Aristotle 🍎

Tiny but mighty! We are actively hiring agent engineers and generalists to build the world's best voice-based tutoring platform. Join us! heyaristotle.com/careers

1/ The 2024 Nation's Report Card came out in January and the headline is hard to read. 40% of U.S. fourth graders scored below NAEP Basic in reading. For 8th graders, 33% — the highest share ever recorded. 2/ Both grades dropped another 2 points between 2022 and 2024. This isn't a pandemic story we're recovering from. It's a pandemic story that's still getting worse. 3/ The 4th-grade reading average is now statistically equivalent to 1992. Thirty-two years of work. 4/ Worth sitting with: the bottom of the distribution is moving more than the top. We are not losing ground evenly. We are losing it at the floor. 5/ Interventions that could move that floor at population scale have to be cheap, available outside school hours, and consistent week over week. Very few things meet all three. It's why we follow AI tutoring research as closely as we do. 6/ Source: NCES, The Nation's Report Card — 2024 NAEP Reading Assessment Results, Grades 4 and 8 (Jan 2025). nationsreportcard.gov/reports/readin…

1/ Across 96 randomized trials, structured tutoring delivers an average effect size of 0.37 SD on student achievement. Higher than almost every classroom-level reform we've ever rigorously tested. 2/ Trained teachers and paraprofessionals do best — around 0.58 SD. Volunteers do half that. Parents tutoring their own kids barely move the needle. 3/ Dosage matters: three sessions a week, 30 minutes minimum, was the precondition for all of the above. 4/ The hardest part of high-dosage tutoring has never been figuring out that it works. It's been finding enough qualified tutors, in front of enough kids, three times a week, for long enough to matter. 5/ That math has never closed at scale. It's the math AI tutoring is trying to change. 6/ Source: Nickow, Oreopoulos & Quan, "The Promise of Tutoring for PreK–12 Learning," AERJ 2024; NBER WP 27476. nber.org/papers/w27476

1/ Anthropic analyzed 574,740 academic conversations between university students and Claude. The use cases don't look like the cheating panic. 2/ 39% of conversations: students creating or improving content — drafting essays, generating practice problems, building study guides. 33%: technical explanations and code. The rote-recall, "answer my homework" use case barely registers. 3/ Mapped to Bloom's taxonomy, students spend almost 70% of their AI time in "Creating" and "Analyzing" — the top of the cognitive ladder. They're not offloading memorization. They're offloading higher-order work. 4/ Roughly what we see when a teen sits down with a good AI tutor. They don't ask for the answer. They ask *wait, why does that work?* — and then keep pulling the thread. 5/ The pedagogy that fits is closer to Socratic teaching than spoon-feeding. The take-home essay as an assessment is mostly dead. What replaces it has been around since Plato. 6/ Source: Anthropic Education Report — "How University Students Use Claude" (April 2025). anthropic.com/news/anthropic…

1/ Across 96 randomized trials, structured tutoring delivers an average effect size of 0.37 SD on student achievement. Higher than almost every classroom-level reform we've ever rigorously tested. 2/ Trained teachers and paraprofessionals do best — around 0.58 SD. Volunteers do half that. Parents tutoring their own kids barely move the needle. 3/ Dosage matters: three sessions a week, 30 minutes minimum, was the precondition for all of the above. 4/ The hardest part of high-dosage tutoring has never been figuring out that it works. It's been finding enough qualified tutors, in front of enough kids, three times a week, for long enough to matter. 5/ That math has never closed at scale. It's the math AI tutoring is trying to change. 6/ Source: Nickow, Oreopoulos & Quan, "The Promise of Tutoring for PreK–12 Learning," AERJ 2024; NBER WP 27476. nber.org/papers/w27476

1/ Two surveys, one school year apart. RAND in 2023–24: 25% of U.S. K–12 teachers used AI for work. Walton × Gallup in 2024–25: 53%. 2/ Teachers who use it weekly say they save an average of 5.9 hours a week. About six work-weeks back per school year. 3/ Top use cases aren't exotic. Lesson prep. Worksheets and activities. Modifying materials for students with different needs. 57% of teachers expect AI to improve accessibility for students with disabilities. 4/ The number that should keep school leaders up at night: 68% of teachers got zero AI training in 2024–25. 5/ Teachers are adopting a serious new classroom tool faster than their districts can teach them how to use it well. 6/ A good AI tutor isn't a replacement for what great teachers do. It's a thing that takes load off, so teachers can spend more of their time on what only they can do. 7/ Six weeks of teacher time per year is enormous. The next decade is about whether districts learn to spend it well. 8/ Sources: Walton Family Foundation × Gallup, "Teaching for Tomorrow" (2025); RAND American Educator Panel (2024). waltonfamilyfoundation.org/six-weeks-givi…

vectorizing hand-drawn icons... our app is gonna be unstoppably cozy

1/ Bloom's "two-sigma problem" (1984): 1:1 tutoring with mastery learning could move the average student to the 98th percentile. The catch was always cost. So we never did it at scale. 2/ Three randomized trials in the last two years are quietly chipping at that. Harvard physics undergrads with a GPT-4 tutor learned 2x more than the active-learning control, in less time. 3/ A World Bank RCT in Edo State, Nigeria gave 1,500 secondary students six weeks of GPT-4 tutoring — gains worth 1–2 years of typical schooling. 4/ Bloom's number is 2.0 SD. None of these get there. But Hattie calls 0.40 SD impactful, and the Harvard study landed at 0.73. 5/ The right bar isn't 0.3 or 0.7. It's whatever the best 1:1 human tutor your kid will ever have produces. That's a number we won't have for years. The work is to build toward it honestly. 6/ Sources: Bloom (1984); Kestin et al. 2025 Sci Reports; De Simone et al. 2024 WB WP 11125; Nickow, Oreopoulos & Quan 2024 AERJ. nature.com/articles/s4159…

Source: Epoch AI benchmark tracker; OpenAI & Anthropic system cards (2023–2026). epoch.ai/benchmarks/gpq…

The question we keep coming back to: when the substrate gets this good this fast, what does a great tutor — human or otherwise — actually do with the time?

Three years ago GPT-4 scored 39% on GPQA Diamond — graduate-level physics, chem, bio questions written so Google can't help. PhDs in their own field score ~70%.

I can at least guarantee everyone on 24th st Noe valley at 11am knows about our company

🦾Meet Aristotle Agent, the world’s first autonomous mathematician — live and currently free of charge. We designed Aristotle Agent to solve and formalize the world’s most challenging mathematical research problems. It is now: ☑️#1 in Formal Math: We’re the #1 formal math model according to ProofBench, by @ValsAI, ahead of the closest competitor by 15%. Aristotle Agent can autonomously prove/formalize for up to 24 hrs without human intervention. ☑️Fully Agentic: Give it an English problem and it will prove/formalize from scratch, or it can work and edit files directly inside your Lean project / repository. ☑️Github-ready: Aristotle agent produces repo-quality code; project leads are increasingly merging Aristotle-drafted PRs with no modifications. Now live across both web, CLI, and API. 🔥

most AI "tutors" just give kids the answer faster. that's not tutoring. that's cheating with extra steps.

@EndorsiVal @gabriel1 :)

@gabriel1 an adaptive voice based AI learning platform already exists btw

i could learn any topic in 5 minutes with the mist optimal text & visual explanation, if it's adapting live to what you do and dont understand fundamentally ai has like another 5 gpt3 to gpt4 moments ahead

12,000+ tutoring sessions completed. 96% of students improve within 2 weeks. Every session reviewed by a real human. Aristotle is a voice-based AI tutor built by Stanford learning scientists with 10,000+ hours of real tutoring experience. It doesn't give answers — it teaches. First session free, no credit card: heyaristotle.com

A private tutor costs $50-250/hour. Aristotle gives your kid unlimited 1-on-1 tutoring — voice-based, adaptive, with a real-time whiteboard — for less than a single hour with most tutors. It remembers what your kid knows, adapts when they're confused, and never just gives the answer. 12,000+ sessions. 4.9 parent rating. First session free. heyaristotle.com

Your kid is using ChatGPT for homework. The problem: it just gives answers. Your kid copies them and they learn nothing. The test comes, and they flounder. We built Aristotle to be different. It teaches — step by step, out loud, on a shared whiteboard. It refuses to just hand over the answer. Every session reviewed by a real person. 96% of students improve in 2 weeks. Try it free: heyaristotle.com