Alpha insider here. Going to confine my comments to the “2x learning is bogus” part of this critique.
As noted in footnote 1, my view is that there’s a more statistically precise way of saying this, a way that (a) virtually no one understands and (b) *is in fact very impressive*. A 1.4 to 1.5 sigma effect size, which is what I typically get when I run the numbers, is very large. So the worst-case scenario for 2x learning is that it’s just an incorrect description of what it is in fact an impressive effect.
But how imprecise is “2x learning” exactly? Putting aside the standard deviation for a moment (I’ll return to it), and using my 6yo as an example, the following things are true: she scored a 179 RIT on the fall 2025 math MAP. She scored a 222 RIT on the spring 2026 math MAP. The median student scoring a 179 on the fall test gains 15 points over the year. My daughter gained 43 points.
Moving beyond what’s indisputably true: I don’t think my daughter would have had 43 points of gains in a typical school, or even another good Montessori school. I think it would have been much closer to 15. (If you believe something like this, almost all the other discussion is irrelevant, and it answers the question as to why Alpha is beloved to many parents.)
Alpha markets that to me (in my capacity as a parent) as 2.9x learning. That seems... fine to me? It’s not what I would publish for data wonks, but it gives an impression that is correct: she outperformed, by a lot.
Does that mean that she learned 3 years worth of material? Not necessarily. I don’t think Alpha has ever said that. It’s presented to parents as a testing target. *Grade levels gains are reported separately and clearly parents.* (Incidentally, in fact, she did, almost. She went from mid-K material to 3rd grade material. But this doesn’t always line up with the RIT multiples.)
Returning to the shape of the data. So the median yearlong score improvement is 15 points for a 179 math MAP. That the median is 15 doesn’t tell you anything about how likely it is that a student scores a 43; if the curve is relatively flat, it could be about as common as any other result. We need to know the standard deviation to estimate the effect size (e.g. (gain - median_gain) / sd). I can basically look that up (inferring from some other data that MAP has published). It’s 8.5. (43 - 15) / 8.5 == 3.3. 3.3 is an insane effect.
There are all sorts of reasons one might want to tweak or dispute that, or use slightly different measures, but you’re going to end up in the same ballpark. It’s going to be a big effect, because it is one.
Does this happen every single time you compute the results of 2x (or 3x or...) learning? Not exactly in this way. The SD isn’t always the same across ages and subjects. But this *is* the basic shape of the effect. As far as I can tell, having run this exercise across many contexts: it never turns out that the range is such that a 2x effect is basically nothing to write home about.
What about outlier cases, where you get, like, 20x learning? Again, this happened to my 6yo. Her reading last year went from 201 to 222 fall to winter. MAP exepcted her to gain 1 point; she gained 21. Is that “21x learning”?
Well, I mean... what do we make of the fact that her score might have gone up by 1 in a typical school? She’s a gifted reader; gifted readers are typically not challenged in most Kindergartens, and the Alpha platform absolutely challenged her every single day. It is entirely plausible to me that if I enrolled her in either a decent public school or a decent Montessori elementary program (two of my benchmarks for Montessorium, the Alpha Montessori program she is enrolled in), she would not have been challenged.
To be clear, I wouldn’t put 21x learning on a billboard. But I don’t think it’s fair to say that the numbers are meaningless. (If you run the same math as I just did with math with derived SDs from public MAP data, you get a 2.3 sigma effect. Not incidentally, you are basically breaking the models at this point, which are not designed to be informative about this kind of gain, from 99 point something to a higher 99 point something.)
As far as probably overhyped and/or imprecise marketing claims go, Alpha’s seems minimally objectionable. And the minimum objection is not that it’s *misleading*, not that there’s some sort of false promise embedded in it, but that the effects are quite real but confusingly described for people who want to understand it precisely. And they are confusing because the actual statistical truth is objectively hard to communicate; our marketing department has wisely opted to not communicate in sigmas.
It feels like what a lot of people like about AlphaSchool is less that it’s proactively enacting some amazing set of outcomes, and more that it *lacks* a great deal of the self-imposed downsides of many other schools.
Like a class room with five behavior problems plus six kids with IEP's and one very frazzled teacher. Those sound like real downsides to me. I'm glad my kids are grown.
Naimoli alluded to this in his article, but just to make it more explicit about metrics: if you have a bunch of multipliers, it really doesn't make a lot of sense to average them in a traditional way.
For example, if you have one kid whose score was 2 times the median and one kid whose score was 1/2 times the median, intuitively it should make sense that these multipliers average out to be 1 times the median, but a normal arithmetic mean would say that on average the kids were 1.25 times the median. To actually capture this you can use the geometric mean, where you multiply all n of your numbers and take the n'th root at the end; here, you get sqrt(2 * 0.5) = 1, so it tells you that you 1x'd on average.
What's interesting though is that the arithmetic mean is ALWAYS higher than the geometric mean (https://en.wikipedia.org/wiki/AM%E2%80%93GM_inequality), so if Alpha is reporting this average multiplier by doing an arithmetic mean, it will necessarily be overstating results!
This sounds right to me: the secret sauce is bribery, the test results are good, the software isn’t exportable outside their model, and they either don’t want to share or don’t know how to do careful research.
I’m curious about why their software (and ed tech generally) is still so shoddy after a supposed 100m+ investment. It’s 2026, Claude or Codex can build something polished if you have a clear vision and minor dev skills.
It's nice that a microschooler wrote this article because it means the focus is on what Alpha School actually represents. We've known for a long time that a large part of the function of school is custodial, but it was always thought that there was no way to separate the custodial function from the academic function.
Alpha School's proposition is that the academic function is reduced to 25% of the time commitment or less. Even if that's not really true, it's reasonable to predict that future edtech or other advancements could make it true. What Alpha School should be making us think about is, what is the other 75% for?
Essentially, what should kids do all day if they didn't *have* to spend the whole day in class? Practice a skill? Get a job? Speedrun the curriculum? See the world? Form pint-size gangs and roam the neighborhood?
History, and the existence of homeschooling, have shown that there are markets for all of the above. Maybe this is an area where the market should decide. Or maybe societies will want to make that decision for all their kids. The fascinating aspect of Alpha School, whether it actually works or not, is that we may find out.
I'll admit that alpha appeals to a lot of replacement level intelligence tech bros (me). I'm very much not a genius but I remember just how insanely boring my public school was. Just an insane amount of time was spent on classroom management. Math was the only class I enjoyed in elementary. Some how they managed to even ruin science with endless rote memorization of the words that make up the "scientific method" and an absolute perversion with sigfigs.
I understand why many parents want to avoid that same experience for their kids. If I was making bay area tech bro money, I would definitely consider it.
Exactly, regular school is so slow that getting through the same curriculum in half the time just doesn't sound hard or impressive to me. All you have to do is not drag your feet.
I think a big selling point is that no matter how bad Alpha School is, it only takes a few hours. If they only assign terrible AI generated books, you can give your child some good books yourself and they'll have plenty of time to appreciate them.
In the third world we achieve a similar effect much more cheaply in regular school, which for all its faults at least doesn't take all day like the first world version. The school day used to be so short that there was a morning shift and an afternoon shift that both could use the same classroom, because their schedules never overlapped. Nowadays school takes longer and I can tell that kids are much more strongly affected by whether or not their school is good. Not just in test scores or whatever but also their mood and happiness.
There’s value in reading and discussing a book with a class, led by a competent teacher. Handing a kid a good book is great, but it isn’t the same activity. Personally I find the AI generated snippets depressing. Especially when a lot of the tech world seems to be led by anti-human, broken misanthropes.
I don't remember what the specific value of discussing a book with the class was. I remember it being annoying, and it took forever because we had to wait for the slow kids.
Some of the discussions led by the teacher were good, but not worth having to wait so long.
Discussing with strangers on the internet is more efficient, but then you really feel the lack of a competent teacher. Discussing (human-written) books with an AI works OK to replace all but the most competent of teachers. but then you miss the social aspect.
Sounds like book clubs are not for you. Because a book club is basically discussing the book with a class of your choosing. Those of us who like class discussions join book clubs.
A book club sounds amazing compared to reading books at school.
No waiting for the slow kids, no frustrated teacher trying to coax some kind of comment or answer out of a kid who would rather be doing anything else.
It's similar to language classes. They never work in a classroom because the slow kids and the kids who would rather be doing anything else drain all the fun out of it. But a voluntary club for practicing a language can be very fun.
I think that really depends on the book and how we are discussing it. I love reading, I have over 1000 books on my kindle, and many more in paperback and hardback
But I usually HATE class discussions about them. They tended to be so much BS.
The hypothesis that school is so bad parents are willing to pay $75k per year to shrink it is interesting. Couldn't you pay a part time tutor half that for the same result?
That's how my parents did it. It was affordable even for a poor family because labor costs are very low in the third world, there were plenty of underemployed teachers to go around.
In the first world I think they pay the 75k because private tutors are already not cheap, you can't risk a random cheap tutor messing up your kid's test scores (famously the one thing Alpha School does care about) and also because of selection effects. If you want your kid to hang out with the kids of other excentric rich families, Alpha School is invaluable.
If only they didn’t use AI slop to teach reading, I could get past all of these other issues. 2hrs in a mediocre app that allows you go at your own speed + the rest of the day for extracurriculars of your choice is a significant step above the standard public school around here. (Although it wouldn’t be a step above the standard public school if public schools simply offered tracking…)
I can’t get past the AI slop thing. Am I just a snob? Is it better for a gifted 6 year old reader to rapidly level up her reading ability on AI slop, or sit around and twiddle her thumbs while her peers learn to read at all? I don’t know!! But I don’t want my kid to learn bad taste!!
We home school, the regular book learning is usually done in 2 to 3 hours. then yep we do things like music, karate, just go play outside etc
We also very much encourage separate reading outside it. I think my daughter was reading through the first couple of Harry Potter books by 7 or 8
But one of the biggest reasons is because you get individualized instruction. Especially when learning to read. It's really more of a one on one activity. Same occurs if there is a problem understanding math.
If I need to spend 10 or 15 minutes (or even an hour) with one kid going over a math concept it's fine. We can explain it it many different ways, and keep practicing problems till we get it.
I don't have any context for Alpha besides this article. It's really the first thing I've read about it. But from everything you've said I come across with a very positive impression. Because if they're even getting the same results as normal schools for equivalent students but spending much less time on boring sitting around in class, that's a huge win. Even if the win is just realizing that 80% of stuff we did in school wasn't needed
Picking out metrics you can apply to any reality that sound like you’re making real and sustainable progress while you wait for the institution to be shut down is good preparation for corporate America
Alpha insider here. Going to confine my comments to the “2x learning is bogus” part of this critique.
As noted in footnote 1, my view is that there’s a more statistically precise way of saying this, a way that (a) virtually no one understands and (b) *is in fact very impressive*. A 1.4 to 1.5 sigma effect size, which is what I typically get when I run the numbers, is very large. So the worst-case scenario for 2x learning is that it’s just an incorrect description of what it is in fact an impressive effect.
But how imprecise is “2x learning” exactly? Putting aside the standard deviation for a moment (I’ll return to it), and using my 6yo as an example, the following things are true: she scored a 179 RIT on the fall 2025 math MAP. She scored a 222 RIT on the spring 2026 math MAP. The median student scoring a 179 on the fall test gains 15 points over the year. My daughter gained 43 points.
Moving beyond what’s indisputably true: I don’t think my daughter would have had 43 points of gains in a typical school, or even another good Montessori school. I think it would have been much closer to 15. (If you believe something like this, almost all the other discussion is irrelevant, and it answers the question as to why Alpha is beloved to many parents.)
Alpha markets that to me (in my capacity as a parent) as 2.9x learning. That seems... fine to me? It’s not what I would publish for data wonks, but it gives an impression that is correct: she outperformed, by a lot.
Does that mean that she learned 3 years worth of material? Not necessarily. I don’t think Alpha has ever said that. It’s presented to parents as a testing target. *Grade levels gains are reported separately and clearly parents.* (Incidentally, in fact, she did, almost. She went from mid-K material to 3rd grade material. But this doesn’t always line up with the RIT multiples.)
Returning to the shape of the data. So the median yearlong score improvement is 15 points for a 179 math MAP. That the median is 15 doesn’t tell you anything about how likely it is that a student scores a 43; if the curve is relatively flat, it could be about as common as any other result. We need to know the standard deviation to estimate the effect size (e.g. (gain - median_gain) / sd). I can basically look that up (inferring from some other data that MAP has published). It’s 8.5. (43 - 15) / 8.5 == 3.3. 3.3 is an insane effect.
There are all sorts of reasons one might want to tweak or dispute that, or use slightly different measures, but you’re going to end up in the same ballpark. It’s going to be a big effect, because it is one.
Does this happen every single time you compute the results of 2x (or 3x or...) learning? Not exactly in this way. The SD isn’t always the same across ages and subjects. But this *is* the basic shape of the effect. As far as I can tell, having run this exercise across many contexts: it never turns out that the range is such that a 2x effect is basically nothing to write home about.
What about outlier cases, where you get, like, 20x learning? Again, this happened to my 6yo. Her reading last year went from 201 to 222 fall to winter. MAP exepcted her to gain 1 point; she gained 21. Is that “21x learning”?
Well, I mean... what do we make of the fact that her score might have gone up by 1 in a typical school? She’s a gifted reader; gifted readers are typically not challenged in most Kindergartens, and the Alpha platform absolutely challenged her every single day. It is entirely plausible to me that if I enrolled her in either a decent public school or a decent Montessori elementary program (two of my benchmarks for Montessorium, the Alpha Montessori program she is enrolled in), she would not have been challenged.
To be clear, I wouldn’t put 21x learning on a billboard. But I don’t think it’s fair to say that the numbers are meaningless. (If you run the same math as I just did with math with derived SDs from public MAP data, you get a 2.3 sigma effect. Not incidentally, you are basically breaking the models at this point, which are not designed to be informative about this kind of gain, from 99 point something to a higher 99 point something.)
As far as probably overhyped and/or imprecise marketing claims go, Alpha’s seems minimally objectionable. And the minimum objection is not that it’s *misleading*, not that there’s some sort of false promise embedded in it, but that the effects are quite real but confusingly described for people who want to understand it precisely. And they are confusing because the actual statistical truth is objectively hard to communicate; our marketing department has wisely opted to not communicate in sigmas.
It feels like what a lot of people like about AlphaSchool is less that it’s proactively enacting some amazing set of outcomes, and more that it *lacks* a great deal of the self-imposed downsides of many other schools.
Tech bros inventing rich private schools from first principles
“Guys guys I figured it out”
Like a class room with five behavior problems plus six kids with IEP's and one very frazzled teacher. Those sound like real downsides to me. I'm glad my kids are grown.
Yes
Naimoli alluded to this in his article, but just to make it more explicit about metrics: if you have a bunch of multipliers, it really doesn't make a lot of sense to average them in a traditional way.
For example, if you have one kid whose score was 2 times the median and one kid whose score was 1/2 times the median, intuitively it should make sense that these multipliers average out to be 1 times the median, but a normal arithmetic mean would say that on average the kids were 1.25 times the median. To actually capture this you can use the geometric mean, where you multiply all n of your numbers and take the n'th root at the end; here, you get sqrt(2 * 0.5) = 1, so it tells you that you 1x'd on average.
What's interesting though is that the arithmetic mean is ALWAYS higher than the geometric mean (https://en.wikipedia.org/wiki/AM%E2%80%93GM_inequality), so if Alpha is reporting this average multiplier by doing an arithmetic mean, it will necessarily be overstating results!
This sounds right to me: the secret sauce is bribery, the test results are good, the software isn’t exportable outside their model, and they either don’t want to share or don’t know how to do careful research.
I’m curious about why their software (and ed tech generally) is still so shoddy after a supposed 100m+ investment. It’s 2026, Claude or Codex can build something polished if you have a clear vision and minor dev skills.
Iirc they used a lot of off-the-shelf EdTech software for the first couple of years and have now moved to their own bespoke solution
It's nice that a microschooler wrote this article because it means the focus is on what Alpha School actually represents. We've known for a long time that a large part of the function of school is custodial, but it was always thought that there was no way to separate the custodial function from the academic function.
Alpha School's proposition is that the academic function is reduced to 25% of the time commitment or less. Even if that's not really true, it's reasonable to predict that future edtech or other advancements could make it true. What Alpha School should be making us think about is, what is the other 75% for?
Essentially, what should kids do all day if they didn't *have* to spend the whole day in class? Practice a skill? Get a job? Speedrun the curriculum? See the world? Form pint-size gangs and roam the neighborhood?
History, and the existence of homeschooling, have shown that there are markets for all of the above. Maybe this is an area where the market should decide. Or maybe societies will want to make that decision for all their kids. The fascinating aspect of Alpha School, whether it actually works or not, is that we may find out.
I'll admit that alpha appeals to a lot of replacement level intelligence tech bros (me). I'm very much not a genius but I remember just how insanely boring my public school was. Just an insane amount of time was spent on classroom management. Math was the only class I enjoyed in elementary. Some how they managed to even ruin science with endless rote memorization of the words that make up the "scientific method" and an absolute perversion with sigfigs.
I understand why many parents want to avoid that same experience for their kids. If I was making bay area tech bro money, I would definitely consider it.
Exactly, regular school is so slow that getting through the same curriculum in half the time just doesn't sound hard or impressive to me. All you have to do is not drag your feet.
The apparently hard and impressive thing about it is that it's almost impossible to not have public schools drag their feet on almost everything!
I think a big selling point is that no matter how bad Alpha School is, it only takes a few hours. If they only assign terrible AI generated books, you can give your child some good books yourself and they'll have plenty of time to appreciate them.
In the third world we achieve a similar effect much more cheaply in regular school, which for all its faults at least doesn't take all day like the first world version. The school day used to be so short that there was a morning shift and an afternoon shift that both could use the same classroom, because their schedules never overlapped. Nowadays school takes longer and I can tell that kids are much more strongly affected by whether or not their school is good. Not just in test scores or whatever but also their mood and happiness.
There’s value in reading and discussing a book with a class, led by a competent teacher. Handing a kid a good book is great, but it isn’t the same activity. Personally I find the AI generated snippets depressing. Especially when a lot of the tech world seems to be led by anti-human, broken misanthropes.
I don't remember what the specific value of discussing a book with the class was. I remember it being annoying, and it took forever because we had to wait for the slow kids.
Some of the discussions led by the teacher were good, but not worth having to wait so long.
Discussing with strangers on the internet is more efficient, but then you really feel the lack of a competent teacher. Discussing (human-written) books with an AI works OK to replace all but the most competent of teachers. but then you miss the social aspect.
Cool, there is still value even if you don’t remember it.
Sounds like book clubs are not for you. Because a book club is basically discussing the book with a class of your choosing. Those of us who like class discussions join book clubs.
A book club sounds amazing compared to reading books at school.
No waiting for the slow kids, no frustrated teacher trying to coax some kind of comment or answer out of a kid who would rather be doing anything else.
It's similar to language classes. They never work in a classroom because the slow kids and the kids who would rather be doing anything else drain all the fun out of it. But a voluntary club for practicing a language can be very fun.
I think that really depends on the book and how we are discussing it. I love reading, I have over 1000 books on my kindle, and many more in paperback and hardback
But I usually HATE class discussions about them. They tended to be so much BS.
The hypothesis that school is so bad parents are willing to pay $75k per year to shrink it is interesting. Couldn't you pay a part time tutor half that for the same result?
That's how my parents did it. It was affordable even for a poor family because labor costs are very low in the third world, there were plenty of underemployed teachers to go around.
In the first world I think they pay the 75k because private tutors are already not cheap, you can't risk a random cheap tutor messing up your kid's test scores (famously the one thing Alpha School does care about) and also because of selection effects. If you want your kid to hang out with the kids of other excentric rich families, Alpha School is invaluable.
If only they didn’t use AI slop to teach reading, I could get past all of these other issues. 2hrs in a mediocre app that allows you go at your own speed + the rest of the day for extracurriculars of your choice is a significant step above the standard public school around here. (Although it wouldn’t be a step above the standard public school if public schools simply offered tracking…)
I can’t get past the AI slop thing. Am I just a snob? Is it better for a gifted 6 year old reader to rapidly level up her reading ability on AI slop, or sit around and twiddle her thumbs while her peers learn to read at all? I don’t know!! But I don’t want my kid to learn bad taste!!
We home school, the regular book learning is usually done in 2 to 3 hours. then yep we do things like music, karate, just go play outside etc
We also very much encourage separate reading outside it. I think my daughter was reading through the first couple of Harry Potter books by 7 or 8
But one of the biggest reasons is because you get individualized instruction. Especially when learning to read. It's really more of a one on one activity. Same occurs if there is a problem understanding math.
If I need to spend 10 or 15 minutes (or even an hour) with one kid going over a math concept it's fine. We can explain it it many different ways, and keep practicing problems till we get it.
I don't have any context for Alpha besides this article. It's really the first thing I've read about it. But from everything you've said I come across with a very positive impression. Because if they're even getting the same results as normal schools for equivalent students but spending much less time on boring sitting around in class, that's a huge win. Even if the win is just realizing that 80% of stuff we did in school wasn't needed
Picking out metrics you can apply to any reality that sound like you’re making real and sustainable progress while you wait for the institution to be shut down is good preparation for corporate America
Goodhart's law and cohort effects always reign supreme in education reform gains.