Teach Your Past Self

The description states that Teach Your Past Self is an AI-powered educational tool designed to help students learn by teaching a simulated AI student. That AI student is programmed to fail exactly like the user did, based on misconceptions extracted from their own incorrect homework.

• It accepts a photo of handwritten homework.
• GPT-5.6 extracts a structured misconception profile (e.g., "believes (f·g)′ = f′·g′").
• This profile is implanted into an AI student via a belief-state engine with three conviction levels: held → shaken → corrected.
• The AI student takes a baseline exam, fails like the user did, and then can be taught by the user.
• A final exam shows whether the user's teaching led to conceptual correction.
• The system includes a live sidebar showing belief status and conviction updates.

Inference The system is designed to simulate the protégé effect — where teaching something improves one’s own understanding. This is claimed to be measurable via score deltas between baseline and final exams.

The project positions itself as a tool that makes learning measurable by simulating a user's own past mistakes in an AI student. It claims to use the "protégé effect" — a well-known educational theory — but applies it through a novel mechanism: using real, personal homework to generate a digital twin of one’s misconceptions.

• The tagline states: “An AI student implanted with YOUR real misconceptions from a photo of your wrong homework — it fails exactly like you did, until you teach it. Its score = proof you truly understand.”
• The authors reference prior work such as Betty’s Brain and ALTER-Math.
• The project is described as a solution to the problem that generic teachable agents don’t reflect personal errors.

Inference The positioning implies a focus on personalized, self-directed learning through AI feedback loops. However, no evidence is provided that this approach has been validated in practice or scaled beyond the hackathon prototype.

The description does not explicitly define a target customer or ideal customer profile (ICP). It suggests the tool is for students who want to learn by teaching — particularly those with homework they’ve gotten wrong and want to understand why.

• The system is described as useful for math, with examples of calculus misconceptions.
• It includes a “common-misconceptions mode” for future development, suggesting broader applicability.

Inference The ICP likely includes students in STEM fields who are looking for personalized feedback on conceptual understanding. However, no explicit segmentation or targeting data is provided.

There is no evidence of a business model or pricing structure in the description.

• The project is described as a hackathon submission.
• No mention of monetization, subscriptions, or paid features.
• No indication of whether it’s intended for individual users, schools, or institutions.

Inference If this were to become a product, it might target educational platforms or SaaS tools for students. However, no such plans are evident in the description.

The project is described as built entirely with Codex and GPT-5.6 across five focused sessions, using structured outputs and spec-first prompting.

• The system uses a belief-state engine to maintain AI student stubbornness.
• It includes a quiz generator, grader, and extractor modules.
• The app is built on Next.js 16 + TypeScript + Tailwind on Vercel.
• Features include KaTeX rendering, localStorage session restore, and live belief-state sidebar.

Inference The technical architecture is described as modular and engineered to avoid common AI pitfalls like self-healing or knowledge leakage. However, no evidence of production-grade infrastructure or scalability is provided.

There is no evidence of traction, revenue, or adoption beyond the authors’ own claims.

• The project is a hackathon submission.
• No mention of users, customers, or usage metrics.
• No data on retention, engagement, or product-market fit.

Inference The tool is described as complete but lacks any evidence of real-world usage or market validation. It remains an experimental concept, not a product in use.

The description references prior work such as Betty’s Brain and ALTER-Math, which are known to explore teachable agents in education.

• The project claims to improve on generic agents by using personal homework data.
• It is positioned as a novel way to apply the protégé effect through AI simulation.

Inference The competitive landscape includes educational AI platforms, but no specific competitors are named. The tool’s novelty lies in its personalization via homework photos and belief-state simulation.

Several risks and red flags emerge from the self-reported description:

• No traction or validation: The project is a hackathon submission with no evidence of real-world adoption.
• Unproven scalability: The system is described as built in five sessions using Codex — not a scalable engineering approach for production use.
• AI behavior risks: Maintaining AI stubbornness and preventing knowledge leakage are described as hard challenges, but no evidence of robust solutions beyond testing.
• Limited scope: The tool is currently focused on math; future expansion to CS/DSA is mentioned but not demonstrated.

Inference The tool may be a promising concept but lacks real-world testing or commercial viability. It remains in early-stage development.

1. What are the actual user acquisition and engagement metrics?
2. How does the system handle edge cases like unclear handwriting or no misconceptions found?
3. Is there any data on how often users actually teach the AI student correctly vs. incorrectly?
4. What is the plan for monetization, if any?
5. How do you intend to scale beyond the current prototype and hackathon environment?
6. Are there any partnerships or institutional pilots planned?
7. How does the system ensure that teaching is effective and not just a placebo effect?