I am starting an AI+Education company
twitter.com
twitter.com
I'm skeptical that the product described in the announcement will result in a business able to generate VC-level returns. Based on my experience as a tech entrepreneur with edu-focused products across multiple startups and decades, scaling sales revenue to significant levels in edu is extremely challenging for unique and non-obvious reasons. None of the edu-specific challenges are directly related to the effectiveness or relative value of the product itself. They are structural problems related to the three main types of buyers, which are: 1. Schools, (K-12, Univ), 2. Companies (employee training, compliance, etc), 3. Individuals (self-learners). There are other types of buyers but they don't represent nearly as much aggregated potential revenue as these three. Note that these are the economic buyers not the product end-users.
Each of the three types presents different challenges from the others, however there recurring meta-themes. The main one is that their purchasing decisions are often driven by factors unrelated to the product. For example, the primary focus of schools is, sadly, not maximizing net student learning. While that is a consideration, it's swamped by other external requirements, many of them wickedly complex. For companies, employee training, whether a regulatory requirement or purely to increase employee effectiveness, isn't core to their actual business - meaning it's a cost not revenue. Individuals on the other hand break into two types, those doing it to actually learn something for themselves, and those seeking certification or external validation like a degree to increase their future employment earnings. Learning lovers are a pure consumer, non-essential, discretionary spending-type sale with all the same well-known LTV and customer acquisition cost challenges. For certification/validation buyers, they aren't primarily focused on the learning but rather checking a particular box that has high perceived value to external parties.
While I agree that creating a new large and growing type of customer in addition to these three is possible, it's a huge lift that's risky, expensive and will require substantial sustained investment over time - all ahead of meaningful revenue. So, I wish Andrej et al the best and I'm rooting for them but I'm not optimistic they'll build a high-growth, high-revenue sustainable edu business around AI as the primary value differentiation.
But there is another buyer that your post leaves out: parents. This is not the same as individuals/self-learners (the buyer is not the user, exactly) — but the buyer has strong incentives that align with student performance that are different than the others.
Historically, this has been a very challenging market to scale up well — exceptional 1:1 tutors are rare, and hard to train, they’ll be expensive and often form a relationship with their tutees that incentivizes disintermediation of any platform that tries to capture some fraction of the value by offering a two-sided marketplace (YC has funded at least one of these!)
However, AI has the potential to dominate in this market. Most tutors are actually not amazing, but 1:1 tutoring is just so much more effective than basically any other kind of educational context (two-sigma improvements!) that it doesn’t super matter. AI could be a much more effective tutor — at a cost substantially lower than a human tutor. Of course, wealthy families may still want to hire the well-pedigreed, role-model-level, truly exceptional tutors. But folks who are lukewarm about their private (or small-group) tutors may just give the AI tutors a shot—and find that they’re better than what they were paying for.
Or not! That’s what makes it a risky bet. :)
But it’s a large, $50B+ market today, and growing fast — not one to be discounted.
Self motivated learners who will be willing to pay you can also be groupped in a few groups - the ones seeking future employment in tech sector, looking to understand the technology better and showcase it on their resume with some projects that your platform will allow them to complete and ones doing this purely for fun or to keep up, I believe the latter group is very small.
Platform alone, even with tons of learning materials will not be enough to get this off the ground, you will either need to leverage your existing audience, bringing us back to the fact that everything is an attention economy now, or, you'll need to be very advertisement heavy on YouTube, which you can bruteforce with VC investment, see skillshare for example. NordVPN took off the same bruteforce way, although in a different product category.
My platform will combine technical learning with in-material challenges, sort of a sprinkled leetcode if you will. The idea to implement a LLM side-tutor also crossed my mind.
I think the only hope would be to run a pilot and show such an improvement in student testing levels that it would be fast-tracked or mandated for approval.
I don’t have an X account and as such might miss additional information from subsequent replies but there is nowhere in the post mention about building a high growth/high revenue business? Maybe Andrej just enjoy teaching and sharing knowledge and want to bring like minded, well off people together rather than creating an unicorn?
Or do you mean by your post that it is not (or is very hard) to even reach non negative revenue to keep it sustainable on its own?
If anyone else had done this, I'd be very skeptical but give Andrej's penchant for teaching and how good he is really at it, I am super excited about this endeavour.
Let's not forget the inbound and organic interest this would drive. For the intro LLM Course, people are already submitting PRs and Andrej is requesting them not to. Imagine that!
https://github.com/karpathy/LLM101n
Update June 25. To clarify, the course will take some time to build. There is no specific timeline. Thank you for your interest but please do not submit Issues/PRs.
- Classroom learning doesn’t cater to individuals
- Corporate learning programs that I came across were 100% useless
- Personal learning is unbelievably hard to organize (persistence, learning path, real world application…)
I had basically five wonderful experiences:
- Karpathy on YT ;-)
- Dash and Dot from Wonder Workshop
- Crafting Interpreters
- Gilbert Strang
Hope eureka labs becomes the sixth. Good luck Andrej!
Your analysis of the market structure and the challenges is consistent with my experience - I'd like to add 2 points though:
1. Building on what you've said about VC-level returns, while edtech (and industrial edtech) may not necessarily be fields where VC funded unicorns can actually form and then survive to profitability, I believe there is sufficient need and space for companies that are investable and can grow into profits / valuations in the 10s /100s of millions - they can follow the startup format, etc.; just not VC investable, as you've noted.
2. In the space of skill development (i.e., adults / professional learning), I would like to add that for many skills, as industries get more sophisticated (I am taking of manufacturing, automotive, energy, oil & gas, infrastructure, petrochemicals, infrastructure), there is value for the company in upskilling their staff - simply because that does directly convert to profitability if done well; and at least puts a floor on issues that poor or absent training may cause - like safety, quality, etc. So even if it is a cost, it is a cost that can directly affect their topline / bottomline, and I believe there is an opportunity here.
/off soapbox
What about parents? That would seem like a natural market for a product that could replace expensive private school tuition (or paying high property taxes in order to be in a good public school district). I recently saw a literacy product aimed at parents, and which costs $50/month. They justify their pricing based on its substitutes being a tutor or private school tuition. While this is a bit of a stretch for a literacy tool, it would be a very fair comparison for a very good AI tutor that teaches the child and has a dashboard/summary for parents.
I do wonder if recent (Covid onwards) trends toward increased home schooling blurs the line between 1 and 3, or perhaps manifests a new segment that could be meaningful; homeschool pods seem to be quite outcome oriented and don’t have any of the bureaucracy of schools. Also I assume for now they are non-material from a TAM perspective, but interested if you have any thoughts on that trend.
Didn't Apple make a great deal of their revenue in the early days (1976-1985) by selling to schools? I seem to recall that schools were their main focus. A $3000 Apple II in 1980 would be $12,000 in today's money, and schools often bought many units. Is Karpathy's product really that different from an Apple Computer in how it would be marketed to schools?
Having a scalable technology that can teach and reinforce learning for humans will change the world. The human potential of teachers will be freed and human intelligence could rise all minds.
If it works. Nothing similar in "ed-tech" has actually offered value thus far.
That's true for literally any business and nothing specific to edu.
We don't even need AI. I'd love to see something like Khan Academy mixed with space repetition and an DAG of knowledge dependencies. If you get a question wrong, the spaced repetition algorithm should present material about the underlying dependencies. Unfortunately, Khan Academy is focused on conforming to the linear paths (curriculums) that schools and universities have chosen.
https://www.dealii.org/developer/doxygen/deal.II/Tutorial.ht...
i.e. you learn about "addition". Then you repeat your knowledge of "addition" when you learn about "multiplication", because it is just "addition" but with extra steps.
I think the more important a concept is, the more spaced Repetition you will experience. If you only repeat a piece of information very rarely, it may not be very essential, and you should consider dropping it from the curiculum.
Also there are information, which only exist as stepping stones to understand more complex topics. And once you understand the more complex thing well enough, you can forget about some of the stepping stones. For example: I once learned a proof, that basically says that all computers can solve the same set of problem in roughly the same amount of time (within a linear factor). I forgot all the details about the proof, but in my every day life as a programmer it is important to know, that the benefit "better hardware" can only get more so far, and for really strong performance-gains, I need to improve my algorithms.
Only the most dedicated of teachers will review addition with you yet again before moving on with multiplication lessons. After all, there are dozens of kids in the class, they can't all delay the lesson just for your sake.
Look at the ways we change teaching
In my opinion, the whole AI trend is another round of industrialization making low-quality options for the poor.
- You're too poor for meat? Here is some hyper-processed, growth-hormoned burgers.
- You're too poor for a proper house? Here's a shitty pre-fab with walls so thin you can smash through them with your bare fists.
- You're too poor for fancy clothes? Here's a Bangladesh-sweatshopped T-shirt that won't make it past 3 washing cycles.
- You're too poor for school? Here's a computer with an LLM on it.
You could say that at least poor people have houses, steaks and education, but in the end, it's just a way of paying people less, cause they don't need much to send their kids to college anymore...
In any case, you can be sure that Harvard kids won't get any of that crap.
In addition to that, think in most fields there's an increasing need for education that is more flexible than multi-year degrees, simply because requirements for skills change so fast these days.
Everyone uses the same smartphones (the poor people use phones about as good as rich people used less than a decade ago), everyone drinks the same soft drinks, everyone takes the same cold medicine (the generic ones are pretty good), everyone uses the same social media, the same operating systems (even the richest Microsoft customers have to put up with their adware)
I bought them about 7 years ago for 1 pound each. I think they're 2.5 pounds now.
1. LLMs are in their infancy
2. LLMs are better at explaining things than A LOT of teachers.
3. LLMs don't mistreat/shame students for lack of understanding
If I could get a degree just by chatting with an LLM, I'd do it in a heartbeat. School is about gaining understanding and skills, not something else.
I don't need a human to hold my hand along the way, I need an efficient, fair, and cheap system to prove I know 2+2=4.
I see no reason that chemical molecules and their reactions cannot be transformed into music. Mathematics, programming, physics could turn into songs and registered to memory better than any other way. A.I. could help with that as an assistant.
[1] https://www.independent.co.uk/tech/harvard-chatbot-teacher-c...
I think a lot of the real issues with solving problems in education is that they have trouble applying to the larger picture of compulsory education.
It is self-evident to me that this is not true in the U.S.
There is very little about education that is egalitarian. An inner city grade school in Michigan is in no way comparable to a grade school in the Palo Alto suburbs.
> Such that no one is left out or treated better if you didn’t happen to go to an “AI” school makes for a tough hill to climb.
To drive this point home, in the real world we do not hold this standard for air conditioning. Tens of thousands of schools in the U.S. do not currently have adequate HVAC systems. If we are not currently applying this standard to A/C, I don't know why we would choose to selectively apply it to AI.
Attempting to gate innovation behind complete social conformity and universal adoption doesn't strike me as a rationale stance.
Part of the cost is localized like salary, utilities costs.
Detroit isn't necessarily being under funded compared to Palo suburb when you normalize regional cost differences. It does affect the quality of teachers one can attract but so does the location in general. Bumping up Detroit to Palo Alto's spending wouldn't have much effect on the education for the average student.
Palo Alto is home to many immigrant from many places who self select and place education as extremely important for the next generation. For the average Detroit kid education isn't cherished and valued over everything else. Until you can equalize those believes things won't change. You won't be able to buy a higher educational rate by painting the walls twice as fast or by hiring more educational assistants.
A smart kid with a phone in Kenya should be able to have a shot. Wikipedia allows this. Khan Academy helps. This in the extreme is Khan Academy but potentially for every course a kid (or adult) could want, generated on demand from scratch if need be.
The next optimal study session generated for your current state of knowledge, energy level, learning goals, by the teaching agent who is provably the best at delivering that lesson for you, using the optimal mediums for your context. Video at home or in a class, audio when on a bus, printouts if you want. Add in spaced repetition and a few fun quizzes, gamify it. Let each person fulfil their potential unhindered by the resources of their school, parents or country.
They just installed some state of the art AI-enabled, "smart" mega drawing screens w at my daughter's schools touting all the supposed immeasurable benefits it will bring, and most of the parents, including myself, just rolled our eyes.
There are all kinds of self-interested reasons educators might resist some of these technologies. But this also seems to be one of the areas where people from the tech world impose some idea that could potentially work for their limited domains of expertise but don't work at all for others.
This is the approach I'm taking.
There are a few psychological and technological tricks here or there, but in my opinion the hardest part is getting the incentives right.
It's very easy to build an educational system that incentivizes for the wrong things. And even if we all agree what the "right" things are, something like universal education is a collective action problem with all the difficulties that come with that.
Educational goals are to educate people not make money. Education requires countless failures. Business models are composed to be failure averse or reduce risk.
Education has and will likely continue to be a money and time sink because it is an area that requires constant failure from its participants to grow in knowledge.
Thank you for setting this.
Why?
I've noticed that many people in tech seem to disregard or disrespect educational institutions. So I'll turn it back on you. What draconian reason could you possibly have to make the argument that we shouldn't try to give every child an equal opportunity for a high quality education? Do you hate living in an educated society that much? Are you interested in living in a malthusian nightmare?
The human species has only existed as such for ~100,000 years. Almost all human societies have failed.
We should spread knowledge far and wide for the same reason adaptive mutations spread through a population: shit happens.
Rocks fall out of the sky.
The Earth is jelly with a thin "crust" of congealed goo on top. So-called "solid ground" is thinner relatively than the paint on a globe. It shakes.
(As a kid I lived through Loma Prieta[1], I've seen the earth roll like Santa's belly. We are small!)
We should go through life like "an old man crossing a river in winter", and things both precious and free, like knowledge, should be the treasure of every person, no matter how poor or weak, for tomorrow they may be all that's left.
That's why Wendy, that's why.
[1] https://en.wikipedia.org/wiki/1989_Loma_Prieta_earthquake
I’m a software engineer at major US research university developing AI-powered software to improve critical reading and writing skills in higher ed. The idea is to provide immediate, high-quality feedback to students, closing the “latency” of submitting something and waiting to hear back from you professor.
I do genuinely think AI can reshape teaching and learning, but it will be a slow iterative process. We can use it scale what works (personalized learning and tutoring, helping students develop mastery/automaticity on topics, targeting areas where they struggle). It can also automate time-consuming tasks that bog teachers down.
If you're interested in pedagogy, AI, and tech, please reach out.
Oftentimes, the root cause of the critical reading problem is the quality of the writing that students are subjected to. My daughter recently showed me one of her economics readings, and said she couldn't understand it. It was 40 pages of convoluted academic writing like this:
Wibbels argues that developing countries face an inherently disadvantaged position in the world economy due to their dependence on foreign capital and an undiversified base of commodity exports as primary sources of hard currency. This dependent position relative to capital markets prevents developing countries from borrowing to engage in counter-cyclical aggregate demand management.
Is such language the optimal way to express ideas for comprehension by peers, students, and policymakers?
I hope your mission to improve writing skills in higher ed addresses the source of output - professors, teaching assistants, journal editors, and others who continue to promote outdated, inconsistent, and counterproductive academic writing styles.
“Wibbels claims that developing countries are at a disadvantage in the global economy for two main reasons:
1. They rely heavily on foreign investment.
2. They depend on exporting a limited range of raw materials to earn foreign currency.
Because of this weak economic position, developing countries struggle to borrow money when needed. This makes it hard for them to boost their economies during economic downturns, unlike wealthier nations that can more easily borrow and spend to stimulate growth. “
Wibbels says that poorer countries are in a tough spot in the global economy. They rely a lot on money from other countries and mainly sell raw materials to get cash. Because they're so reliant on foreign money, they can't borrow much to manage their economy when things are going bad.
This is a pretty basic prompt too, you can add qualifiers to fit your general audience: "easier to understand for a high school student". You could provide context of previous and subsequent passages as well. Yeah its not perfect, but the right prompt could provide at least a consistent output style.
Those are three completely different groups with completely different needs.
I have a roommate who is a weak non-native English speaker. However, he needs to write and submit scientific papers in English. He uses the "pro" version of ChatGPT to improve his written English. He said it is like having a 1:1 English language tutor because he gets nearly instant feedback when trying to rewrite a sentence or paragraph. I am native in English. He showed me some before and after examples. His message remained unchanged, but the revised versions were so much smoother. My point: He is not using ChatGPT to "cheat", rather to improve delivery of his message. I wrote about this previously here on HN. It received very mixed reviews.
I think what your roommate is doing is fine. Being a native English speaker is a bit of a cheat anyhow. I very much appreciate the challenge of learning another language, so am not going to fault someone for using different tools to help improve their language. So long as the ideas are (relatively) original, the specific wording seems less important.
https://www.youtube.com/watch?v=hJP5GqnTrNo
How AI Could Save (Not Destroy) Education | Sal Khan
---
edit: Khanmigo has its own site now
Most learners, the world over, are not self-motivated. The pandemic showed us exactly what children would prefer to do, when they don't have a physical teacher standing over them, which is bugger all. We send kids to school, in the hope they get some education, but the reality is that we use schools for free childcare while we work. If parents have to additionally monitor their child's learning, it breaks down pretty quickly.
I see AI being more of a teaching assistant, rather than a replacement for teachers. Having been in the education game for over twenty five years, I know the difference in impact when comparing virtual learning to in-person training.
The issue is not engaging teachers. The teachers we have here in BC are excellent and love their subjects (my wife and many of my friends are teachers). The issue is behaviour, which has deteriorated significantly since COVID, though the changes have many other contributors.
Try asking an AI to engage with 30 kids who are on their phones with earbuds in. You absolutely need a human as a teacher.
That said, AI teaching could be a great teaching assistant.
In my opinion, the best enforcers generally are charismatic yet firm and come from a similar community/background to that of the students. The best teachers have an infectious passion for their subject, but oftentimes that trades off with their ability to enforce.
this seems like a bizarre conclusion. In my experience, most people, the world over, are in fact self motivated. You won't see that if you have a very narrow definition of what is it that they're supposed to be learning
kids aren't motivated to do boring math drills, because they don't see why it matters to their life (the real answer is: it does not, they are not wrong).
I appreciated hearing this echoed by Conrad Wolfram in a recent PIMA episode: https://freakonomics.com/podcast/why-do-we-still-teach-peopl...
In your experience? The world over? Can you tell me your experience. I've been a teacher for a long time. I've worked in the UK, the USA, PNG, and Kenya.
The vast majority of kids in the developed world don't really care about education. A few do, and they get great grades. Most care more about social status, their cliques, or just surviving the jungle that is school.
School is important. It teaches you how to deal with other people. It teaches you how to deal with people in authority. You can't get that at home, in front of a screen. Learning stuff is secondary. I'm sure there are plenty of people here that are not working in whatever they majored at.
Most kid athletes are also not self-motivated to run laps, or do boring repetitive drills, when they know from experience that these activities help them win games within the next few months. Usually need a coach to force them to do them. Same for young music players. Practicing scales endlessly does make you a better musician. But they won't do it till forced.
The primary reason kids don't like running laps or playing scales or doing math drills is because they are boring.
I’ve heard that kids in upper middle class circles are totally different in this regard though. Maybe they want to do more on average.
The real real answer is that it probably does, but on a much longer timescale that we generally consider and it is really hard to explain why. Something like better math skills lead to better life outcomes. Maybe due to a better model of the world and sharper thinking, but I am just guessing.
I think you are partially right in that the dryness of much of math teaching hides a lot of the underlying material's applicability to life. I think one thing AI could do is help design rich situational lessons that could are prompted, vetted, and updated by teachers and then taught to the class. It could be trivial to create incremental difficulty of problem materials tailored to each student's progress and goal.
So you are a teacher?
Upon reaching a certain threshold of technological dependence, the need for rational thought (which includes calculation) is tied to the need for food. The actual yield may be low based on other factors, but it is absolutely necessary for survival.
The alternative you suggest, is where technology no longer advances.
Logically then, population growth hits a malthusian trap, the old crowd out the young since they have the most influence, and then a depopulation occurs as the old naturally die off, and replacement births cannot sustain those dependent systems used to feed the masses.
You get a dragon-king event where everyone its a free for all over food and bare necessities, farming no longer becomes possible (because of looters), and the world order collapses to pre-agrarian levels, assuming the environment isn't destroyed in the chaos (i.e. MAD and Nuclear Fallout).
There are much better ways to calculate than are currently taught in schools, Trachtenberg System and Vedic Maths have worked well in many places.
Mental math has been around for quite some time, and the principles of math are all about finding uncommon knowledge or information that is not immediately apparent (though it becomes so via various mathematical transformations).
The current pedagogy of math is all about sieving and exclusion, and rote-authority based teaching, since it is a requirement for any specialized area of science (and is only taught in relation to mathematical concepts, instead of intuitive approaches). This is why they adopted a burn-the-bridge strategy right around trigonometry at the grade school level (intended to cause PTSD/suffering/torture), to safeguard against disruptive innovators at the source.
Algebra -> Geometry -> Trig
1 -> 2 -> 3
What do you suppose happens when the passing grading criteria in 1 is changed from just following the process (but not correct answer) to 2 (separate unrelated material which is passed) to 3 (correct process and correct answer).
If they fail Trig, and the problems are from Algebra (not something a teacher paid bupkiss will bother to look at), how do they go back if they passed Geometry? The students not knowing why they are failing are simply told, well you maybe you are just not good at math and should consider other paths if you can't do it.
This structure is called burning the bridge because it makes it so you can't go back from a progression standpoint. Ironically, this structure was adopted at the request of representatives from the National Teachers Union in the late 80s/90s, and largely remains the same today.
There are several other progression sieves embedded in academia intended to make it almost impossible for us as a society to develop a large number of creative people who reach einstein-level achievements in math and science (outside-self study, or specific environments/private schools).
This broad push largely started in the 1970s in publishing, and expanded from there.
[1] https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4895486
There are also a lot of things that can come in before you build a full on tutor. One example is being able to tailor word problems (transform the nouns) to subjects interesting to the particular student. They could also be used to help understand where students are struggling. We are still at the early phases of useful AI, optimism is more appreciated, especially as contemporary times have become so pessimistic
Sal Khan provides a more optimistic take and demo: https://www.youtube.com/watch?v=hJP5GqnTrNo
This is not true. The pandemic showed us exactly what children who are accustomed to being force-fed information and whose natural learning mechanisms and curiosity have been suppressed in favor of a generalized one-size-fits-all approach do when suddenly removed from the only learning paradigm they've ever been exposed to.
My kids (not yet old enough for school) are extremely self-motivated to learn and explore the world around them. So am I, and that never went away over the course of a full homeschool education.
You and your kids are not typical of society at large.
There’s teaching students like. There’s teaching where students learn. Sometimes the two intersect. Will an AI education company optimize one that students enjoy or one where they learn better?
Layperson coverage: https://news.harvard.edu/gazette/story/2019/09/study-shows-t...
Most people aren't academically inclined so it follows that most people aren't academically self motivated. Therefore among those that are academically inclined it is important to provide them with all the tools necessary because they're the ones that will most likely excel in an academic environment.
It is odd that the curriculum tends to accommodate people that aren't academically inclined at the expense of those that actually want to learn.
People that aren't academically inclined should not be forced to learn, or at least forced only insofar as they're baseline literacy so that they function in today's world.
Teachers like gifted kids because they'll be successful no matter what they do, and the teachers can test out all of their dingbat social and pedagogical theories with no consequences. They can start with elite kids, finish with elite kids, yet somehow take the credit. Not impressed. Make dumb kids smart, then I'm impressed. You might even be holding back the smart kids, but they're probably smart enough to see through you and do well anyway.
That being said, there are some people who are motivated to learn entirely by the desire to impress teachers and other authority figures. They need attention to develop. However, I do not think that most people are like this, and I honestly think those people should be in therapy.
You can't run a billion dollar chip industry on passion alone, you are going to pull in people who may be working for external reasons. What matters then I'd that the education they receive is effective regardless.
Isn't that by definition? Most xes are x.
> Most people aren't academically inclined
Is that so?
> It is odd that the curriculum tends to accommodate people that aren't academically inclined at the expense of those that actually want to learn.
Well, if what you say is true, isn't it fair that the program is catered to the majority, who are apparently not academically inclined? For one size fits all mass education, catering to the largest mass is the best you can do.
> People that aren't academically inclined should not be forced to learn, or at least forced only insofar as they're baseline literacy so that they function in today's world.
Isn't that what the curriculum already accommodates then? Didn't you just say that?
What a teacher provides is a sometimes customized, sometimes flexible schedule, that (sometimes) pays individual attention to what aspects of a concept a student is falling behind at, and (sometimes) comes up with personal recommendations and alternative approaches to break down a student's involuntary resistance to a concept. This might be doable with A.I.. It's not doable with actual teaching anymore because class sizes are too large. A.I. will be cheaper.
And I'm not saying that teaching is so simple that A.I. can do it, I'm saying that teaching is so complicated that it might be that only A.I. is sufficient to largely replace it. I think that what I'm arguing against is that the idea that teachers could be replaced by glowering scarecrows, or fur-covered wire armatures like they once used in experiments to replace animals' mothers.
I don't think that teachers make as good parents as parents do teachers. I don't think most people are mostly motivated by the approval or judgement of their teachers.
What people need is constant, helpful, personalized guidance, and that is very expensive to get from employees.
And I'm not trying to make a general argument against in person training. But I think the details of how virtual learning happens matters quite a lot. AI can make it much more personalized and make tutoring relatively affordable. Don't you think?
It's been extremely effective for me, where reading a math textbook/wikipedia article seemed like too much effort, but a friendly conversation with my AI tutor was just fine.
You cannot replace that with a machine.
Amend to: The pandemic showed us exactly what children's parents would prefer to do, when they don't have a physical teacher standing over their children, which is bugger all.
This is the exact opposite conclusion and methodology of Maria Montessori (and her schools with the same name). Children are naturally curious and want to learn, but they may not want to use a poor education system designed to mark grades in a hyper specific focus.
even when given access to school choice, less affluent and minority parents do not choose montessori and there is absolutely a reason for that.
That is exactly what he says in the tweet.
I think the problem with traditional teaching, as in any skilled profession, is often in short supply and underpaid, not happy, and unable to keep up with 25+ kids in a class. The world needs orders of magnitude more teachers that are highly competent and more easily accessible.
AI could massively scale high quality teaching with still a teacher in the loop.
So far, AI can't replace good teachers. But there aren't that many good teachers. In my experience, GPT4 is better at explaining advanced concepts than 70% of college professors. Unfortunately, education is often oriented around this horrifyingly archaic method of instruction, which prevents people from imagining what an AI oriented system could look like.
AI based education might or might not be "MOOCs 2.0". Even for the less good teachers, having a real human in the room is one of the features that lots of people appear to be ready to pay lots of college fees for.
> Khanmigo is an AI-powered personal tutor and teaching assistant from trusted education nonprofit Khan Academy.
We also send kids to school to learn social skills they can't learn by themselves.
My kids sometimes watch science shows (on TV as well as online) and tell me all kinds of fascinating facts about black holes and the human immune system and {insert_huge_list_of_stuff_I_don't_fully_understand}. That's the easy bit.
"Getting along with other people" isn't something you learn ... by yourself.
Most teachers today aren't experts anyway, we just pretend they are. So I'm not sure "replaced by AI" is the right way to frame the conversation. Instead, it may change education.
Experts in what, grade school math? Do you mean professors?
Lmao what
> hope they get some education, but the reality is that we use schools for free childcare
Exactly, teachers will be less and less pedagogs and more and more wardens.
So would fusion power, unfortunately such a thing does not exist yet, nor close to.
Teaching is not the end goal of education though, the educated student is. Or so I was taught.
Part of the reason why teaching is considered noble is because it is an act of assured replacement, inspiring not dependency imparting skills of self-motivation and will power.
Obviously someone like Andrej will totally crush it.
The future of education is the playful gamification of relevant skills, knowledge, and behaviors.
i’m not trying to be pedantic, but anytime someone implies a human, particularly a kid will be at all predictable shows an incredible lack of understanding of people. the vast array of moods, time of day, quality of sleep the night before, are they hungry, mood of the parents when they drove them to school, how did their school/work day go, how was their social day, and on and on and on.
again, apologies, i’m not trying to be pedantic but i think in this particular topic it reeeeaalllly matters.
For instance, I’ve had trouble understanding exactly how heat pumps worked. Sure I knew the basic concepts of condensation and evaporation but not the nuances of pressures and boiling points at various stages. I asked chatGPT to explain it to me from the perspective of the refrigerant. It started with “I am R-134a, a refrigerant just leaving the evaporator…”, and proceeded to give me the most thorough understanding of heat pumps I could imagine, complete with working pressures, boiling points, pressure differentials at the escape valve etc. Follow up questions led me down interesting paths where it came up with a brilliant comparison to quantify the greenhouse potential of the refrigerant R22 ie 1 pound of R22 has the same greenhouse potential as a human being breathing for 787 days in a row.
That’s what he announced he’s doing. Creating an assistant, not a replacement.
I got a lot more motivated to learn when i learned programming.
during the pandemic, the world was in shock, so of course kids are going to play video games when their parents are anxious and filled with cabin-feever.
For example, how do you create an AI language teacher without creating an AI that can make learning languages obsolete? If you've got an AI that can, in real time, hear other languages and translate them (as you might have for an AI language teacher), then why would a human need to spend countless hours learning this other language? Just hold your phone up and let AI do the work for a fraction of the effort.
For a harder, non-solved problem, consider math. For an AI to do math will require something unknown at this point, if it can ever happen. But assuming it does, why would we want a human ever to "do math" ever again when we have the AI that can teach it just do it for us? The AI will almost certainly do it more cheaply and with more skill than a human IF it can be done at all.
It's this sobering realization that I struggle with. If someone can tell me where I'm wrong I'd be greatly pleased
You need to have your phone on you all the time, otherwise you can't communicate. There is always going to be at least one clause worth of latency in each direction of translation due to differences in word order and semantics, so you'll be at least a sentence or two behind the conversation - whereas if you learn the language, you're interacting in real time. Also do not underestimate the goodwill that comes from being willing to learn a language - people recognize that it's hard, they're going to be much warmer to someone who spent over a thousand hours to reach B2 in their second language vs. someone who downloaded an app instead of putting the effort in.
Near real time speech-to-speech machine translation is super cool if you're a tourist visiting a country for a couple weeks, or an employee visiting a factory in a country you don't speak the language. It isn't a replacement for learning a language, though.
Coding, for example, is just a small subset of the tasks necessary for software engineering, which in turn is a small subset of the tasks necessary for SaaS company.
We can break down human endeavors into subjects that are "AI teachable" and without ever needing the AI to be able to use them in a sophisticated fashion.
We already do it like this, but with humans.
If the former, how are people learning unteachable moments right now.
If you only want some basic communication, sure use a translation app, but learning a language is also about learning a culture. Learning about new music and literature and poetry that you'd never otherwise get exposure to.
It's like asking what's the point of learning a musical instrument when I'll never be a great musician.
It's to benefit myself so what would be the point of just getting a translation. They're never the same.
AIs can also teach us to do things they cannot themselves do, for instance you could have a driving-test tutor. It could teach you a lot of things, despite us not having full self-driving AIs.
That's not necessarily the case. A good tutor can be more "curator" than "creator of course content"
I can learn math by reading existing math textbooks. Imagine having an ai that is able to judge my knowledge of that math and assign reading or problem sets accordingly. Imagine an AI that is compelling to talk to and keeps me on task given my stated learning goals.
None of that requires superhuman anything.
Would I still want to learn linear algebra even if the frontier of math is being advanced by some super intelligence? Sure, why not? Isn't the frontier already being advanced by many people smarter than I am?
However, creating many individual systems which are better than a human in a specific area each is very different from one unified and integrated system which combines all of them. The latter will take a lot longer to achieve. Until then there is value in teaching humans.
In the language example, people still want to learn to read/write other languages despite many translators being available. The tech to teach might be less sensitive to latency than the skill in humans, or it might be very expensive, or it might be useful to non-tech-literate people.
If such an AI teacher style becomes widespread, this means that they have the potential to replace the parental relationship (in the same manner AI girlfriend/boyfriends threaten romantic relationships).
I see people talk about the dangers of the AI girlfriend/boyfriend, but not the dangers of introducing AI teachers to (especially young) kids. Nominal adults are already being affected by this (see Replika and company) and they are not even the "best".
If I wear my cynical hat for a second, I'm willing to bet that this parental replacement is a certainty, as an extension of the "screen" parenting that already exists. But this time, it might actually be helpful for the child so it will be socially acceptable and encouraged.
Spot on. It would be interesting to see how one would train a model to do this
My guess is that current models are not yet able to provide consistent, effective motivation to learners over the long term. Like a lot of new educational technology, they might be fun and engaging in the short term but will lose their effectiveness as motivators once the shine has worn off.
But the models might become better motivators as they become more multimodal, so that they are able to respond in real time to the student’s tone of voice and facial expressions, and as they acquire longer context windows, so that they are able to adapt their interactions to what the student has done weeks or months earlier.
The biggest issue, I think, is what the OP raises: Whether the bots will be able to—or allowed to—acquire sufficiently human-like personalities and indentities so that they can motivate learners as human parents, teachers, and mentors do—by making the learners want to please them, to be praised by them, to avoid angering them, to emulate them.
At school, there existed some classmates who hated math as a school subject, but nevertheless loved my ramblings about mathematical topics (well, they were at least more interesting than some other school subjects). Nevertheless, I guess my style of motivating people to like and do math would not be loved by parents: it it rather of the style "I'll explain you stuff about plant and process engineering so that you can build a weed farm that will be harder to detect by the police" for math, i.e. explain what subversive stuff you can do if you know math. I don't want to go into the details here.
This is deeply motivating for particular kinds of kids (with punk traits) to become quite interested in mathematical topics, but of course many parents would hate it because my teaching methods for math turn the child into a bad citizen. ;-)
While I agree that this approach to learning (with a bot) may not be a good fit for everyone, I am absolutely sure it can be for some.
I don't see the harm in exploring this path.
This is not like education was a solved problem, it is extremely wasteful, inconsistent and costly.
I think there is a path for more projects/interest based, less competitive and more individually tailored learning system when we remove all the constraints attached to a classroom and teachers, many of them being incompetent or miserable.
We had an economics teacher, that when you would bring up "the pope" topic, she could go on a 1+ hour rant about what is wrong with the whole catholic institution, etc.
Of course us kids, would somehow be able to get this topic into the lesson, to trigger her rant.
If you want to know various ways on how to break AI, let some kids, who are forced to interact with it, loose on the thing. I'm sure they'll figure out how to get it to say the most crazy shit in no time.
And only barely an inconvenience.
Also solar panels in the floor lmao
_solar freaking roadways_ ptsd
The idea that one could be taught physics by the virtual Feynman... oh my! There will be no better live person in the next hundred years, who could be better at the task than even fake Feynman.
Yes, I'm fully aware that the tech is not there yet and there are miriad of both technical and social problems to be solved but the mere idea of it really excites me.
This can play out in two ways:
- Teachers free up time/make time to give more personal tutoring to kids
- Teachers do this but schools start hiring fewer of them in the expectation of doing more
I don't think you will find anywhere on earth where teaching is a job with a great job market, perks or benefits. It's underpaid, underappreciated and understaffed.
People wielding AI like a bulldog at a glass store should be more mindful of the socioeconomics. But I'm not gonna hold my breath for that.
I was surprised to see the relatively high salary in Luxemburg, Germany and Switzerland, too: https://read.oecd-ilibrary.org/education/education-at-a-glan...
Actually, this seems to be absurdly beyond any of the recent progress in generative AI. This sounds like the kind of thing people say when their only deep knowledge is in the field of AI engineering.
The example is obvious marketing hyperbole, of course, but it's just not going to happen beyond a superficial level unless we somehow create some kind of time-travelling panopticon. It's marred by lack of data (Feynman died in 1988), bad data (hagiographies of Feynman, this instance included), flawed assumptions (would Feynman even be an appropriate teaching assistant for everyone?), etc.
I wonder if AI fans keep doing this thing in hopes that the "wow factor" of having the greats being emulated by AI (Feynman, Bill Gates, Socrates, etc.) will paper over their fundamental insecurities about their investment in AI. Like, c'mon, this kind of thing is a bit silly https://www.youtube.com/watch?v=og2ehY5QXSc
I mean, the guy isn't saying that it's going to 100% happen. He's saying that the problem feels like it might be doable at all. As Andrej has a background in physics, the phrase of 'feels tractable' would mean that he thinks that a path might exist, possibly, but only a lot of work will reveal that.
For instance, Synthesis[0] is an instructor designed, AI supplemented site for early math. https://www.synthesis.com/
It really seems like the distinction for these kinds of AI-education ventures comes down to the human educator(s) involved.
These would be charcutiers of these people with nothing but their most famous equations and examples and quips letting you know "this is supposed to be a famous person"
I think you mean caricature, but I do love the idea of Feynman making a mean charcuterie board
Great teachers know how to motivate students and keep them excited, which is crucial because many people give up during tough times.
But doing that at scale wasn't possible - until now.
Self-learning platforms had low completion rates, and cohort-based courses (CBCs) were seen as the future during COVID. But it turns out drop-offs are similar in CBCs since everyone learns at a different pace.
If we can learn from AI versions of the best teachers, it would be a game-changer. We could see massive improvements based on outcomes.
This is the closest we've ever been to matching the results universities can predictably produce. It'll be interesting to see if this works.
I'm a process oriented learner. While I can read any fact or trivia and retain it with no effort, things that require steps I only really learn by repeatedly doing them. So if there's an AI, for example, that can create problem sets for me to do with something that I am attempting to learn, and then explain to me the steps after I've attempted to find the solution, then I am more likely to retain that knowledge in the long run.
This would have been especially beneficial to me in my college statistics class, for example, where I had a prof who was... not much of an educator.
imho ai-based education is, in its own way, as hard to truly achieve as self-driving cars. the mechanical aspects can be mimicked for both with the sota, but the ideal version requires something intangible (call it agi shit for all i care).
i wish the hype and his reputation helps him get going for this long-term goal.
Btw, the founder also came out from Tesla, he helped Elon founded the school which taught kids from Telsa employee familys, I am wondering if they know each other lol, I think yes.
Kind of ironic an AI isn't building it. But that's an example of the current state of AI being more a remixer of knowlegde than a planner of actions.
> The teacher still designs the course materials, but they are supported, leveraged and scaled with an AI Teaching Assistant who is optimized to help guide the students through them.
So the AI isn't expected to build the materials.
Kids everywhere will be learning nothing except how to hack these chatbots to get passing grades.
Parents- get your son an AI girlfriend with a strong degree.
Karpathy’s company scales in a different way because students can get an individualized learning experience. If the product is excellent they might get a hundred million customers if the price is low enough.
I love Neal Stephenson's "The Diamond Age," but can narrative-based learning work as the main way to instruct a student?
Education is a part of what gets adopted into each and every student's self concept (identity). There is an uncanny valley in communication where any inconsistent distortion of reflected appraisal can lead students down a path of irrational madness, in a way that they themselves cannot perceive. This is well known in certain circles.
AI always has the biases which are programmed into it by the creators of the AI, whether this is explicit or implicit, the fact remains, there are biases.
AI in education will be proven to be harmful because there is no possible way that AI can remain consistent in its interactions at all times. We already know this from hallucinations where experts went and fact checked it and found it had lied, where lawyers were disbarred or censured.
Put plainly, this is a safety critical system involving our children.
Given the level of care to date that's occurred with AI in industry, there is no room for discussion or adoption of these tools, unless the actual unspoken intention is to drive your children crazy; where they self eliminate later as a result.
Rationally, good people don't create a world of intolerable suffering and then magically expect that somehow, someone, will figure out a way to fix everything you've broken but were too blind to see; at the time when knowing would save lives.
For those who enjoy cinema, a perfect analogy of this might be the Day After Tomorrow, where the main scientist turns out to be correct, but was ultimately ignored until devastating losses and risks were nearly insurmountable.
Leadership didn't want to hear the facts when knowing would have saved lives, and so they had to take a triage on the battlefield approach (abandoning half the country). While it is a fictional story, it sufficiently demonstrates what lack of scientific or rational foresight can lead to for those that can't imagine it themselves (people who are blinded like that should never be in positions of influence or power).
Please expand.
Hey, I don't have a twitter account (long story). Can anyone post those links here?
But it's Karphathy... and he's way smarter than I am.
We've collectively lost our minds with this AI hype
First of all, what is "AI"? I mean that in a literal operational sense: can you even explain the math and machinery that you're talking about? Because if you cannot, then you are advocating strapping children into Black Boxes (containing that rugby mascot!) that you don't even understand what they are, and that means you're nuts and should not be allowed anywhere near children.
Obviously, with Andrej Karpathy that's not the problem, but assuming that a person actually knows WTF they are even talking about with "AI", "AI in education" is still a bad idea.
First and foremost, consider the question of what even is a human being? And consider it in light of the contrast between us and our nearest living relatives the chimpanzees. What makes humans different than apes?
Well, it's thought and language and higher reasoning, and self-awareness, eh?
We do not fully understand how human brains do this, make these phenomenon happen. We do know that the process of becoming a human being is partially automatic (e.g. learning language) and partially the result of deliberate self-work.
The thesis of the "AI in education" people seems to be that this yet-mysterious process can and should be turned over to automatic machines, typically for economic (rather than social or philosophical) reasons.
Children taught by machine will think like that machine, to the extent that they think at all.
Metaphorically, it is as if these folks want to strap children into wheelchairs on the theory that that will make them better runners.
This brings me to the second objection: use it or lose it. If we rely on external devices to do our thinking for us our brains turn to jelly. It started with the written word, as noted by what's-his-name a couple thousand years ago, turning the faculty of memory into the mere faculty of recall. (If you don't know the difference between memory and recall you are doubtless the victim of just such a degenerate "education" system as these folk are proposing to intensify with automation.)
It's not difficult to learn to use your brain well. Unfortunately the knowledge of how to do this is still not recognized by mainstream conventional education systems. E.g.: you don't need memory aids if you just use self-hypnosis and mnemonics to "just remember". You don't need to learn to spell, you just need to learn the correct spelling strategy and then your brain spells for you, correctly, every time (you can even spell things backwards without any effort.)
There is a lot of low-hanging fruit in terms of ways to make education better, but automation is completely useless for it, because things that can be automated are not important human activities. Just as it's pointless to make a robot jog for you or eat for you it is pointless to make a robot think for you.
To sum up:
1) Becoming human is something only other humans can help you do. Machines can't make you human, they can only train you.
2) Using artificial aids rather than one's own natural faculties is regressive. (It's like using a wheelchair when you legs work fine.)
"AI in education" is oxymoronic, pernicious, and even dangerous. People who want to strap children into mysterious black boxes should be kept away from them (the children, not the black boxes.)
Like... is that a satirical choice or what?
1. Behavior:
From my experiences and in speaking to others, teaching any subject varies not because of the material, but the behavior. When behavior is a problem no teaching, no matter who or what is doing it will get done. Behaviors can range from sleeping in class because dad made you work at his out of home construction company till 2 or 3 am to get a project complete, or it could be your mom is tired of you going to school and instead wants you to “go hit up them streets because that’s where the real motion (money) at.” These two might sound like isolated cases, but they aren’t. There are many kids who live out of cars, sell themselves to get buy and other atrocious life circumstances. Learning usually can’t happen when you’re preoccupied with keeping yourself alive or afloat. Combining AI into the educational landscape can’t and won’t fix this disinterest in learning. You will have to fix home life first. The problems arise because behaviors, in the classroom, work on a spectrum. They are better some days and others not at all. This is because behavior follows home life. It is usually modeled from home and then is acted out away from home. When Nathan (name changed) see his brother and brothers friends carrying around guns, drugs, insert something other activity not safe for most students, he will naturally mirror the act. When these behaviors present in class, a good teacher can maneuver a class back into an educational focus. But being a good teacher requires many exhausting abilities such as knowing the current generations lingo, figuring out who is a natural leader in class, who are the other kids afraid of or respect, knowing their material and presenting it to the students in a way that fits their life circumstances. Being good at teaching is a multi-headed venture across multiple disciplines. This makes managing behaviors in class a challenging and exhaustive time sink, which tends to be the main issue with teaching being interrupted or difficult in a class. One child from a difficult place can make learning a challenge for a class full of goody-two-shoes. Teachers are unnecessarily pulled away because we have no other recourse but to handle and teach children who need more support than what can be offered in a class. Behavior is one of the main issues that many people overlook.
2. More Tech and AI tools than I could want:
When I sign into my Microsoft portal I am presented with more 3rd party teaching tools than I could ever want. I use none of them because they are yet another tool to learn and be trained on, they help mostly in presenting educational material, but have nothing to do with the flow and behavior management of class, they have their own technical problems that I don’t have time to deal with because behind my back Timothy has begun passing his shoe around the room for others to sniff, and the applications are gamified learning that trivializes or simplifies the educational materials. There are many more reasons I and most of the teachers I know and have met don’t use these systems. Work with teachers to learn how much teaching environments vary and how we can implement tech tools into that classroom in a way that prioritizes education. But know many teacher’s priorities will not be for VC attention, but because we want to educate student into society. Listening to us gives you the opportunity to hear and see what educating people is actually like rather than the inert surveys and data points collected on students that are disconnected from their lives experiences. Listening to the people with boots on the classroom means possibly making a system that actually does something for classes than just become another novelty the principal and the district can feel good about.
3. Education is controlled by the wider community and home environments:
Diane Ravitch, a leading thinker in the education realm, writes about all the changes she has been a part of and seen occur. In the end, her response to these failed reforms in school is that much of what happens in communities controls what occurs in schools. If you have a society that is splintered into social groups then that will typically be reflected in the classroom. First we must adjust society standards which will then alleviate the malignant issues in our schools. But as many teachers have said to principles and school officials, “Hey, what do I know, I’m just a teacher.”
His great instructional videos on YT tell me that he is passionate for both AI and education, so I'm all for him trying to mix the two. More effort on education is always welcome in my book.
[1] https://github.com/karpathy/LLM101n [2] https://eurekalabs.ai/
> However, with recent progress in generative AI, this learning experience feels tractable.
If you can create the Feynman AI, you shouldn't bother with an education company.
A sharp, domain-specific Feynman-flavored LLM would be broadly useful and still worth bothering with.
The thing is, when the AI art generator makes a mistake and draws a person with 3 arms, that is obvious to the student and they can take the output with a grain of salt.
But when the AI physics tutor generates some physics result that's the equivalent of a person with 3 arms, that will not be obvious to the student. They will take the words of the AI credulously. I see it all the time in programming as well, where the AI just invents APIs, semantics, and syntax.
I don't know how to solve this.
This is the same issue I have with large language model as coding assistants, since you're effectively not in the driver seat - you're acting more like a code reviewer, and I think that that passivity eventually causes critical observational skills to atrophy.
I am very excited for this and can't wait to try their products!
As an aside, odd to see people here are so pessimistic about the idea. Education is hard to scale because tutors are hard to scale. ChatGPT is already an ad-hoc educator for me, in a similar jump that Wikipedia was. Having a purpose-built product with a hand-crafted curriculum, that's a killer idea.
Instead we just accept really bad education.
Until a chatbot is provably correct and ethical in its output, it must not be used to teach.
Case in point: the slop image attached to the announcement has the typical malformed hands and ghoulish faces problems.
I like this idea. What if you could study geometric function theory with Kaczynski? You could study chemical engineering with Thomas Midgley Jr!
Creative writing with L Ron Hubbard!
But if you opted for the mega ultra premium gemini pro max++ model, then you get a minimal ad free experience. No wait, forget a monthly subscription. Think, micro subscription model on a per usage basis.
In your death bed and need life saving measures? Augmentation returns "402 Payment Required" before nanobots can proceed to excise the root of the issue. 5M SHIB tokens required. Unfortunately, 20G signal is not reliable in the Wilderness Zone, thus old school EMS services are dispatched to the scene. Unfortunately, EMS services do not accept your insurance and private EMS leaves the scene and dispatches public EMS services. The wait is 2 hours given the WZ.
You fail to receive life saving measures (ie, tPA) in time, thus resulting in impaired motor functions and decreased quality of life. The flashing 402 Payment Required prompt is forever imprinted in your augmentation.
Well obviously the target audience is actually AI-agents as part of the Super Intelligence ramp-up!
To get an AI-based program adopted would be an uphill battle against all forces. Almost Herculian, I would argue, and almost guaranteed to not be worth it.