I'm curious to see how Alpha School develops over the next few years, both in terms of the way we can use machines/technology to create optimal support for learning and mastery, and how any potential success might generate possibilities for what schools do. If (and it is an if at this point, not a given) machines can do what you have outlined, then the question becomes what do schools offer beyond this? One outcome might be one where schools can lean into their mission or identity with greater clarity. Texas Sports Academy is an example. And I think Zach Groshell, in his conversation with Craig Barton (https://www.youtube.com/watch?v=ZZz4HCtHWxI), talked about his choice as a parent as to the type of school he wanted for his child. I think this type of future might be possible, but it isn't hard to see challenges that might come when you attempt to implement this at the level of a public school system.
My other thought is similarly badly formulated at this point, but I'll take a risk and share it anyhow. I wonder whether, once we control for all the cognitive elements of learning (CLT, attention, scaffolding, spacing, interleaving, etc.), Alpha and the researchers will be as interested in looking at the affective side of learning? The interaction between confidence and competence feels like something Alpha could look at over a period of time, and we might learn something interesting and of general value. I wonder about the variation between affective learning when at a screen engaged in mastery learning versus affective learning during the other six hours of the day, when not engaged in carefully scaffolded learning?
This is a brilliant essay, as always, Carl - I appreciate your time writing it. I don't always agree with everything but you've connected with a big tension I have been feeling lately with AI in the classroom. It feels like something to protect our students' learning from AND a missed opportunity if we keep it from them. It extends to teachers too, I mentioned this in a note I just shared but when young teachers are baselessly relying on Claude, they are missing the development of pedagogical tact, yet teachers who are already expert pedagogues are able to use these tools to do really amazing things. I want to both protect young learners / novice teachers from that loss and also let those with expertise run with it. In our research we've seen many schools adopt the "red, yellow, green" rule for activities and assignments - red means no AI, yellow means some restrictions on AI, green means go for it. There's something there with that idea, but we need to clear up how we know when a student has developed enough 'expertise' to warrant a yellow or green activity.
I will say, though, my bigger concern these days is the over-corporatization of schools that is already happening. AI tools for the most part are products of just a few powerful corporations, what does it mean for education when OpenAI, Anthropic, Google, and SpaceX are in complete control of the algorithm? Not to mention the building tension between the government and these corporations. I worry, broadly, about how much epistemological power these entities, let alone how much power they'd have in the classroom.
This is such a good point, I was thinking the same. I'm wary of the constant messaging that "AI is inevitable", "Adapt to AI or fall behind", etc, both surrounding who is saying these things and why. There are a myriad of other large concerns regarding AI that I wonder if introducing it in education at this point in time is a good idea at all.
In addition to the corporations that have control over this market and technology, AI has been seen to amplify preexisiting issues, especially as we lack a lot of data regarding the problems we are attempting to fix (to me this problem is most apparent in healthcare, especially for women's and POC's health). Not to mention, AI data centers excacerbating problems such as climate change, wealth inequality/poverty, etc, because they require much more energy and resources to function when compared to previous data centers.
Of course, there are preexisting studies which show the impact of these problems on humans and their ability to learn (as a downstream effect), and thus will outweigh any supposed benefit any AI program can do to enhance a student's learning plan or focus. There aren't enough regulatory frameworks in place, or enough priority being put towards understanding what is necessary on a government scale, that is proportionate to the scale at which AI has been pushed into many spheres of life, including schools. A lot of governments seem to be adapting to corporations' will instead of protecting their people.
This is obviously a very elementary summary of ethical concerns surrounding this technology, but a good starting point. It may be more beneficial to engineer AI more, and have stricter policies surrounding it for the time being, so that negative impacts can be minimized as much as possible while it is being better understood. Kids do not have the agency to regulate their own AI usage, and cannot be blamed when they have been bombarded with this sort of messaging (which was pointed out in the article).
Additionally, 'AI' is such a broad term which encompasses so much, so I believe it would do good to have more narrowed down definitions of it to be used in varying contexts and capacities to clarify its use cases, abilities, and intention. Are they using generative AI, analytical AI, or something else? What other ways can AI be defined more specifically? Lessening the scope will also make it easier to understand, because it will break it into parts.
"It is also the part that can never be looked up, because it is the very thing that makes looking up intelligible in the first place." This statement haunts me because it feels so familiar--overly familiar--the perfect description of why so many AI queries make so little sense.
I really enjoyed reading your post - thank you. It reminded me of Beverly Park Woolf's book, Building Intelligent Interactive Tutors: Student-Centered Strategies for Revolutionizing E-learning. Written back in 2009, it predates deep learning and large language models, yet she heavily influenced modern AI tutoring and had a significant impact on my own ideas and studies whilst at LKL/UCL. It is encouraging to read about modern AI systems implementing what were predominantly theoretical concepts in 2009. Woolf brilliantly integrates educational theory with computational architecture and AI reasoning.
The magenta line reminds me of our phone navigation. It's a great tool for figuring out the best way to get somewhere; but some people have also followed it into a lake. We need to make sure people remember how to navigate - even if GPS navigation is a great tool.
Some research has shown that people are rapidly at sea when the navigation device is removed. Another form of losing capacity via outsourcing to the ‘machine’.
Please, will someone - anyone - who is familiar with Alpha’s learning delivery platform enlighten us as to:
1. How exactly it functions, as a permanently ‘teacherless’ (non-human) system for the academic side of instruction.
2. What is actually taught - eg subjects - at each level (apart from SAT prep which we are told it “crushes”).
3. Exactly how the relentlessly incentivised learning system (via overt material rewards, such as cashable tokens etc) actually works.
Its interesting to speculate re the efficacy of the model were this system withdrawn.
I ask out of genuine interest as a long time educator who is not looking to uncover trade secrets, rather to understand how the system holds up as a true rounded “education”.
One drum we should keep beating is this novice/expert tension—both for returning culture back to common sense about knowledge building and for building paradigms for when/how to use AI in schools.
This lands hard: solving a problem and learning from it are not the same process.
The magenta-line frame also maps cleanly onto brain-training apps. Autopilot drills that feel productive can leave people dependent on the drill format, with little transferable judgment when the interface changes. Constrained practice that keeps the cognitive work on the learner (retrieval, desirable difficulty, feedback after attempt) is closer to what you describe at Alpha than open chatbot scaffolding for novices.
I build IntelligenceMax (adaptive reasoning practice). Disclosure: founder. We try to keep the effortful part with the user and stay honest that practice metrics ≠ clinical IQ. Short overview if useful: https://intelligencemax.ai/guide
What will happen when it's easier to train a scalable AI than teach a group of unscalable humans? Particularly when quadrillions of dollars are invested to the former instead of the latter?
The scarce resource is not time, it's attention" — yes, and I suspect most readers will scroll past that line to fight about Alpha.
One push. You call the tracking — where each child is, what they just got wrong, what they need next — the clerical substrate beneath the human part. I'd offer that in a real room it isn't clerical. It's the attention itself. That noticing is what the good teacher does with her eyes, and it doesn't quite detach from the relationship without becoming a different act.
Which makes me watchful of the verb. Once "knowing this child" gets filed as substrate, it can be handed to a system, and that handoff always arrives dressed as liberation, freeing the teacher for the real work. That's the sentence every measurement regime has used for a century, and the freed-up time rarely shows up.
Maybe Alpha's the exception. But the tell will be in the guide's attention, not the student's — whether she's watching the child, or watching the dashboard tell her about the child.
I've never understood why Alpha School's Instagram has such a shallow, anti-academic attitude. It keeps repeating myths and outright nonsense.
Because it’s effective at getting reach & resonates with their target parents.
Yes, probably - because it is populism.
This needs a reply from the author.
I'm curious to see how Alpha School develops over the next few years, both in terms of the way we can use machines/technology to create optimal support for learning and mastery, and how any potential success might generate possibilities for what schools do. If (and it is an if at this point, not a given) machines can do what you have outlined, then the question becomes what do schools offer beyond this? One outcome might be one where schools can lean into their mission or identity with greater clarity. Texas Sports Academy is an example. And I think Zach Groshell, in his conversation with Craig Barton (https://www.youtube.com/watch?v=ZZz4HCtHWxI), talked about his choice as a parent as to the type of school he wanted for his child. I think this type of future might be possible, but it isn't hard to see challenges that might come when you attempt to implement this at the level of a public school system.
My other thought is similarly badly formulated at this point, but I'll take a risk and share it anyhow. I wonder whether, once we control for all the cognitive elements of learning (CLT, attention, scaffolding, spacing, interleaving, etc.), Alpha and the researchers will be as interested in looking at the affective side of learning? The interaction between confidence and competence feels like something Alpha could look at over a period of time, and we might learn something interesting and of general value. I wonder about the variation between affective learning when at a screen engaged in mastery learning versus affective learning during the other six hours of the day, when not engaged in carefully scaffolded learning?
This is a brilliant essay, as always, Carl - I appreciate your time writing it. I don't always agree with everything but you've connected with a big tension I have been feeling lately with AI in the classroom. It feels like something to protect our students' learning from AND a missed opportunity if we keep it from them. It extends to teachers too, I mentioned this in a note I just shared but when young teachers are baselessly relying on Claude, they are missing the development of pedagogical tact, yet teachers who are already expert pedagogues are able to use these tools to do really amazing things. I want to both protect young learners / novice teachers from that loss and also let those with expertise run with it. In our research we've seen many schools adopt the "red, yellow, green" rule for activities and assignments - red means no AI, yellow means some restrictions on AI, green means go for it. There's something there with that idea, but we need to clear up how we know when a student has developed enough 'expertise' to warrant a yellow or green activity.
I will say, though, my bigger concern these days is the over-corporatization of schools that is already happening. AI tools for the most part are products of just a few powerful corporations, what does it mean for education when OpenAI, Anthropic, Google, and SpaceX are in complete control of the algorithm? Not to mention the building tension between the government and these corporations. I worry, broadly, about how much epistemological power these entities, let alone how much power they'd have in the classroom.
This is such a good point, I was thinking the same. I'm wary of the constant messaging that "AI is inevitable", "Adapt to AI or fall behind", etc, both surrounding who is saying these things and why. There are a myriad of other large concerns regarding AI that I wonder if introducing it in education at this point in time is a good idea at all.
In addition to the corporations that have control over this market and technology, AI has been seen to amplify preexisiting issues, especially as we lack a lot of data regarding the problems we are attempting to fix (to me this problem is most apparent in healthcare, especially for women's and POC's health). Not to mention, AI data centers excacerbating problems such as climate change, wealth inequality/poverty, etc, because they require much more energy and resources to function when compared to previous data centers.
Of course, there are preexisting studies which show the impact of these problems on humans and their ability to learn (as a downstream effect), and thus will outweigh any supposed benefit any AI program can do to enhance a student's learning plan or focus. There aren't enough regulatory frameworks in place, or enough priority being put towards understanding what is necessary on a government scale, that is proportionate to the scale at which AI has been pushed into many spheres of life, including schools. A lot of governments seem to be adapting to corporations' will instead of protecting their people.
This is obviously a very elementary summary of ethical concerns surrounding this technology, but a good starting point. It may be more beneficial to engineer AI more, and have stricter policies surrounding it for the time being, so that negative impacts can be minimized as much as possible while it is being better understood. Kids do not have the agency to regulate their own AI usage, and cannot be blamed when they have been bombarded with this sort of messaging (which was pointed out in the article).
Additionally, 'AI' is such a broad term which encompasses so much, so I believe it would do good to have more narrowed down definitions of it to be used in varying contexts and capacities to clarify its use cases, abilities, and intention. Are they using generative AI, analytical AI, or something else? What other ways can AI be defined more specifically? Lessening the scope will also make it easier to understand, because it will break it into parts.
"It is also the part that can never be looked up, because it is the very thing that makes looking up intelligible in the first place." This statement haunts me because it feels so familiar--overly familiar--the perfect description of why so many AI queries make so little sense.
I really enjoyed reading your post - thank you. It reminded me of Beverly Park Woolf's book, Building Intelligent Interactive Tutors: Student-Centered Strategies for Revolutionizing E-learning. Written back in 2009, it predates deep learning and large language models, yet she heavily influenced modern AI tutoring and had a significant impact on my own ideas and studies whilst at LKL/UCL. It is encouraging to read about modern AI systems implementing what were predominantly theoretical concepts in 2009. Woolf brilliantly integrates educational theory with computational architecture and AI reasoning.
The magenta line reminds me of our phone navigation. It's a great tool for figuring out the best way to get somewhere; but some people have also followed it into a lake. We need to make sure people remember how to navigate - even if GPS navigation is a great tool.
Some research has shown that people are rapidly at sea when the navigation device is removed. Another form of losing capacity via outsourcing to the ‘machine’.
Please, will someone - anyone - who is familiar with Alpha’s learning delivery platform enlighten us as to:
1. How exactly it functions, as a permanently ‘teacherless’ (non-human) system for the academic side of instruction.
2. What is actually taught - eg subjects - at each level (apart from SAT prep which we are told it “crushes”).
3. Exactly how the relentlessly incentivised learning system (via overt material rewards, such as cashable tokens etc) actually works.
Its interesting to speculate re the efficacy of the model were this system withdrawn.
I ask out of genuine interest as a long time educator who is not looking to uncover trade secrets, rather to understand how the system holds up as a true rounded “education”.
I think the best nuts & bolts explanation of Alpha School is here: https://www.astralcodexten.com/p/your-review-alpha-school
Thanks, it looks detailed & I’ll take a look. Its a mystery as to why it’s so hard to get a response from people who write about the school. Cheers.
An instant classic! Thank you!
One drum we should keep beating is this novice/expert tension—both for returning culture back to common sense about knowledge building and for building paradigms for when/how to use AI in schools.
Carl- it would be helpful for the readers if you gave specific examples of how AI tools could be used effectively by teachers and/or students. Thanks
Excellent writing except for your unnecessary calling out of critical pedagogy as slop.
This lands hard: solving a problem and learning from it are not the same process.
The magenta-line frame also maps cleanly onto brain-training apps. Autopilot drills that feel productive can leave people dependent on the drill format, with little transferable judgment when the interface changes. Constrained practice that keeps the cognitive work on the learner (retrieval, desirable difficulty, feedback after attempt) is closer to what you describe at Alpha than open chatbot scaffolding for novices.
I build IntelligenceMax (adaptive reasoning practice). Disclosure: founder. We try to keep the effortful part with the user and stay honest that practice metrics ≠ clinical IQ. Short overview if useful: https://intelligencemax.ai/guide
What will happen when it's easier to train a scalable AI than teach a group of unscalable humans? Particularly when quadrillions of dollars are invested to the former instead of the latter?
The scarce resource is not time, it's attention" — yes, and I suspect most readers will scroll past that line to fight about Alpha.
One push. You call the tracking — where each child is, what they just got wrong, what they need next — the clerical substrate beneath the human part. I'd offer that in a real room it isn't clerical. It's the attention itself. That noticing is what the good teacher does with her eyes, and it doesn't quite detach from the relationship without becoming a different act.
Which makes me watchful of the verb. Once "knowing this child" gets filed as substrate, it can be handed to a system, and that handoff always arrives dressed as liberation, freeing the teacher for the real work. That's the sentence every measurement regime has used for a century, and the freed-up time rarely shows up.
Maybe Alpha's the exception. But the tell will be in the guide's attention, not the student's — whether she's watching the child, or watching the dashboard tell her about the child.