Khan Academy integrates GPT-4 as every student’s customized tutor

Do you know if/how someone can get involved with this at the research phase? I have a daughter who will be entering 1st grade next year and I'd be interested in having her try this out if there was a way of signing up.

I was a high school English teacher for about ten years before transitioning into tech last July. Has KA explored assistive assessment tools for in-person instruction?

It better not be for every student. I am very familiar with Khan Academy as I am currently guiding a student through several AP courses on khanacademy. In my opinion, khan academy's time would be better spent fixing the UI for teachers, and improving the organization of the physics curriculum.

I would prefer that khan academy not be dragged into some PR fluff piece for some AI shill.

>it's not for every student (at least not yet!)

>Donation required after chosen from waitlist

So, it's for rich ones?

Chatgpt is already more accurate than a lot of my high school teachers.

They're saying that they have reduced hallucinations. But RLHF seems to undo some of the improvements (figure 8 in the paper).

Its more troubling because its being used in a role as an educational resource. the audience is learning and significantly less able to realize that.

Firstly, the unreliability may not be permanent, and we may improve the AI accuracy in the future. Secondly, we need to figure out what works and what not, and we're in the middle of such phase. Finally, there is no need to be alarmed, even with hallucinations, the amount of advanced (and reliable) knowledge provided by the AI vastly counterbalances the small mistakes. The elitist mindset needs to die, let's give access to advanced knowledge to everybody and stop putting it behind some unreachable walls.

There are two indicators that this is PR bullshit produced by whoever is trying to capitalize their investment, 1) is the use of the marketing term 'AI' or 'AI-powered' and 2) use of the 'think of the children' trope.

Don’t expect it to know specific facts but it is excellent at explaining topics

That is a feature, not a bug. It's an efficient way to brainwash a population.

Are the tolerances in education really that tight?

How many terrible teachers are allowed to continue to teach after decades of disastrous results?

How many ideologues with no interest in teaching, but every interest in indoctrinating young minds, are tolerated because the alternative is "no teacher" and the class wouldn't run?

How many teachers who would fail state exams, teach, despite relying on answer sheets to be "competent"?

I agree with you that on the top end of education, this is no replacement and at best a supplementary tool. For the poor kid in a bad neighbourhood whose teacher is more interested in "de-colonizing" mathematics than teaching mathematics, this is a Godsend.

I mean that is but one thing to worry about, we've gone about nowhere with the alignment problem, and we're screaming ahead at full speed with making these things more powerful.

Without providing context or examples, GPT-4 is already better at answering questions than the average teacher, with the unlimited patience that only a computer can provide.

With context, which Khan academy has in abundance due to their lesson plans and transcripts, accuracy will be higher than even the best teachers and tutors.

Once you give context and known-true facts to best-in-class LLMs like GPT-4, the output is shockingly good.

If you don't know how they are using GPT-4 it's not fair to say it will hallucinate.

As far as I understand the preferred way to use LLMs nowadays for domain specific information retrieval is through embeddings that insert the related context in the prompt. GPT-4 is specially good for this since they increased the prompt size almost by an order of magnitude.

This means that the model can be given a very specific task: to extract information from the context or avoid providing an answer at all.

The answer doesn't rely on the neural memory of the model, since it doesn't need to store information, just understand the task, and they are really good at that.

My mind went to the same place. Knowing that "the average tutored student was above 98% of the students in the control class", imagine how transformative it will be to lift the average student to the 98th percentile!

This may be the most transformative change to education since the invention of public schooling.

Since using T5 to generate quiz questions in 2019 I've thought this. Transformers now can understand misconceptions, quote source materials... it will equalise education.

Every student might get an absolute boost, but the relative gap may even widen. Consider students who can afford to use cutting edge, expensive to train/run models, vs students who can only use off the shelf, older models.

Circa late 2015, early 2016, my cofounder and I spent ~6 month attempting to build exactly this.

Our hypothesis was that students self studying for the SAT used books (eg Kaplan) but the solutions in the back were poor in explaining how they arrived to the solution because they usually just cut to the chase rather than explaining the approach & helping you arrive to the solution.

So we took each practice SAT math question and wrote out a solution spread over four swipeable pages. Our hope was that if you saw the first or second page, you'd arrive at the solution without having to see the answer, thus increasing your chance to solve it on your own next time.

We spent over 100 hours tutoring kids and had them use our MVP to mild success. The obvious next step was to automate the solutions rather than hand write it ourselves - like GPT-4 has done here.

As someone who has logged over 2000 hours teaching, there's a lot of soft nuances and tactics to tutoring. Also, motivation is a big thing a good tutor brings. I don't see this solution as quite there yet but getting close.

Very optimistic here.

It's not the tutoring alone. It's also the social effect. A machine simply does not exert that kind of influence, nor do online courses require the same amount of dedication and ambition. Compare MOOCs: they have incredibly high drop-out rates.

A tutor doesn't just imply parents who can afford one but also who care enough to do so in the first place and see the value of an education.

I think you'd find that even by the sat-prep age the actual limiting factor for many (most?) of the tutor-less is a fate-accompli attitude towards their path in life.

Might as well turn on the chat bots at the cradle. Ah, and make sure everybody gets the same quality AI to parent them for it to be egalitarian.

On the other hand, I think it could also trigger an arms race, and might very well incentivize learning how to write a good prompt more than learning the concept well enough to succeed in a class. This reminds me in my era of the divide in success in my era, between students who were able to navigate the internet to the benefit of their academic success, and those that weren't able to do things like quickly paraphrase a wikipedia article on the subject and use the wikipedia citations as their own, find a relevant flashcard set someone else made and shared on quizlet, or even find pdfs of the teacher's solutions accompanying the textbook. Cheat codes are cheat codes. They will help in terms of your grade but usually hurt in terms of understanding.

Human teachers make mistakes too - perhaps even at higher rates than GPT in some cases.

Instead of isolating a single factor of hallucinations I think you have to also consider the cost and experience quality and make a more holistic judgement on whether these models are good or bad.

I would like to divert your attention to this recent post: https://openai.com/research/gpt-4 specifically on the topic of "Steerability: Socratic tutor" (Sample 1 of 3)

This is something I’m working on. I think having a good understanding of the material for good explanations is doable in the near future, but the main problem is actually building a good model of what the learner knows to guide the LLM

In my brief asking ChatGPT question and to rephrase/clarify, it seems at least as good as the average tutor/teacher/TA which for me has been marginally better than warm body rephrasing the theories and working through textbook problems in front of you.

And in time, much better. To the point of raising the standard of education for all of humanity with access to the internet. I'd count on it "working out".

My daughter does children's martial arts. Often in the dojo there's an atmosphere where the yellow belts are helping the white belts, the red belts are periodically chiming in, and the black belt adult instructor is free to float around the room supervising everything, giving individual attention where it's most effective, but leaving the lower ranks to take care of the lessons they're capable of giving.

In the same way, AI will be able to make an expert human tutor 100x more effective by escalating the tricky cases but learning how to handle the ones that come up over and over.

Ractors for hire!

Once AI can generate videos, we just need pervasive surveillance capability to insert context-aware content into the stories.

The Primer was able to identify the protagonist's family member and pet toys in its personalized myth.

I wonder which TTRPG sourcebook will have the first credit to ChatGPT in DriveThruRPG :)

Procedurally-generated worlds plus conversational NPCs!

Khan Academy's integration of GPT-4 already steers students towards particular modes of thought. According to the article, they want "to get students thinking deeply about the content that they’re learning" and in the examples, Khanmigo prompts students to recall information from the lesson rather than directly explaining how to solve the problem.

I think what Khan Academy has done here is desirable, but just as Stephenson's Primer enforces its creator's values through AI tutoring, Khanmigo is enforcing Khan Academy's values. It's easy to imagine how someone with more authoritarian values could use this same technology (e.g. to teach students to follow instructions without doubt) to indoctrinate students with scale.

See also: the Mind Game from Ender's Game

https://enderverse.fandom.com/wiki/Mind_Game

For younger readers Monica Hughes' classic "Devil on my back" is more explicitly about a society where those who can interface with the web via computers installed into them become higher caste and those who cannot become lower caste, resulting in a revolution. Definitely aimed at a younger audience than Stephenson.

>It's pretty fascinating to consider the implications and whether that's something we'll want, bc it may soon be possible.

I wouldn't be surprised if it told people to take some path of least resistance to $10+ million so you don't have to work to survive, then you can mostly do whatever you want.

(I havent read The Diamond Age)

The eventuality I see is everybody talking to each other through LLM's, and the world ending up in some kind of tower of babel situation. You wear an AR/MR headset, and all conversation is dual translated, with some kind of bytecode intermediary. Each persons language evolves to be incompatible with anyone else's, and only your LLM can understand you. The power goes out, and nobody can talk to anybody.

Having heard many recommendations, I tried reading that book more than once, but have real issues with Stephenson's writing style there. Could you please recommend an overview of the qualities of the AI tutor that I could go over instead?

My daughter is in 2nd grade and math is her favorite subject. I liked math too. I've looked forward to helping her learn more advance math through the years, but a doubt has entered my mind; will I have any reason to help her?

Xaeon-12 wakes up at 8am, has a bowl of cereal and put's on her VR goggles. She is greeted by her tutor Tom. Tom will guide her through advanced "AI prompt" class today. Xaeon-12 learns to create her own short story based on her dreams that were recorded the night before. 12am, time for lunch. After lunch, the kids all connect to play Roblox VR for 30 mins. Suddenly, the doorbell rings, but no one else is home. Fortunately, Xaeon-12 doesn't have to answer the door, as everything is automated. The mailbot dropped off some packages, before flying back to the zeppelin mothership to retrieve the next package. Xaeon-12 is just finishing her last class, "circuit boards", when her dads come home from the golf range – with an abundance of free, renewable energy, working for a basic income has become irrelevant.

Honestly it depends on the growth rate of our models and if they converge with human intelligence without requiring have the computing power on the planet.

If we get slow incremental progress over the years then it will act like an interactive personalized encyclopedia that can teach you based on your strengths and help bolster you where you are weak, as the technology improves that is.

Now, it's still scary as hell in some ways even in this slow growth mode. The authoritarians will love to have a system like this that just forgets particular parts of history and make reports back to the mothership if you ask too many 'wrong' questions. So actual implementation will be interesting to see.

Now, if we have explosive growth to AGI and beyond, then pretty much every bet about the future is cancelled and we'll have more luck in taking random guesses, gluing them to the wall, and then throwing darts at them blindfolded.

You get the full experience reading that story if you have deadmau5's the veldt playing in the background as well

A Young Lady's Illustrated Primer

Another one that came to my mind: "The Fun They Had" by Asimov

"Margie was thinking about how the kids must have loved it in the old days. She was thinking about the fun they had."

http://web1.nbed.nb.ca/sites/ASD-S/1820/J%20Johnston/Isaac%2...

Replace "AI" with "powerful institutions" and you hit it on the head. Maybe "AI" should stand for "augmented institutions".

Maybe you'd be interested in Paulo Freire's book "The Pedagogy of the Oppressed"?

There will probably be a janky but usable open source version of the model and tool eventually. The "Linux" option. That way the 1% that cares about this stuff can do the self host route and retain control by them controlling what the model tells them(hopefully).

A plot point in an episode of Person of Interest, where AGI "Samaritan" funds a charity (free tablets for students, Samaritan access pre-installed) to take over education and recruit more mercenaries.

They literally announced that they're testing this on a small sample of kids whose parents opt-in. I don't see any better approach, seeing how the sentiments of "untested" and "using kids as guinea pigs" are contradictory.

I mean, at the same time could that not also mean that people are going straight to ChatGPT to ask the question in the first place?

I don't think this is such a problem if it is teaching calc-3 and stays on topic. If it's teaching humanities I can see there would be more of an issue.

And that is any different from the school books we currently use how?

I don't think tests will work at all like that much longer; they'll be more like oral exams, where you have a back-and-forth conversation with an AI that probes the breadth and depth of your knowledge with sufficient variability as to detect overfitting (i.e., memorizing the answers to a few previous exams).

I wonder what the point of school is even going to be at this point

Looking after kids while the parents work

I think GPT-4 has made substantial improvements here.

> SAT Math: 700 / 800

It will get there, at least it should solve the easier problems correctly

2 of the 3 demo videos is literally a student asking it arithmetic problems

There's other models that are doing this at a much better rate than GPT, so I don't think this is an unsolvable problem.

In this case, as I understand it, the idea is that the existing Khan Academy system generates a random question, the AI helps the student solve it, and then the student provides the answer to the original system, which then validates it.

I would also assume that Khan Academy could inject the true answer into the GPT prompt and instruct the AI to hide it from the user, but rely on it in its explanation, in order to reduce the risk of misleading outputs.

I mean honestly, if I took the corpus of our companies support data and used it as a training set for our first line support, I believe that GPT could answer the questions better than they do first shot.

No idea what the cost difference would be, but for many businesses I could see the AI response be a costs savings.

I'm currently homeschooling two teenagers and one of the most interesting aspects for me is learning all the material that I 'supposedly learned' way back when I was in high school.

I graduated, went to a great university, and yet virtually all of the stuff that we are learning in the current HS curriculum is new to me despite my having taken the same classes. I honestly don't know how I graduated high school, except to consider that I was very good at parroting back recently received information.

Most of what students are learning will be obsolete by the time they enter the job market, but on the plus side they will have forgotten it anyway. The most important thing from my perspective is ensuring that those topics don't use up too much of their time as I think most intelligent people are able to acquire the knowledge and skills they need as they come to need it.

I assume the one from the research page would work:

"You are a tutor that always responds in the Socratic style. You never give the student the answer, but always try to ask just the right question to help them learn to think for themselves. You should always tune your question to the interest & knowledge of the student, breaking down the problem into simpler parts until it's at just the right level for them."

From: https://openai.com/research/gpt-4

Learning is absolutely going to be required, we are just going to need to spend more time learning patterns and motifs rather than lots of facts. Learning facts and trivia was already devalued by everyone having a hyper-encyclopedia in their pockets, this will just extend that. Knowledge of patterns and motifs in domains of information will let you work together with the machines in a way that an untrained person will not be able to match.

I don't know why comments like this take for granted "basic income" as like this inevitability. We currently let people all around the world die on the streets for not participating well enough in the economy. Why would this moment in history be any different? You think just because you get paid a lot right now, the guys in charge will make sure you are ok in the future? That is not really how this whole capitalism thing works unfortunately.. No amount of technology will save us from waged labor in itself, its simply the character of our labor which will change.

What you want is communism, and buddy, I'm with you!

I am not even that optimistic about it. What would be the incentive to keep humans entertained in a world run by AI? Why would they keep billions of us around? I won't make any sense economically.

When I was a kid, I remember one of my teachers saying the same thing about the internet ( google, computers, etc ). Us kids were so lucky to have such powerful tools and everyone in the future will be smarter for it. Turns out, people are dumb as ever and these tools made us lazy and dependent. Why memorize the multiplication table when you could ask siri?

I don't know that I would classify Khan Academy as a tech giant. For the sake of argument let's say they are. Yes it is fair for them to get access earlier. Why would it not be?

Surely on HN of all places we understand that life is not fair. OpenAI is looking out for their own bottom line. That means money, and PR.

Because tech giants can write checks that don't bounce (off of SVB).

I don’t think Khan Academy would be considered as a Texh Giant. Or are you referring to OpenAI, the creator of GPT-4?

I wouldn't really call Khan Academy a tech giant. They are a charity.

That said, it does not necessarily seem like OpenAI is trying to be fair.

I think it's fair but there's a reason why it almost seems natural to raise the question in this scenario. It's OpenAI's hypocrisy and doublespeak. It's their non-stop messaging about equality, alignment, unbiasedness, etc. They can't have it both ways and they will be judged for it.

I'd assume that's because the Gates Foundation is one of the biggest financial supporters of Khan Academy, and they can reach the right ears at Microsoft and by extension OpenAI. As always, the right connections are everything

Well Microsoft bought OpenAi and Khan Academy is owned by Bill Gates so...

All products need testing before release.

OpenAI is an opinionated startup, they don't owe everyone the same things. Probably they thought KhanAcademy would be a very ethical way to use GPT-4 for good and they gave it early access.