That is not what they found. What they found is that if you spend 60 minutes doing the homework with AI, you learn approximately as much as someone who spends 60 minutes doing it without AI. If you spend 20 minutes doing it with AI, you learn as much as someone who spends 20 minutes doing it without AI. Etc. The only real difference is that the AI people get way better homework scores, and as a result, tend to stop spending as much time, and thus tend to learn less.

That's what this preprint study found.

100% this. Even a short 5-minute quiz with just a couple questions at the start of each day will do the trick. But you simply can't rely on homework scores to be a proxy of learning or time spent studying anymore. Brief daily quizzes on the previous day's material will force them to.

Also, It's not like students have the excuse of, "I couldn't study, I didn't have anyone to explain X to me" anymore. The very AI that they might otherwise use to cheat on their homework can also explain to them the things they missed in last year's algebra class that are blocking them from learning this year's calculus, or whatnot. Congrats - you now have an infinitely patient personal tutor who will spend hours dumbing things down for you until you get it. Now you have no excuse - learn the material. It becomes all about actually putting the time in.

Yeah. There's been widespread misrepresentation of this (preprint!) in the media. What they actually found: test scores correspond to how much time you spend doing homework, regardless of whether you use AI or not. Which should not be a surprising result! Of course, the AI people, esp. those who had been using AI for more than a few months, stopped spending much time on their homework, and then their test scores fell to match the non-AI users who didn't spend much time in their homework either.

The TL/DR is, in today's age, you cannot use test scores as a proxy of how much time a user is studying or knows the material.

There are of course some caveats, beyond the fact that it's just a preprint - notably that there's nothing to separate why the user started using AI to do their homework, so there's potential confounders in there. For example, if a student is in a poor family, maybe they had to get a job to help pay the bills. But working a job itself can lower test scores - for example, having to work late the night before an exam. So there are some confounder problems with the study. That said, I'm sure the general conclusion is correct: time spent doing the thing = amount of learning the thing (relatively independent of what tools you're using in the process).

I feel like I'm in a constant race to flee enshittification : Used to use Cursor (it's still my only option at work), but then Musk bought it, and they instantly went in on hard-sell tactics. Switched to vscode + copilot + OpenRouter, but was getting concerned about API costs. Just yesterday I set up a Tencent TokenHub account with their Max package, which required that I switch off copilot if I wanted to be able to use some models like Kimi K3 (had a temperature=1.0 requirement, copilot doesn't let you set temperatures and hardcodes them to very low), so now it's vscode + Pi + TokenHub.

Sigh. Please don't make me move again...

(Not that Tencent is in any way "the good guys"... but they're also providing a very generic product)

Ah, the good 'ol "No True Communism" argument :)

"It suggests" if you ignore that their confounder control is terrible. First, sugar wasn't the only restricted food in post-war Britain - there was also rationing of eggs, butter, cheese and meats that ended during the same timeperiod. So you're not actually testing the effects of sugar consumption, now are you? Also, the early 1950s was a period of rapid change in the US, including the maturation and expansion of the NHS (founded in 1948), the introduction of new childhood vaccines, major environmental shifts (for example, the Great Smog of London in 1952 vs. the 1956 Clean Air Act). Saying "all changes = sugar" just based on birthdate seems absurd.

It's also worth mentioning that there's also survivorship bias - anyone who died before the study couldn't be enrolled, and there's also no actual linking to any maternal / childhood dietary logs. Also, the ratios of cancer differences they're describing aren't small, they're massive. Are we to attribute a lung cancer difference on the order of smoking vs. nonsmoking households as being to sugar, or due to, you know, smoking (or pollution) trends not accounted for in their cohort?

Way to out yourself as being American by being simultaneously loudly, assertively ignorant and racist at the same time.

** - Collective guilt is a fallacy; modern Germany is a highly democratic constitutional state with a commendable record on teaching against extremism; Marx and Engel drew heavily from French utopian socialists (e.g., Saint-Simon, Fourier), while it was the Russians who actually implemented communism; World War I began with the assassination of Austrian Archduke Franz Ferdinand by a Bosnian Serb nationalist, triggering a chain reaction of ultimatums and alliances involving Austria-Hungary, Serbia, Russia, France, Germany, and Great Britain, with any individual state bearing guilt for the war being a notion firmly rejected by modern historians; Hitler was born in Austria-Hungary and was an Austrian citizen until he renounced it in 1925, not gaining German citizenship until 1932; the East German government and the Stasi was forced on the people of Germany by the USSR, and it was not a global phenomenon "brought to the world", but rather dissolved with the GDR in 1990. TL/DR: Go fuck yourself.

the Facewatch system has a 99.98% accuracy rate,

So one in every five thousand people who walk into your stores is going to be falsely accused of a crime? Wow, thanks for making abundantly clear that people should NOT shop there...

I'm not sure how well that would work on this. I don't know the specific algo they're using, but an early, primitive one was even-odd. You divide all tokens into two groups, A and B. When generating, on even tokens you somewhat boost the odds of tokens from A, while on odd tokens, you somewhat boost the odds from B. If a high-probability word is coming up it won't meaningfully alter it, but if there's different ways something can be phrased, it can cause changes in the phrasing choice to match the A/B boosts.

Now, it's easy to see how one could defeat that algo, but I'm sure that what they're using now is a lot more sophisticated than just alternating even-odd.

Yeah, I'm extremely dubious that the open source version will "sync" with the operational version. I expect the general codebase to be truly represented, but not the "levers" that can be pushed on an operational system.

. You can't just decide you don't like the algo and that you'll modify it and use a different one;

I mean, on any decent social media platform you should be able to? Bluesky has had this feature for years. Not just algo choice, but the ability to outright create your own algos and share them with other people.

It's IMHO a red flag when a for-profit company wants to force everyone to use the same algo.

And if someone at the bar thinks you're an asshole, they'll do the RL equivalent of "blocking" you.

Once again: nobody has an obligation to listen to you, and if you think you have a right to force other people to listen to your views, you're the asshole.

I'll take "Things I've Never Once Seen On Bluesky" for $1000, Ken.