This sentence is out of nowhere, did ChatGPT write this article? How could someone think something so new is suddenly negatively impacting students across the board.
Yes, I know the information isn't always 100% accurate. Neither are the books, blogs, courses, etc., I pay for. They also don't let me ask "what about this scenario..." This gives me such a strong base to build from which I can then use other sources to verify.
Those are two completely different things. ChatGPT has no obligation to be correct, and isn't even trying! It is a chatbot. Its prime directive is only to sound believable. Those books, blogs, courses, etc are at least trying to be correct and provide you with reliable information on a topic.
Try using ChatGPT on a topic you know really well.
And read what it says carefully. On many occasions I've seen it say something in a subject I know well and it's started correct so I've mentally autocompleted and assumed it had the whole thing right.
Eg. This for "What is Gell-Mann amnesia?" - contrast to the actual meaning. [1]
> Gell-Mann amnesia is a term coined by science writer Michael Crichton to describe the phenomenon where people tend to forget information that contradicts their preconceptions or beliefs. The term is named after physicist Murray Gell-Mann, who won the 1969 Nobel Prize in Physics for his work on the theory of elementary particles.
> According to Crichton, Gell-Mann amnesia occurs when people encounter information that contradicts their beliefs, but rather than updating their beliefs to reflect the new information, they simply forget the information and continue to hold onto their preconceptions. This type of cognitive bias can be a significant barrier to learning and can lead to flawed decision-making.
Learning from ChatGPT as if it's similar to "books, blogs, courses, etc" really doesn't seem like a good idea.
[1] https://en.m.wikipedia.org/wiki/Michael_Crichton#GellMannAmn...
My bootcamp Discord feels useless in comparison. ChatGPT is always available, and while it's true sometimes gives beliable but wrong answers in my experience if you have basic knowledge on the topic it's easy to spot.
It saves me sooooo much time it's amazing. The time I spent correcting or figuring out ChatGPT mistakes is minuscule in comparison of the time I'd be spending scanning horrible documentation or testing stuff from Stack Overflow.
And if you feel it doesn't provide useful answers, then just and use your search engine.
And step by step solving of exercises is great for students too, they can verify with the textbook's answer key.
The key is having a source of truth on hand.
Like i think if you're aware that it can be confidently wrong then it can help you explore areas you don't know.
To think of it differently it felt like learning via comments on Reddit. Which is to say a large portion of Reddit posts are shockingly confidently wrong. But with ChatGPT you can inspect those "comments", ask from various angles, etc.
I have this feeling that ChatGPT could, even in its current form, be useful for learning the larger complex picture. Very bad at reciting facts definitely, but ideas maybe not so bad.
Either way i'm still wanting it to improve. But i still think it's shockingly impressive. I will be paying happily if they can make "small" improvements to how it understands information.
On topics I know really well, ChatGPT is wrong more often that courses, blogs, books, etc. However, I don't think it's prudent to put them in two different categories, where human books are reliable and LLM answers aren't. Both make many mistakes, with a difference in degree that currently favors human books, but without either really being in a different category.
Newly published books (let alone blogs) are frequently less reliable than even Wikipedia. They are written by a handful of authors at most, get a limited period of review, and the first edition is then unleashed upon unsuspecting students until the errata list grows long enough that a 4th edition is needed.
The prime directive for LLMs with RLHF is a combination of giving the answer that best completes the prompt, and giving the answer people want to hear. The prime directive for authors is a combination of selling a lot of books, not expending so much time and energy writing the book that it won't be profitable, and not making so many mistakes that it damages their reputation.
Neither books, blogs, nor ChatGPT have any obligation to be correct. Either way, the content being reinforced (whether through money, or through training) is not The Truth straight from the Platonic Realm, but whatever the readers consider themselves satisfied with.
And that's the difference! Human authors are incentivized to return reliable information. Reliability is not ChatGPT's concern at all, believability is. It can't even cite a source!
It's not even 25% accurate when you start asking anything other than "what is a lion" and "what is 2+2"
[1] Sourced the salary for teachers near me from Google's linked data. Sourced the Lowe's part time cashier range from an actual job listing in my area.
Edit: fixed footnote, sentence fragment.
Thankfully, in such a world, AI is available to the "teacher" side as well and can serve as a way to both check that the student is doing their job and also to answer their questions, like a personalized teacher of sorts.
I'd have a set of N questions that shouldn't take long to answer but demonstrate that the material has been read. Yes, ChatGPT can answer them.
The second part would be in class, after the homework had been handed in pull a name and number out of a hat and have that student answer that question again and answer a followup question on that material.
The oral part of it makes it so that having another do the homework for the student and not do the reading means that the student would show a lack of understanding of the question or its followup.
(note: my sophomore English teacher did this... and I messed up answering questions about the Miller's tale which I hadn't read)
And yes, the first thing she did was see if it could literally answer it for her. It did. However, once that was done with she proceeded to massively expand the scope of the original homework... I've rarely seen anyone her age as enthused with learning, they're usually already tired of school.
Yes, I was standing by to fact-check it, but it didn't actually make any mistakes. Which is a little unfortunate, I guess... I'd been planning to stress the importance of cross-checking. Oh well. It generally seems to get everything right at the level of questioning children are usually capable of.
(Half an hour later she was asking it why water expands when it freezes. Fifteen minutes after that we were reading a wikipedia article on molecular physics... I dare say this was a productive evening.)
---
I think ChatGPT is an incredible resource for learning, and trying to keep it out of schools is throwing the baby out with the bathwater. It does better at teaching than the teachers often manage.
I don’t think it’s bad, per se, but ChatGPT has effectively made it pointless to do certain kinds of assignments now.
A lot of professors have been teaching the same way for many years. It’s a reckoning.
Even assuming all the students do it in good faith, there are factors other than knowing the material that can affect their performance on coursework. For example one person may have access to a well equipped, ergonomic, quiet space and plenty of undisturbed time in which to complete the assignment, while another may not.
The major benefit of exams is that everybody is taking them in, as far as possible, consistent and controlled conditions. It's not ideal, but I think it's better than the alternatives.
Imagine using it interactively as a study partner or tutor rather than a homework cheat?
I was playing a bit with ChatGPT and I wondered how much actual knowledge could be stored in that model. So I asked ChatGPT whether it knew the song by Franz Schubert called "Am Feierabend" and if so, if it could tell me the subject/meaning of the song. "Certainly I know this song", responded ChatGPT and then gave me -- with total confidence -- a completely wrong answer about the meaning of the song. In fact I was a bit baffled by the authority with which is spouted this nonsense answer. A truly intelligent system would be aware about the limits its knowledge, right?
So I think that with a minimum of creativity teachers can come up with questions that can easily stump ChatGPT.
ChatGPT created a pretty good essay for my daughter's home assignment. She had to write a fictive autobiography from the perspective of a 16th century noblewoman. Is that too generic? (Side note: ChatGPT did it in Hungarian.)
> A truly intelligent system would be aware about the limits its knowledge, right?
That's a question of definitions. If you ask me, ChatGPT is a truly intelligent system that is ridiculously unaware of its own limitations. It doesn't look like a contradiction per se, I've met very smart, highly functioning megalomaniacs.
No but as you stated it's fictive... I can also tell you a lot of fictive science facts, or fictive president names, or anything fictive for that matter
The goal is to use your _own_ imagination, of course chatgpt can align sentences in a semi cohesive manner, it's its whole purpose
It's not about awareness or limits of knowledge. From the point of view of a language model, it doesn't matter whether it was Todd or Walter White who killed Lydia, or whether it was Kinbote or Shade who invented a phrase. It only tries to generate a response to your input, such that it is a plausible continuation.
Not what they meant, but it may be somewhat of an existential threat to the profession of teaching, the thing is a killer tutor in subjects that it is confident in. It is a perfect fit for language, especially, since it can carry a full conversation with you and correct every single one of your mistakes.
Obviously it has major issues with inventing things out of whole cloth and accuracy problems, but human teachers also aren't 100% reliable or knowledgeable about everything, so it's not like it has to be perfect.
You don't need to know about physics molecules until you face a task related to it. And ChatGPT can direct you to learn more about it.
I am in my bachelor's right now and I see how I can skip a lot of knowledge regarding some low-level computer science skills until I really need them. And at the moment if I can't go further without learning how to effectively optimize ML algorithms using CPP I will hear from ChatGPT that I need to go that way. Otherwise: skip. It is not good enough to provide a whole curriculum for sure. But it is sufficient to get directions to which parts I need to learn.
The education is going to be transformed into "tree" | "unstructured" | "NoSQL" instead of a clumsy set of blocks everybody needs to go through.
This is my opinion
I don't think ChatGPT will really help you learn about actual 'tasks' related to molecules except generating short text paras about them
Wide availability of these tools will create a new "AI Native" generation. What this will do I don't know. Some effects will be good, some bad.
Narrowing to just me, I find the real value I get while learning is asking questions along the lines of "is this concept like this other concept that I know?" When I get an affirmative, then ChatGPT often goes on to tell me something else that I didn't know. These kinds of interactions are very valuable to me. Example from last night in the middle of a long chat:
"How is the query representation computed? Also, is the dot product between the query and key representations similar to computing the cosine similarity between two vectors?"
"The query representation is typically computed by applying a linear transformation to the hidden states of the input words, often represented as vectors. This linear transformation is typically implemented as a fully connected layer, or a matrix multiplication, followed by a non-linear activation function, such as a ReLU. The resulting query representation vector has the same dimension as the hidden state, and it's used to measure the similarity between the word and the other words in the input sequence.
The dot product between the query and key representations is similar to computing the cosine similarity between two vectors. The cosine similarity is a measure of the similarity between two vectors in a multi-dimensional space, and is defined as the dot product of the vectors normalized by their magnitudes.
The dot product of the query and key representations can be seen as an un-normalized version of the cosine similarity, in the sense that it computes the dot product of the two vectors. The result is a scalar value, which represents the similarity between the two vectors, the larger the scalar, the more similar the vectors are.
The dot product, however, is just one way of computing the similarity, other similarity measure, like the euclidean distance, the L1 norm, or even other more sophisticated similarity measures can also be used. The choice of similarity measure depends on the task and the model architecture, also the dot product has the advantage of being computationally efficient, which is important when dealing with large scale models like transformer based models."
Ironically if it were impacting learning I think it would be specifically if ChatGPT was NOT being used across the board, but by only a subset of students. So some kids doing assignments the hard way while others have the equivalent of a calculator/secretary/genii.
But if every student has the same GPT access then the class can simply raise the expectations for the quality of the student work across the board and it's cool again.
I don't disagree but in a world where every GPU on the planet is as intelligent as PHD student, the next year each GPU is twice as smart as the smartest human on Earth, and not long after that every person on Earth is either dead or turned in pure energy or something and homework isn't on anyone's mind...
You can’t “raise expectations” to address work being done by LLM based on its corpus instread of students based on their knowledge of the material; it fundamentally eliminates both the direct learning use and the assessment use of certain assignment classes, for those doing it, whether it is consistent across the class or not.
Aka welp students prefer to learn from ChatGPT rather than our teachers who have poor knowledge transfer skills.
I've been using ChatGPT daily and it is an incredible tool. I feel that I learned so much more through that time than if I had to use Google and books.
I found ChatGPT to be very useful in two areas:
1. where the problems are simple, and the question requires rudimentary knowledge, but of something very domain specific. Where, I would be 100% able to find a better answer myself by reading up on it, but I can get there 90% of the way, and figure out the 10% that is missing or wrong out of context.
If you don't have the ability to fix or identify that last 10%, you might suffer more for it. You'll get there more quickly, but you end up at slightly the wrong place.
2. When the questions are creative in nature. Suggest an imaginary setting, character, describe properties associated with, etc.
It seems to spill into all areas of your life.
I try to impress upon him too that some things take hundreds of hours of explanation before you know enough for it to be functionally useful to you, so you can't always just Google everything. You have to learn things and build your foundations up. Alas he hates the idea of learning anything.
1. Knowing things lets you know more things faster and they stick: associative memories are very durable because they are mostly groups of references to existing objects in memory. Much less novel information that needs to be encoded. Repetition and memorization matters.
2. Look up latency: Same problem as computer cache misses and L1 vs RAM lookup times. You will take 100x longer to achieve the same result by looking up reference material all the time. Also important questions typically have answers from many different problem domains (social, scientific, philosophical, ethical) and its important to have a lot of knowledge memorized to see a solution to a novel problem holistically.
There so many ways this revolutionizes learning for erudite people.
ChatGPT is quite popular on HN and other similar tech-savvy spheres, but is still far from enjoying mainstream success. Maybe like 1% of students have used it yet, and for a month at most, so it's quite the hyperbole to say that it's been "affecting" them already. Or, conversely, if this kind of thing worries you today, you're in for a wild ride ...
And the fact openAI is still struggling to handle the load in peak hours 2 months after launch tells me they must be seeing pretty big user growth.
I've used ChatGPT maybe 3 times since launch and 2 out of 3 attempts resulted in a "come back later" type message. It would make sense to see which accounts are the heavy users and throttle them in favor of lightly used accounts but I guess they didn't think of that or decided against it for some reason?
This changes how easy it is to scale.
Absolutely! And I honestly think it will be a net positive, after all.
As others have pointed, with or without calculators, GPT or whatever, slackers gonna slack and those who take it seriously will have another great tool in their hands to do great things with it.