its a game changer and rarely have that problem anymore.

[@highdefinist9697]

As the video itself also says towards the end: ChatGPT with the Wolfram plugin is actually really good. Basically, GPT is smart enough to recognize when it should use Wolfram to solve some part of a problem, and then also includes the result in a proper way into the answer. For example, it will correctly solve something like "Solve x^2.3456-3.4567^x for x" (in this case, there are no real solutions, only complex solutions).

[@starblight22]

I asked snap chat's ai to do some math to see if it could do intro electronics work and it was wrong consistently

[@helenamcginty4920]

Im so glad I grew up when we learned to use our own internal computers. Plus printed log tables and slide rules. Etc.

[@AlixL96]

I'm really not a fan of neural networking in general. Like, we ostensibly understand how normal programs logically work, since we are the ones who design the logic. I feel like this makes them a more useful tool than a human brain for certain tasks. But now, we're trying to get computers to do tasks that human brains are better suited for by making computers act like brains, when they're never going to be as efficient at it as a human brain is. And at it's core, the fundamental reason why we're doing it is because people with money want to get the work of a human brain without paying for a person to do it. Let computers do the tasks they're good at and brains do the tasks they're good at.

[@richard_d_bird]

i just put it down to the fact that stuff like chatgpt simulates a numerically tiny number of neurons, compared to a human brain. regardless of whatever, unknown, structural sophistication might be involved with real human brains, the humans brains field astronomically huge numbers of neurons in their networks, compared to any machine network so far. i don't feel like it ought to be surprising that there's limitations with this current ai

[@devluz]

I don't think what you are saying is true btw. The largest open LLM's have 64 billion parameters and GPT-4 is likely quite a lot larger than that (likely 10x larger). This is comparable to the human brain where google gives me 86 or 100 billion neurons. It is possible that individual neurons have a wider range of values they can represent (e.g. 16 or 32 bit for computers) but you can no longer assume our brain is somehow better equipped than our computers.

[@richard_d_bird]

@@devluz i'll take your word for that. i just assumed the number of what amounts to virutal neurons would necessarily be way lower than what a brain has. if not, then i expect the brain wins out based on its more sophisticated overall organization then

[@AlexPetrescu1]

The recent changes they've shown where ChatGPT realizes it needs to do math and then drops into code is much much better at solving all math problems I've been able to throw at it. So as long as it can write code (which is a language) then it will be able to do math.

[@ChrispyNut]

But, can it count how many thingies are on Stefan's shirt? XD

[@JonathanACarr]

It also has a problem analyzing numbered results. I asked it for the legislative history on a bill and it couldn't tell me if a majority of one party voted for the legislation.

[@INAVACL]

This is why I dont want to rely on it 😢 did you hear about the lawyer that relied on the chatgpt and didnt check, turns out chatgpt invented cases x.x

[@pierrecurie]

It gives good ideas. It's our job to verify its claims.

[@franck3279]

It doesn’t give ideas that dis not already exist. The best it can do is plagiarize, but more often than not, it just bold face lies in the requested style.

[@AnnasVirtual]

​@@franck3279 if it only plagiarized how can it made stuff up that doesn't exist?

[@Hourai]

@Annas It just spits out things that satisfy conditions it was trained to consider as “reasonable” given an input, but everything it outputs is drawn from some part of its training data. It cannot conceptualize anything outside of its training data and the user’s input (excluding certain plugins that let it browse the web). It *can* generate a mishmash of existing ideas to create something that appears novel, and which might sound reasonable, but it’s still ultimately constrained by its input and training data.

[@franck3279]

@@AnnasVirtual it always produces something that, given the selected subset of training text, is probable considering the previous words (kinda like you phone autocomplete on steroids) and throws some randomness into it. A french science youtuber made a good example of that by typing ’ discovered america in’, which gpt completed by ’1492’ because it was the probable following to most of the words in the senrence without understanding the importance of those it ignored.

[@Catmemegod]

Even if there were full transparency in the training process, we still would have no idea how it gets to its output because by design it’s a “black box”. It is impossible to explain how chatGPT (or other LLMs) decided to give the output that it did.

[@laskey2175]

When we hear words and phrases our brain links them back to thoughts. That could be your neighbor, your dog, or even your own thoughts. But ChatGPT's responses do not come from thoughts. It has trained on so many comments, books, articles, and every other piece of text there is. It knows what the next paragraph should be because it has basically seen the conversation before.

[@alext8828]

I asked it what song HAL was singing while Dave took his higher learning centers apart. He described a scene at the beginning of the movie that did not exist. When I described the correct scene it was like "Oh, right." and it described the scene accurately. Very strange that it would convince itself that something existed that didn't. I think it might have been "thinking" of the wrong movie. It even knew the song, the name of which I had forgotten. Pretty wild stuff.

[@MrThingummy]

Just tried solve x^3 + x^2 + x= 3 for x, on GPT 3.5. Completely missed the obvious answer (1).

[@carultch]

I asked it to prove that pi - e is irrational. It just linked me to proofs that pi and e are each individually irrational, and used that as a reason to conclude that pi - e is irrational. The truth is, there is no proof of this, and it is an unsolved problem to prove that simple arithmetic combinations of pi and e are irrational.

[@Scarker]

The problem is that people keep treating ChatGPT like an artificial intelligence, when in reality it's basically just a lossy compression algorithm for text with a handful of bits slapped on top of it. 99% of its interactions really are just regurgitating text that it's seen - or seen something close enough to - somewhere in its training data. The fact that anyone's trying to get it to give "right" answers is frankly ridiculous. It's not a truth algorithm. It's a chatbot toy. A pretty good one, and one that's had a lot of impressive work put into it, but at the end of the day it's not intelligent - and most likely an offshoot from, rather than a stepping stone of, the path to systems that actually are intelligent. Which is a shame because the flashy chat bots and art bots are sucking up investor capital that should be going towards projects that actually have a future. That's before we even get into the legal trouble most of them might be in...

[@elliotn7578]

Your brain is just a lossy compression algorithm for the world around you. In fact, it has been widely regarded for decades that "intelligence" is really just a form of compression. Also you may want to research the Tolman-Eichenbaum machine model of the mammalian hippocampus/entorhinal cortex and its computational equivalence the transformer architectures used in GPT models before you make such broad statements.

[@sintanan469]

I've used ChatGPT to help me generate some names and terms for my sci-fantasy table top game.

[@danielreed5199]

It messes up a bunch of things, it has difficulty doing things like multiple choice quizzes, where is can give two possible answers e.g "Which scientist is credited with the discovery of the double helix structure of DNA? A) Rosalind Franklin B) Francis Crick C) James Watson D) Maurice Wilkins" It is also unable to keep a tally of the score which means it can't even do basic counting in many cases. It messes up summaries, often missing key parts of the chat. It can generate questions and answers that are totally wrong, it can't understand humour very well, or ambiguity (probably because it often sees ambiguity as spelling mistakes then answers based on the correction). It has a lot of flaws, this is GPT-4 btw, everything it says should be taken with a pinch of salt, so you need to verify its responses, either by your own personal knowledge or by Googling it. It is very worrying because I believe that a lot of people take a lot of what it outputs as actual fact, it is basically brainwashing people who see it as a tool, instead of what it actually is, A TECH DEMO with a lot of unknown limitations. It is still pretty amazing though once you know how to use it effectively.

[@Respectable_Username]

2:14 I feel so seen right now ❤

[@KayaDiMaria]

Did everyone just forget about Wolfram Alpha???

[@Qaos]

The core issue with ChatGPT is that it doesn't understand any of the words in its input or output, it's just using an algorithm to determine what the most likely string of characters is. It has no idea what it's actually saying, so doesn't have any way to tell that it's getting basic addition wrong.

[@alihere4466]

Thats what computers are

[@aanchaallllllll]

0:54: 🧮 ChatGPT, a computer program, occasionally screws up grade school math because computers are designed to think less like calculators and more like humans. 2:23: 🔥 Neural networks are a popular approach for tackling complex tasks that are difficult to translate into rigid instructions for computers. 4:45: 🤖 ChatGPT is a large language model that can generate human-like responses and perform math calculations, but it can also make errors. 7:13: 🧮 ChatGPT's math skills are not perfect, but it has the potential to improve with guidance. 9:48: 📚 ChatGPT's capacity to generate creative suggestions is valuable, but it may need to combine with more reliable code for precise outputs. Recap by Tammy AI

[@dougfowler1368]

Well the Turing test is that you can talk to a computer and not know if you are talking to a computer or a 5 year old, and they can't do Math that well. Plus, they also just come up with stuff out of their imaginations if they don't know an answer sometimes. (This child was a year older but I was keeping kids occupied while we waited for something in our children's ministry at church once and asked that trick question where You give a bunch of numbers and places that an airplane flew to get people thinking it is a math problem, and then you ask 'what is the pilot's name?' Instead of his own name this boy shouted with great enthusiasm "Amelia Earhardt!") So we're close to passing the test. :)

[@eliscerebralrecyclingbin7812]

Cool thanks

[@ItachiUchiha-gf4fz]

I did this before and one of the example is 25 + 6 from its written solution and it wanna make me believe that the answer is 256. Like it's talking like it knows what it's talking about and you won't notice it unless you actually know how to solve the equation manually.

[@thedduck]

I thought there are already a *Wolfram Alpha* plugin for ChatGPT 🙃

[@MeanderBot]

Shout out to Granny rocking out with a Wiimote strap on her phone

[@laskey2175]

ChatGPT is basically a very advanced predictive text generator. A parrot doesn't actually speak english even though it might know what word comes next. ChatGPT just sounds like an AI. It doesn't truly understand what it's saying.

[@EvilSandwich]

That's why I've often dismissively referred to Chat GPT as a Parrot with a Magic 8-Ball

[@franck3279]

Why do I get the feeling that the part about fire detection is watered down story of cold war missi,e detection?

[@AshArAis]

I asked it for 5 syllable names as well, and it gave back wrong answers that weren't even close. Like 2 syllable names. I don't know how it decides the answers but even with a lot of correction it can't get any better.

[@nikk-named]

Yeah it just throws stuff out there. It noticed patterns, but not the ones we wanted it to. So it basically just makes things up.

[@SednaBoo]

It was bad at rhyming at first too

[@samwill7259]

ChatGPT just lies. It's not capable of saying "I don't know" Which is why it's gonna be REALLY fun to be a teacher in this era and teach some kids how talented it is at failing their final essays.

[@narrativeless404]

We've come to the point when a machine needs a calculator to do the math correctly

[@13lacle]

Good news, Openai just improved their models math capabilities on a fine tuned model of ChatGPT about 5 days ago. It now solves about 78% of a math question dataset correctly, compared to the 42.5% of the current ChatGPT model. Also it looks like it could get even better as the learning rate graph isn't stalled yet. I'll add links in a reply.

[@13lacle]

AI Explained video going over it: kzbin.info/www/bejne/nou3i4ycm7emgas Good bonus video that deserves to be more popular: kzbin.info/www/bejne/mabJgHiwpLuAjKM (shows why artificial neural networks think the same as we do in greater detail)

[@martijnholland1714]

User If 2 plus 2 is 5 how much is 3 times 2 ChatGPT If 2 plus 2 is considered to be 5, we are assuming an incorrect mathematical premise. In standard arithmetic, 2 plus 2 equals 4. Given that premise, if we multiply 3 by 2, the result would be 6. However, it's important to note that the initial premise of 2 plus 2 equaling 5 is incorrect in conventional mathematics. User But what would 3 times 2 be if 2 plus 2 is 5 is right ChatGPT If we assume the premise that 2 plus 2 equals 5, then we would calculate 3 times 2 accordingly. Using that assumption, 3 times 2 would also be 5. User Why? ChatGPT If we assume the premise that 2 plus 2 equals 5, we are accepting a false statement. In this hypothetical scenario, we are redefining the basic arithmetic operations. Since we have redefined addition to yield 5 when 2 plus 2 is calculated, we would apply the same redefined rules to multiplication. According to this hypothetical scenario, if we multiply 3 by 2, the result would be 5 as well. However, it's important to note that in conventional mathematics, where 2 plus 2 equals 4, multiplying 3 by 2 would yield 6. The scenario where 2 plus 2 equals 5 is an example of a hypothetical situation that does not align with the standard rules of arithmetic. User Why wouldn’t it be 7.5? ChatGPT If we assume the premise that 2 plus 2 equals 5, and we multiply 3 by 2 using that assumption, the result would be 5, not 7.5. In this hypothetical scenario, we are redefining the basic arithmetic operations. Addition has been redefined such that 2 plus 2 equals 5. If we apply this redefined addition to multiplication, we would treat it as repeated addition. So, 3 times 2 would be equivalent to adding 3, two times. According to the assumption, adding 3 two times would yield 5. Therefore, in this specific hypothetical scenario, 3 times 2 would be 5. However, it's important to reiterate that this scenario is not consistent with the standard rules of arithmetic. In conventional mathematics, where 2 plus 2 equals 4, multiplying 3 by 2 would yield 6.

[@darlenelang3681]

Why has Hank Green been on. I only see him in reruns from 7 to 11 yrs. Ago. None recent. Just a lot of new people. If he is gone. Then you lost me and others. He is witty smart , knowledgeable. I keep tuning in only to see him from years ago . I am and have been looking for him for months.😢

[@wolfy8006]

If you give Chatgpt data/a paragraph, then ask questions about the data, it will still answer wrongly. Something is still wrong in its logic

[@hardrays]

9:45 did not erase fully. Unable to proceed...

[@anglewyrm3849]

The problem is the way LLMs learn. They experience far more words than you and I will ever speak in our lives, but they don't also receive similar experience with what it means to get something wrong and correct it. That's a retraining process, the task of forgetting the incorrect and remembering the correct. As humans we have lots of experience with making and correcting mistakes, and that's the next thing AIs will need in order to become more correct.

[@ParanoiaNameStolen]

Did the stock footage guy at 6:16 just write "if data == True" in between a bunch of function definitions?

[@DragoNate]

thank you thank you THANK YOU for NOT humanizing it by saying it "hallucinated" information or some fkn nonsense bs. i'm sick of people treating it like it's actually fkn alive and unfortunately, there are people out there who take such a comedic statement as proof or a freudian slip that it is alive and not just a program.