Did you know Bloom's very famous "two-sigma problem" is widely misinterpreted - and often intentionally by EdTech apps. 1. Bloom performed zero experiments himself. He simply summarized two (small) PhD dissertations by his own students (Anania and Burke). These authors are rarely even mentioned or known. One dissertation was never published at all; the other has been cited only 77 times. Neither student ever followed up with replications or further tutoring research. 2. The iconic +2 sigma bell-curve graph everyone quotes is not real data, it's a hand-drawn illustration Bloom created to “show what it might look like.” 3. Bloom’s data distorts even his own cited source (Herb Walberg’s meta-analysis from one month earlier). Walberg reported ordinary tutoring at just 0.40 sigma. Bloom relabelled it narrowly as “peer tutoring,” and then inserted the 2.00 value to put tutoring at the top. 4. In his student's studies, topics were both deliberately narrow and also brand-new to the students involved (probability for 4th/5th graders, cartography for 8th graders). Of course novices on unfamiliar material produce outsized short-term gains, but do they generalise with wider topics and/or long term? 5. Tests were custom, researcher-created, and tightly matched to the exact 3-week lessons. They were not standardised or broad (SAT, state tests, etc.). Effects on such narrow tests are routinely 2–3× larger than on real standardised measures. This is 99% of EdTech summed up: apps say users are learning great to keep them paying and sharing with others. Duolingo will race to tell users that they are approaching fluency. 6. The entire “intervention” lasted only 3 weeks (!) (11 × 40-minute lessons) with immediate post-testing with no long-term follow-up 7. The “tutoring” condition was not pure one-on-one tutoring. It was a mix of one-on-one plus strict mastery thresholds. Roughly half the 2-sigma gain (~1.1 sigma) came from the mastery-aspect and informative feedback alone 8. Modern meta-analyses of actual randomised studies find average effects of only 0.33–0.37 sigma across 65–96 RCTs. Zero large-scale studies have ever replicated 2 sigma. 9. +2 sigma is crazy in real terms outside of very specific contexts (new material, cocktail approach, very short term): it would turn a median student into the 98th percentile or deliver 5+ years of school learning in three weeks. 10. He conflates formative assessment (short-term feedback quizzes) with summative claims about broad, long-term achievement. Once again, EdTech. 11. +2 sigma gains do exist, but only at the system level over several years via excellent whole-class instruction (e.g., England’s top Progress 8 schools on broad subjects across 600,000+ students). Not from short-term individual tutoring interventions. 12. Post-COVID governments poured huge amounts of money into programs they thought were explicitly justified by Bloom. Results: 0.06 sigma in most districts, at best 0.16–0.37 sigma in the best structured programs educationnext.org/two-sigma-… substack.nomoremarking.com/p…

Unfortunately many teachers (and thus students) have no clue on what effective studying looks like (see last pic) Borderline criminal when you actually think about it. Instead there's hours every day wasted on fluff techniques that barely move the needle but feel effective. Where else can you find a legitimate 2-3x increase in A-grades from 1 oft-disregarded method? shaeda.blog/s/research

Have they published data on its estimation accuracy for non-trivial meals? Fundamental problems such as lack of scale (is it 100g or 300g?) and hidden ingredients (butters, oils, sauces, food items underneath the top layer etc) are not AI problems they're camera problems. How can an LLM solve that? Genuine question.