Glad to hear about an AI made just to help you learn things that can be a 100% positive tool for kids. I'm sort of interested in using it myself to try and learn some maths subjects
We've banned this account for posting lots of unsubstantive and flamebait comments and ignoring our requests to stop.
If you don't want to be banned, you're welcome to email hn@ycombinator.com and give us reason to believe that you'll follow the rules in the future. They're here: https://news.ycombinator.com/newsguidelines.html.
That is significantly lower price than ChatGPT Plus; and since 90% of my value from ChatGPT is learning/tutoring, this seems exceptionally good value, for my specific use case. I wonder how economics work out for them...
I think the people who automatically had a negative view of AI didn't think it through, when ChatGPT came out in Nov 2022, the realization was that something like this tutor friendly AI would be on the horizon.
Personally, I am looking forward to AI enabled healthcare, my sister has been dealing with health problems and she has to wait up to a month between appointments. At the very least AI can be used to speed up all the paperwork that doctors have to complete.
I worked with my uncles law firm and know that the process for getting medical records is a huge pain. Back then I never understood why there couldn't be some better system for getting medical records between authorized parties.
Growing up, my family friend who is a trauma surgeon always told me she hated all the paperwork she had to do.
1) Transcript of the patient and doctor conversation.
2) Creation of possible icd10 diagnosis and treatment codes from the transcript. With suggestions of which ones are most likely to be properly processed the first time per the insurance company of this patient.
3) Quick review of the transcripts of previous visits to determine if there are any overriding concerning trends for this patient.
4) suggested additional tests to reduce uncertainty in the diagnosis and prognosis according to a pre programmed cost to benefit ratio.
I could go on. The local personal health information (phi) approved ai could greatly reduce time spent on paperwork and double work like resubmitting procedure codes to billing.
Eh, on the surface this seems useful but I'm not sure it makes a dent. Transcript might be useful occasionally, but they aren't going to be referred to often - things still have to be distilled in the doctors notes, including things they don't say during the visit. They already have dictation systems for this, but none of the NLP systems are good enough to do it without checking, so time is about same.
Coding isn't done by the doctor anyway. Replacing a fairly low paid employee with a bot might net out but it's not obvious.
Then on the diagnostic stuff, we've been working on this for decades, I think there is real promise there but it's not a quick fix either for getting the system performance good enough to deploy more widely than e.g. CAD systems, and for getting systemic uptake. They aren't nearly good enough yet, we'll see what the next decade of improvements gets us.
The key here is not "a decade of improvements" but rather "can current tech be regulatory approved". The transcript provides context which is phi and invaluable. Doctors care a lot about billing codes
Current tech [for reasonable versions of current] can be approved although not all configurations (basically you have to fix the model for a release cycle). It isn't good enough yet for most uses in the space you describe. Improving that state of affairs is more a data limitation than a technology one, probably, but medicine is also very conservative and will take it slowly to prove out.
Doctors care about getting paid, they mostly don't care about coding per se, and they don't do it themselves. It's not at all clear that a NLP AI would be more cost effective at it than the people who do it now. Where you will probably see it is auto-rejects, from the insurance side, forcing more iterations.
None of this stuff is off the radar of (at least some of ) the people who build medtech systems, they've been incorporating aspects for a long time. The idea that there is a sudden radical change to be made with todays tech is naive I think, lacking context of both the history of AI systems in medicine and the current techs' limitations...
Paper work no, but last week she was being told two different things from two different doctors. None of them talked to each other/between each other and she had to physically drive from one office to another to get her own medical records.
The industry tried using AI to help with that last year, and it was quickly shut down[0] after UnitedHealth / Humana were sued for using blatantly flawed AI projections. It could definitely be used to improve aspects of healthcare, but the profit motive will be at odds with this goal in most cases.
Versions of this have been in use since the 90s, but haven't really made a dent in the systemic issues that cause these delays. What do you think has changed to make that different now?
I've done a lot of experimentation in this space with ChatGPT4 and also the Wolfram plugin. I've mixed but generally good results when working through basic physics problems though you have to be careful about how you prompt. In particular, you want to break down the problem into smaller bite size chunks and eliminate ancillary information. Interestingly, even when it gets the math and algebra wrong, I still find it useful because it gives me hints about how to approach the exercise. Sometimes having several parallel conversations with the Wolfram plugin, for example, can set you on the right track. I expect there will be significant improvements in this arena in the short term.
That doesn't sound anything like a product suited for young learners, many of whom are unprepared to practice the finesse you're talking about and many of whom will want to put no more effort in than strictly necessary.
That sounds like something an especially patient autodidact might use to automate some busy work or help them explore the basics of a new topic, which is fine, but not what the article is trying to champion.
I believe your insights are accurate. Nevertheless, in the age of AI, it's evident that critical thinking remains indispensable. The value of a liberal arts education, which fosters the age-old practice of intellectual scrutiny, cannot be overstated, LLMs or not.
I have fought this my whole life, and only in my adulthood, and self guided learning, have I found the freedom to accept it: I learn better from being given the answer and working backwards. (Even literally reading textbooks “backwards”)
I hope there’s more nuance to this than this quote suggests, because I shudder to imagine a student experiencing what I did but now with a tireless unempathetic agent refusing to just tell a struggling student the answer and instead continually looping on suggestions to “look at it from another angle” in every way but the one “angle” I’m explicitly asking for.
Aligning digital teachers with our broken education system that thinks “answers” are the important part of education is a terrible terrible idea.
That quote should have read ~“the agent has been designed to work tirelessly to do anything to ensure the student understands the material well enough to wield it”
I feel for you, but I would humbly suggest trying to get to the root of this. In all sincerity, therapy is good for this sort of thing. I have similar desires, but I have come to understand them as quite dysfunctional, and not something that should be catered to by the system around me.
Because the real/natural world doesn't work that way. it's very ill-adapted to real life learning. Only in a constructed and false environment can anyone give us the answer to help us learn. The point of learning is not just the acquisition of information, but also to learn how to acquire future information. Leaning into your current tactics (and advocating ppl be supported to learn this way) will leave the learner very ill-equipped to learn in the real world and real jobs, where no one has the answer to give. Learning to thrive (or at least cope) in that situation, that's important
There are 100% at least some ways of learning that ill prepare us for the world we're eventually going to be thrown into
I mean, we're talking "learning". Even negative reinforcement and beatings are a way that was a "learning style" at one point, that ppl defended as "worked for me". "Every way of learning is fine" is not a sensible way to enter the conversation about education
Maybe my rationale was just unconvincing to you (and I respect that), but I said it in the next paragraph. If you felt inclined to offer what's unconvincing about that, I'm open-minded about changing my perspective, or maybe I can better articulate beyond those words
EDIT: I realize maybe I prematurely submitted and it was just the first paragraph for a minute. It was all one thought, so apologies if that's what happened
Thank you, but I did see the whole answer (the second paragraph) when I responded. I sort of know what you're saying about not spoon feeding people, but learning happens best for a person according to that person, and working backwards from an answer can be as illuminating for me as it was for the other guy (and it can be illuminating for the wrong reason – revealing that it was a poor question in the first place).
I don't think you can be prescriptive about how people should learn. In the end we'll have to be able to operate without someone giving us the answers, but in the beginning however we get the learning process we need, that's all that counts.
I guess you could amend the old adage to “on the internet everyone thinks you’re a dog”.
I am decades removed from advanced university degrees.
I am saying how I, as in me, learn best, and acknowledging that it’s different.
My point was why restrict a “generalist engine” to only teach one specific way.
All the old restrictions that pander to: this is how the text book is written, there is only one teacher and 30 students; are baggage and should be discarded in their absence.
I made my educational intentions clear:
> That quote should have read ~“the agent has been designed to work tirelessly to do anything to ensure the student understands the material well enough to wield it”
But I do like and agree with one point you made:
> but also to learn how to acquire future information.
What’s the student learning in regard to acquire future information? To hope some megacorp will steal all of humanity’s creative output to build an llm to teach it to them?
Hm. Y'know what... I regret my comment. It felt ungenerous and obnoxious to write, and I tried to change the tone in edits, but it still feels gross.
Maybe just consider my comment self-talk. I guess your comment hit a nerve with familiarity, and I ended up talking to myself more than speaking to you. I apologize for that. And thanks for the original sincere comment. I'm sure others have seen it and at least some of them felt a little less alone in the weird world of our own brains <3
Getting the answer and working backwards teaches you to use what others have created within very narrow guide rails. Figuring out the answer for yourself (with some minimal guidance) teaches you to create new things. We should strive for the latter.
Getting the answer and working backwards teaches you to use what others have created in novel ways to find new solutions to old problems. Figuring out the answer for yourself (with some minimal guidance) teaches you toy problems with answers have answers. We should strive for the former.
> Aligning digital teachers with our broken education system that thinks “answers” are the important part of education is a terrible terrible idea.
Your position is inherently incorrect and a conflation of what you think is being done.
Being provided the answer might be a decent method to learning X concept. But the answer is ALWAYS downwind of the techniques, which you are actually being taught every step of the way, in every single teaching system. If you learned+mastered the technique or concept, getting to the answer is just proving it with those things, and mostly mechanical while you write down the steps.
How would you solve a problem that no one has ever touched before, and that no one knows the answer to? Where are you going to be provided the answer?
But if you are provided the answer, you can work backwards to figure out the technique. There is a large contingent here on HN that learned electronics by taking things apart, not by reading linearly through textbooks.
Reverse-engineering the technique from the answer is a different skill from creating the answer though. The latter requires more creativity, the former is deduction.
I'd say test your theory, breaking electronics apart need only be a test to reverse the order in which you took them apart. That is only about as good as a memory test.
Test your theory on something that requires quite a few years of intuition building and understanding.
From my days as an MLE: Why does stochastic gradient descent, work?
From that question, you know the answer - it works! Great, now why should I believe you? I.e, prove it.
The knowledge to answer that question only really requires some understanding of stochastic processes, randomness, and vector calculus. But you would be pressed to get an answer like I did out of numerous people who did ML research and put things into all the types of conferences, and these people were not dumb at all, yet all of them used SGD in some form without really getting into why it works, effectively they do not get the tools they're using.
> How would you solve a problem that no one has ever touched before, and that no one knows the answer to? Where are you going to be provided the answer?
This is very near and dear to my heart.
We are talking about “teaching” and I doubt anyone who supports the current education status quo is advocating to teach using unsolved problems.
That said, I do feel, personally, and have long advocated for, teaching using unsolved problems as a better technique, because you learn strategies for approaching a problem much better than: here’s theorem 13.4 heres an exercise problem that uses the theorem, what’s the answer?
It reads like you see my opinions are black and white and yours as spectra.
Using your own reading on your own comment the logical conclusion is “a system that teaches to an answer will be unable to answer unanswered questions because the students were only taught to answer questions that already have answers.”
A great example of the answer being the most useless part of one’s own personal knowledge is Timothy Gowers YouTube content:
> It reads like you see my opinions are black and white and yours as spectra.
No, we're talking past each other.
Your claim is the current system is "broken" because you think teachers believe the "answer" is the ultimate goal. I refute that. The teacher does not believe the "answer" is the ultimate goal, they are teaching you techniques on how to solve problems, and the answer is the standard upon which you are being measured in correctness.
This is why "show your work" exists, your steps to get to the end (the answer) is employing the techniques, they measure how effective you are at getting there. The best teachers even mark you down if you make a mistake somewhere, but still get the right answer without proving it fully.
I'm seeing a lot of conflicting claims about the factual accuracy this thing is able to achieve. If companies really are able to wrangle it into not making up random bullshit, then consider me impressed and my mind changed. If not, then it's a little concerning how it's already being rolled out to teach kids (but not that concerning, since the bar for American public education is already pretty close to the floor).
Maybe blocking the AI from doing math itself and just redirecting those queries to a normal calculator is enough to get the accuracy acceptably high, at least for grade school topics that it's been extensively trained on. At the very least, I'm newly convinced that the whole thing isn't necessarily a doomed idea.
AI for improved education is noble, but what's the point of learning when AIs are going to do everything better than us in a few years?
I already struggle with some intellectual laziness after subscribing to GPT4, which is fairly out of character for me. Why should I expend the mental energy on something when it can do the task for me? Now imagine GPT5, 6, and so on.
I wonder why this comment has been downvoted. IMHO it's a legitimate concern. We already see first symptoms of this incelectual laziness. People using navigation apps for the simplest routes and getting lost or even killed due to over-reliance.
Programmers who replaced probem solving with Google/Stack Overflow searches. Polce officers using apps instead of critical and analytical thinking and simple leg-work. And so long ...
Socrates: At the Egyptian city of Naucratis, there was a famous old god, whose name was Theuth; ... his great discovery was the use of letters ... This, said Theuth, will make the Egyptians wiser and give them better memories; it is a specific both for the memory and for the wit.
Thamus replied: ... In this instance, you who are the father of letters, from a paternal love of your own children have been led to attribute to them a quality which they cannot have; for this discovery of yours will create forgetfulness in the learners’ souls, because they will not use their memories; they will trust to the external written characters and not remember of themselves. The specific which you have discovered is an aid not to memory, but to reminiscence, and you give your disciples not truth, but only the semblance of truth; they will be hearers of many things and will have learned nothing; they will appear to be omniscient and will generally know nothing...
_The Phaedrus_, Plato, circa 370 BC
This conversation has been going on for a very long time. :-)
God this first paragraph made me want to throw my computer across the room:
> Remember when people were furious about kids using ChatGPT to cheat on their homework? If I may be so bold as to psychoanalyze the United States of America, it’s possible the hysteria was not entirely about homework but also the speed with which artificial intelligence was infiltrating our daily lives. How else do you explain so many people defending the sanctity of a thing none of them ever wanted to do?
"When you were 13 years old you wanted to cheat on your homework, did you not? But now that you're 45 and your son is using ChatGPT to cheat on his homework, suddenly you're upset. Curious. Perhaps the issue is your own discomfort with technology your feeble brain can't understand?"
Meanwhile, I spend a non-trivial amount of time every week deflecting questions that could reasonably be solved by a well-crafted Google query...
Schools would be much better off fostering curiosity, self-reliance, tool use, and practical problem solving rather than the all-too-common busy work. So I can relate to what they're saying (to a degree) about AI.
This acutely reminds me of the transition from basic calculators & lookup tables to scientific calculators to graphing calculators to mathematica/maple. At every step, people said "that's cheating". Now we just accept them as tools. Life moves on.
We don't "accept them as tools" in all academic contexts! In many cases using a TI-89 in a calculus class truly is cheating. Likewise with using ChatGPT in an English class.
69 comments
[ 4.5 ms ] story [ 134 ms ] threadIf you don't want to be banned, you're welcome to email hn@ycombinator.com and give us reason to believe that you'll follow the rules in the future. They're here: https://news.ycombinator.com/newsguidelines.html.
0. https://www.khanmigo.ai/
Personally, I am looking forward to AI enabled healthcare, my sister has been dealing with health problems and she has to wait up to a month between appointments. At the very least AI can be used to speed up all the paperwork that doctors have to complete.
How much do you know about the healthcare/medical system? Where specifically in the process would AI be used to speed up paperwork?
Growing up, my family friend who is a trauma surgeon always told me she hated all the paperwork she had to do.
I could go on. The local personal health information (phi) approved ai could greatly reduce time spent on paperwork and double work like resubmitting procedure codes to billing.
Coding isn't done by the doctor anyway. Replacing a fairly low paid employee with a bot might net out but it's not obvious.
Then on the diagnostic stuff, we've been working on this for decades, I think there is real promise there but it's not a quick fix either for getting the system performance good enough to deploy more widely than e.g. CAD systems, and for getting systemic uptake. They aren't nearly good enough yet, we'll see what the next decade of improvements gets us.
Doctors care about getting paid, they mostly don't care about coding per se, and they don't do it themselves. It's not at all clear that a NLP AI would be more cost effective at it than the people who do it now. Where you will probably see it is auto-rejects, from the insurance side, forcing more iterations.
None of this stuff is off the radar of (at least some of ) the people who build medtech systems, they've been incorporating aspects for a long time. The idea that there is a sudden radical change to be made with todays tech is naive I think, lacking context of both the history of AI systems in medicine and the current techs' limitations...
[0] https://arstechnica.com/science/2024/02/ai-cannot-be-used-to...
That sounds like something an especially patient autodidact might use to automate some busy work or help them explore the basics of a new topic, which is fine, but not what the article is trying to champion.
I have fought this my whole life, and only in my adulthood, and self guided learning, have I found the freedom to accept it: I learn better from being given the answer and working backwards. (Even literally reading textbooks “backwards”)
I hope there’s more nuance to this than this quote suggests, because I shudder to imagine a student experiencing what I did but now with a tireless unempathetic agent refusing to just tell a struggling student the answer and instead continually looping on suggestions to “look at it from another angle” in every way but the one “angle” I’m explicitly asking for.
Aligning digital teachers with our broken education system that thinks “answers” are the important part of education is a terrible terrible idea.
That quote should have read ~“the agent has been designed to work tirelessly to do anything to ensure the student understands the material well enough to wield it”
Because the real/natural world doesn't work that way. it's very ill-adapted to real life learning. Only in a constructed and false environment can anyone give us the answer to help us learn. The point of learning is not just the acquisition of information, but also to learn how to acquire future information. Leaning into your current tactics (and advocating ppl be supported to learn this way) will leave the learner very ill-equipped to learn in the real world and real jobs, where no one has the answer to give. Learning to thrive (or at least cope) in that situation, that's important
I mean, we're talking "learning". Even negative reinforcement and beatings are a way that was a "learning style" at one point, that ppl defended as "worked for me". "Every way of learning is fine" is not a sensible way to enter the conversation about education
I'd be happier to accept this if you'd explain why it's dysfunctional.
EDIT: I realize maybe I prematurely submitted and it was just the first paragraph for a minute. It was all one thought, so apologies if that's what happened
I don't think you can be prescriptive about how people should learn. In the end we'll have to be able to operate without someone giving us the answers, but in the beginning however we get the learning process we need, that's all that counts.
I guess you could amend the old adage to “on the internet everyone thinks you’re a dog”.
I am decades removed from advanced university degrees.
I am saying how I, as in me, learn best, and acknowledging that it’s different.
My point was why restrict a “generalist engine” to only teach one specific way.
All the old restrictions that pander to: this is how the text book is written, there is only one teacher and 30 students; are baggage and should be discarded in their absence.
I made my educational intentions clear:
> That quote should have read ~“the agent has been designed to work tirelessly to do anything to ensure the student understands the material well enough to wield it”
But I do like and agree with one point you made:
> but also to learn how to acquire future information.
What’s the student learning in regard to acquire future information? To hope some megacorp will steal all of humanity’s creative output to build an llm to teach it to them?
Maybe just consider my comment self-talk. I guess your comment hit a nerve with familiarity, and I ended up talking to myself more than speaking to you. I apologize for that. And thanks for the original sincere comment. I'm sure others have seen it and at least some of them felt a little less alone in the weird world of our own brains <3
Actually, the real/natural world is perfectly adapted to real life learning.
Your position is inherently incorrect and a conflation of what you think is being done.
Being provided the answer might be a decent method to learning X concept. But the answer is ALWAYS downwind of the techniques, which you are actually being taught every step of the way, in every single teaching system. If you learned+mastered the technique or concept, getting to the answer is just proving it with those things, and mostly mechanical while you write down the steps.
How would you solve a problem that no one has ever touched before, and that no one knows the answer to? Where are you going to be provided the answer?
Test your theory on something that requires quite a few years of intuition building and understanding.
From my days as an MLE: Why does stochastic gradient descent, work?
From that question, you know the answer - it works! Great, now why should I believe you? I.e, prove it.
The knowledge to answer that question only really requires some understanding of stochastic processes, randomness, and vector calculus. But you would be pressed to get an answer like I did out of numerous people who did ML research and put things into all the types of conferences, and these people were not dumb at all, yet all of them used SGD in some form without really getting into why it works, effectively they do not get the tools they're using.
This is very near and dear to my heart.
We are talking about “teaching” and I doubt anyone who supports the current education status quo is advocating to teach using unsolved problems.
That said, I do feel, personally, and have long advocated for, teaching using unsolved problems as a better technique, because you learn strategies for approaching a problem much better than: here’s theorem 13.4 heres an exercise problem that uses the theorem, what’s the answer?
It reads like you see my opinions are black and white and yours as spectra.
Using your own reading on your own comment the logical conclusion is “a system that teaches to an answer will be unable to answer unanswered questions because the students were only taught to answer questions that already have answers.”
A great example of the answer being the most useless part of one’s own personal knowledge is Timothy Gowers YouTube content:
https://m.youtube.com/watch?v=byjhpzEoXFs
But who’s cares what some YouTuber is up to? He’s just out here promoting an inherently incorrect education method wrought from a dysfunctional mind.
No, we're talking past each other.
Your claim is the current system is "broken" because you think teachers believe the "answer" is the ultimate goal. I refute that. The teacher does not believe the "answer" is the ultimate goal, they are teaching you techniques on how to solve problems, and the answer is the standard upon which you are being measured in correctness.
This is why "show your work" exists, your steps to get to the end (the answer) is employing the techniques, they measure how effective you are at getting there. The best teachers even mark you down if you make a mistake somewhere, but still get the right answer without proving it fully.
Maybe blocking the AI from doing math itself and just redirecting those queries to a normal calculator is enough to get the accuracy acceptably high, at least for grade school topics that it's been extensively trained on. At the very least, I'm newly convinced that the whole thing isn't necessarily a doomed idea.
yawn
If you wouldn't mind reviewing https://news.ycombinator.com/newsguidelines.html and taking the intended spirit of the site more to heart, we'd be grateful.
I already struggle with some intellectual laziness after subscribing to GPT4, which is fairly out of character for me. Why should I expend the mental energy on something when it can do the task for me? Now imagine GPT5, 6, and so on.
People who "learn" are going to become a rarity.
Thamus replied: ... In this instance, you who are the father of letters, from a paternal love of your own children have been led to attribute to them a quality which they cannot have; for this discovery of yours will create forgetfulness in the learners’ souls, because they will not use their memories; they will trust to the external written characters and not remember of themselves. The specific which you have discovered is an aid not to memory, but to reminiscence, and you give your disciples not truth, but only the semblance of truth; they will be hearers of many things and will have learned nothing; they will appear to be omniscient and will generally know nothing...
_The Phaedrus_, Plato, circa 370 BC
This conversation has been going on for a very long time. :-)
> Remember when people were furious about kids using ChatGPT to cheat on their homework? If I may be so bold as to psychoanalyze the United States of America, it’s possible the hysteria was not entirely about homework but also the speed with which artificial intelligence was infiltrating our daily lives. How else do you explain so many people defending the sanctity of a thing none of them ever wanted to do?
"When you were 13 years old you wanted to cheat on your homework, did you not? But now that you're 45 and your son is using ChatGPT to cheat on his homework, suddenly you're upset. Curious. Perhaps the issue is your own discomfort with technology your feeble brain can't understand?"
Schools would be much better off fostering curiosity, self-reliance, tool use, and practical problem solving rather than the all-too-common busy work. So I can relate to what they're saying (to a degree) about AI.
This acutely reminds me of the transition from basic calculators & lookup tables to scientific calculators to graphing calculators to mathematica/maple. At every step, people said "that's cheating". Now we just accept them as tools. Life moves on.
- the first line is me, being so annoyed at the article that I want to destroy a laptop
- the blockquote is a quote from the article itself, including a really dumb, condescending rhetorical device about "sanctity of homework"
- the part in quotation marks is mocking the author
Am I just old? Is this what getting old and curmudgeonly feels like?