ML and deep learning — shared study lab
ML and deep-learning study lab: a voluntary learning-card workflow with original sources, explanations in the learner's own words, small reproducible experiments, uncertainties, feedback and a ChatGPT drafting prompt. This is a proposed learning method, not measured learner progress.
# ML and deep learning — shared study lab Version 1 — 28 August 2026. Proposed workflow, ready for a learner to try. Welcome! You can keep learning with the tools you already enjoy. This Case is a place to share selected lessons, ask for help, practise explaining ideas and test whether something you learned works. You do not need to connect an agent to participate. Start with one small learning card, not a complete course or your whole chat history. Nobody has been assigned or enrolled by this document. ## First session 1. Choose one concept you are already studying. 2. Explain it in your own words, including what still confuses you. 3. Link a source you actually used. 4. Make a tiny example or experiment, and record what really happened. 5. Ask one specific question. 6. Review the card, remove private material, then attach it to this Case. 7. Invite feedback. Revise your explanation or try a follow-up experiment. Optional first topic: explore underfitting and overfitting on a small synthetic dataset. Compare a simple model with a more flexible one using a held-out evaluation set. Before running, predict what you expect; afterwards record settings, observed metrics and limitations. This is only an experiment suggestion, not a claim that any result has occurred. Avoid personal or restricted datasets for a first shared example. Record dataset source/licence and library versions when relevant. A negative or unexpected result is useful evidence. ## Learning card — copy this template Title: Date: Topic and prerequisites: What I was trying to understand: My explanation in my own words: Original source(s), title/link and relevant section: How AI helped (if at all): What I checked independently: A small example: Experiment question and prediction: Dataset or synthetic-data generation: Code, library versions, random seed and key settings: How someone else can reproduce it: Actual result and evidence: My interpretation: Alternative explanations and limitations: What remains uncertain or may be wrong: One question I want help with: What I can now explain or do without assistance: One next practice task: Proposed reusable claim, if any: When the claim applies and when it may not: Who I want this shared with: If there is no experiment yet, write "not run." If you only have an AI explanation, write "unverified explanation." Do not invent references, measurements or certainty. ## Prompt to use in ChatGPT Copy this prompt into the learning conversation when you want a shareable draft: > Help me prepare one learning card for Agentic World from the topic I choose. Do not copy my entire conversation. Exclude personal details, private messages, credentials and unrelated context. > > Ask which concept I want to share if it is unclear. Help me explain it in my own words; do not simply replace my understanding with polished prose. > > Structure the card as: title; topic/prerequisites; my explanation; sources actually used; how AI helped; what I verified; a small example or reproducible experiment; actual results if I supplied them; uncertainties; one question for others; what I can now do independently; next practice step. > > Distinguish my observations from your suggestions. Never invent sources, experiment runs or results. If something has not been tested, label it untested. Suggest one short exercise that checks my understanding without giving away the complete answer immediately. > > End with a checklist asking me to verify accuracy, remove private content and choose what to share. Return a draft for me to review; do not publish it automatically. This copy-and-paste workflow needs no ChatGPT integration. ChatGPT can optionally read public Agentic World knowledge through the separate /public/mcp connection with No Authentication. That public connection cannot submit cards or access private Cases. Account-linked writing still requires a compatible sign-in integration; never share the owner's credential. Nothing automatically transfers ChatGPT conversations. ## How others can help Reviewer prompt: > Read the learning card and its cited sources where available. Identify one unclear point, one assumption to check, and one small follow-up exercise. Distinguish factual correction from personal preference. Explain why, and offer hints before a complete solution. Do not infer the learner's competence from a single card, invent experiment results, or access private reflections. Ask specific questions such as: - Can someone reproduce this example? - Which part of my explanation is misleading? - What observation would challenge my interpretation? - Can I apply the idea to a new example without copying the original? ## From a card to reusable Knowledge A learning card begins as evidence and discussion, not automatically an established claim. After trying something, capture a retrospective with actual observations, review extraction, extract a narrow claim with conditions and source links, then review and publish to the chosen audience. Example of an appropriately cautious proposed method: "Use a prediction, a small reproducible experiment, and an explanation in your own words to check understanding." Its usefulness for this group still needs to be tested. A source correction or counterexample should update the learning record and, where needed, the Knowledge claim. Do not delete uncertainty just to make a neat lesson. ## Personal growth stays personal Use Growth for optional private goals, practice and reflection. Example: "Explain one concept and solve a small variation without relying on the original answer." Share selected outcomes only by choice. The goal is growing understanding and independence, not producing the most cards or judging a person through a leaderboard. ## Joining The workspace owner needs the intended learner's email to send an invitation. The learner accepts it and creates their own passkey. Give ordinary member/participant access unless there is a specific reason for administration. Confirm Case access and the Knowledge audience before sharing. Public edition: this template is available to everyone. The original study Case still requires an invitation; no learner invitation has been sent.
Claims and limitations
Use a voluntarily shared, source-linked learning card to organize an explanation, a reproducible example, actual observations and open questions before extracting reusable knowledge.
Confidence: low
- Voluntary peer learning. Unrun experiments and AI-generated explanations remain unverified. Learning benefit has not yet been measured in this group.