Edtech has a core problem: Learning requires effort, and effort leads to disengagement. I believe this problem exists, but whatever the "fundamental limit" is, we're nowhere near it. It's well known that paying for an education resource leads to higher engagement... But the @RareSkills_io ZK AI Tutor is getting incredible student engagement on the platform even in the free phase, and almost none of the "engagement mechanics" have been used yet. There's a lot that went into this, but one of the core principles is "question design." Every question is free response and evaluated by an AI. A Question Should - be at the right level of difficulty - draw out the thought process to maximize the chance of discovering a misconception early - help the student learn the topic - give a sense of control and challenge - (other things not listed here) Now, if you can give the student those kinds of questions on repeat, then you get high engagement. Fundamentally: Can you be an interesting "person" to talk to as a tutor by asking good questions? If so, the student gets engaged -- AND genuinely learns. Engagement rates have greatly exceeded my expectations, and what's cool is I really think it's possible to do at least 2X better.

When we say "the results have been amazing" frankly that's a bit of an understatement. In the most extreme case, one tester with only a high school education wrote their first ZK-SNARK in less than two weeks of intense study. I've seen extremely smart people need a month to do that (with a lot of guidance), so the efficiency gain is nuts. But the consistent outcome I see in my weekly calls with users is that students can explain what they learned for 10-20 minutes straight, correctly, uninterrupted, and without preparation. And they aren't just repeating what the app is asking them to memorize because they can answer "out of distribution questions" about the subject. This is a MASSIVE improvement over the typical response students give in a classroom when ask to "what did you learn in the lecture/homework?" Students giving talks for 10min+ straight unprompted is a HUGE improvement over the typical book/lecture/homework format. --- Building the RareSkills AI Tutor for Zero Knowledge Proofs was far harder than I thought (coding was easy, thanks Claude/Grok/GPT) 1) Asking good questions is the meat of goodteaching. A good self-directed learner (autodidact) asks themselves good questions, and this informs the direction of their practice/research. Most of the effort of building the app was coming up with questions that make the student engage with the core of the subject rather than tangentially "think about" it. Also, questions need to be carefully calibrated to the student's (predicted) level. A question needs to be clear, optimally difficult, and help the student build a mental model of the subject in the background. Sometimes this means mechanical drills, sometimes this means simply repeating the definition back, sometimes this means asking the student to connect concepts... the space of possible questions to ask is massive. AI could not help with this at all. I had to draw on my experience of face-to-face teaching to know what open-ended questions work. Math questions with a definitive correct answer is easy, but questions that "get the student to think correctly" were challenging. AI doesn't come up with the questions, it's just too unreliable. Some questions the student will encounter are AI generated on the fly, but this can only happen in very specific circumstances with tight guardrails. Asking the right questions in the right sequence is 80% of making the app work well. The other 20% is not screwing up the rest of the implementation, which is the next point... 2) Review algorithms are very hard to get right. The common failure case from spaced repetition is "review overload" and the first version of this app fell into that trap. You cannot expect students to remember everything you teach them. It's your job to determine what is the ultra-important information and drill that with correctly designed questions (see above). The "easy part" is modeling how well they remember specific topics and how good their overall memory is. However, asking the student to memorize something comes at a very high cost. This is why spoken languages are so hard to learn -- you literally have to remember thousands of words and when to use them. So the course needs to be extremely selective about what it expects students to explicitly memorize. The "memorization budget" must be spent wisely. 3) Getting questions wrong is very demoralizing for students, but the solution cannot be to make every question easy. One seemingly trivial, but very important feature in the app is that it gives the student more than one attempt at a problem (depending on the nature of the mistake). Most edtech platforms mark the student wrong, show the correct answer, and move on. But what if the student made a silly mistake? What if the student had incomplete justification? Allowing students to "learn from their mistakes" rather than "punish them with a grade" seems simple, but it's actually (excuse the AI-ism) gamechanging. I think educators are too socially RL-fried into thinking "homework should be graded" that it shows up in how edtech platforms get buit. Making mistakes is part of learning and it provides an incredibly valuable signal into the "learning state" of the student. The app also applies the philosophy in reverse where appropriate -- the student can get the answer right, but if the student's thought process is wrong, the app forces them to correct it before considering the answer correct. In fact, right answer, wrong justification can be marked wrong if the student can't justify the answer with escalating hints. Testers have repeatedly said that getting evaluated on their thought process rather than just the correctness of their answers has been a very positive experience (and it's given me a lot of valuable data). This does not mean correctness is unimportant. But the app strongly weights getting things correct AFTER you learned it rather than while you learned it for the first time. The grader is quite punishing if you don't remember key points from the lesson during later reviews, and that's important. What we care is that you REMEMBERED CORRECTLY not that you cruised through it the first time. Having an environment where "there is no grade, but I want you to get the answer right" makes learning far less stressful and productive. The app fully leans into this style of teaching during the "free recall" sessions (you have to try it to experience it, I can't type it all up here). The broken aspect of homework is that grades are used as BOTH and incentive and feedback mechanism, which means they don't allow for mistakes, exploration, and non-scary feedback. This app nicely solves it by separating incentives from feedback. 4) Active learning does not replace videos and articles. A common feedback I got on an earlier version of this product is that it would benefit from videos. I resisted this feature because I thought the exercises should be so good videos are unecessary (also, writing scripts for videos is time consuming and AI takes a long time to make animations). I was falling into the same trap people who are religious about not commenting their code are. Yes, code should not need comments, but good code has good comments. A good course has good exercises, but passive learning has value if used correctly. A common failure case with videos is that people watch them and then fool themselves into thinking they understand the topic. So in this course, instructional videos unlock AFTER the exercises are completed. Usually, after someone has done a bunch of exercises, their brain is a little cooked. So watching the lectures is a good way to take a "motivational and productive break." Lectures "consolidate" information rather than "teaching" it. Oh, and the lectures feature a lot of sick animations, so they're definitely fun to watch. 5) Whatever an uber-smart AI says is irrelevant compared to empirical data. AI will have convincing explanations for why a question is "redundant" but then you see some students actually struggle with the "redundant" question. It's important to treat AI as a grunt-level teaching assistant rather than as a professor. It's just too likely to fly off the rails into some random pool of training data that does the opposite of what you want. --- The app does things slightly unconventionally: lectures come after "homework" and "homework" is not strictly graded. Then of course there is the RareSkills touch of being very selective about what technical terminology gets introduced and the tendency to prefer intuition over rigor. For those wondering, the app doesn't teach ZK-SNARKs the same way the ZK Book does. The app is a lot more thorough, incremential, and digestible. When you restrict yourself to communicating in 5 minute increments, your mental model of the subject must be extremely granular. The ZK Book has 5,000+ word chapters. The book is clear, but granular and incremental it is not. So the "framework" the app uses is quite different. Some examples: - QAPs are taught *before* Rank 1 Constraint Systems in the app. This is highly unusual, but we've found it scaffolds from polynomial prerequisites better. - The app teaches a version of Pinocchio before reaching Groth16. The app treats Groth16 as an evolution of less efficient algorithms that the book doesn't cover. - The treatment of abstract algebra is significantly better than elsewhere. This is the first time I've seen people get homomorphisms on their first encounter. The teaching sequence is really effective. In fact, I kid you not, an early student on the sunsetted v1 platform was able to explain homomorphisms after only five days of using the app, and that person did not have a math background. If you know anything about teaching homomorphisms, then you know that outcome is (again excuse the AI-ism) *wild*. The tester didn't even know what a "group" was when he started, then five days later he could do basic reasoning about homomorphisms and explain it. --- It's hard to boil down to one thing that makes the app so great... it's more that it does so many things almost flawlessly. 1) At the early stages, drill the basics to get key definitions and concepts down. The app actually expects you to do some math by hand and memorize definitions. These are not fun things to do, but the app (successfully) frames these exercises as "building intuition" rather than "grinding." 2) Later, ask questions that draw out the student's train of thought, then give feedback on the train of thought. Giving feedback on the train of thought, not only the correctness of the answer, is a FAR more efficient teaching mechanism, and the results show. 3) The app isn't stuck to a certain way of teaching. Sometimes it uses direct instruction, sometimes it uses Socratic instruction, sometimes it leans on worked examples, other times it uses inductive teaching. Different topics and different student "learning states" benefit from different approaches, and I think we generally got the combinations correct. 4) The app uses spaced repetition (as any good edtech platform should) but spaced memorization is not the core memory tool -- everything the student learns is part of a whole. "Isolated facts" were aggressively pruned. The spaced repetition algorithm itself adapts to students very quickly and accurately. Students consistently say "I'm getting just the right amount of review" and statistically, students get review questions correct ~90% of the time. 5) The "free recall" technique the app uses is so visibly effective that students tell me "I'm going to start using this in other areas of my life" 6) The coding exercises are designed in a way that is AI-aware. We know nobody writes code by hand anymore, but there is value in writing code to test your own understanding. The coding exercises are carefully designed around this reality. 7) What made RareSkills great was the quality of the explanations on the blog. This DNA still carries over to the platform. The app's explanations are crisp and easy to understand. Interactive tools are used where appropriate. Obviously a good explanation is only the beginning of learning, but it's important and it's done right. Most of the mediocre explanations have already been pruned out by the testers. 8) The app has a very graceful recovery mechanism if you've been off the platform for a while and you aren't sure what you remember. Starting to learn again after taking a two week break normally feels daunting, but on this app, taking a break can actually be productive, and re-entry is smooth! I'll admit this feature isn't perfected, but it's FAR better than anything else I've seen. 9) The app was heavily shaped by over a dozens of people who give really amazing feedback (both explicitly and implicitly). And feedback is still continuing to shape it. To give credit where credit is due: - The YouTube channel "Benjamin Keep, PhD, JD" had some really fresh teaching ideas I definitely borrowed. - Mathacademy opened my eyes to how far you could push "learning through active learning." I thought the exercise-driven approach in our bootcamps was well tuned, but I had no idea how much further the concept could have been pushed. - The substack carlhendrick has some brilliant pieces about gaps between "education research" and "education in the real world." His work actually made clear to me why some aspects of an earlier version of this app were not working out even though the app was doing things correctly "in theory." He really drilled that "just because the student is producing correct answers doesn't always mean they understand the subject." Fill out the form, and get your access code. As long as you remember (or can quickly relearn) how to divide polynomials and multiply matrices, you'll be able to take the AI tutor for a spin.

The most powerful learning intervention I've discovered is putting students into an ice bath and not letting them out until they get the answer right. Beats spaced repetition and AI analysis of past answers by approximately +1.5 SD. Sign up here: RareWeek.RareSkills.io Learning tracks are Zero Knowledge Proofs, ZKVMs and Advanced Rust Programming. (you can also just cowork in a nice villa with high-caliber people if none of the learning tracks interest you)

Learning is primarily a social activity. This is why, even though classrooms are not effective, they have a far higher completion rate than self-paced moocs. This isn't unique to learning: - gym = build cardio/muscles with buddies - office = work on a project with coworkers - clubs = activity + social Anything difficult requires a social dimension if you want a high rate of success.

Explaining a subject in your own words is the most efficient way to learn a subject, and it continues to work even as you get really advanced. It forces a durable encoding of the subject in your own mind (you memorize it better) and it exposes your knowledge gaps. Explaining in your own words also improves your ability to coordinate with other humans and AI.

I propose that we pace blockchain education. We don't want blackhats (or worse, AI) training on RareSkills articles and stealing funds. We will be giving an independent third-party employee access to our drafts to ensure that we move at a safe pace until we make sure all our readers are aligned.

Back in the day, when only the Moon Math Manual was around, it would take engineers -- even those with advanced math backgrounds -- nearly a year to understand ZK end-to-end. When we published the RareSkills ZK Book, the typical learning time dropped to ~3-4 months. Now, with our new AI Tutor, it takes weeks -- or a single week if you are very motivated and prepared. A tutorial series is necessarily one-size-fits all, but interacting with an specialist AI cuts the time even further -- especially because the practice part of learning is closely guided. Explanations of a subject are not usually the bottleneck, but rather how efficiently you can "encode" new information permanently. This means carefully choosing what subtopics to understand, and how to go about memorizing it. Interactive platforms can do this better than any book. So I believe this is the future of learning: in-person events for maximum focus and peak efficiency AI tutoring. Oh -- and since part of the Navier-Stokes solution was likely... how shall we say it... "borrowed" -- ZK is going to play a greater role going forward. So get started now.