Optimizing Feedback to Learn Faster

post by Towards_Keeperhood (Simon Skade) · 2025-02-26T14:24:26.835Z · LW · GW · 0 comments

Contents

  Theory
    Accurate
    Fast
    Rich
  Exercise
None
No comments

(This post is to a significant extent just a rewrite of this excellent comment from niplav [LW(p) · GW(p)]. It is one of the highest-leverage insights I know for learning faster.)

Theory

To a large extent we learn by updating on feedback. You might e.g. get positive feedback from having an insight that lets you solve a math problem, which then reinforces the thought patterns that lead you to this insight.

Key insight: Feedback should be accurate, fast, and rich.

Accurate

This one is obvious: The feedback should rarely be bad. Learning from an accurate evaluation source is much faster than learning from a noisy one.

Example: If you research with an advisor who gives you feedback on your ideas as you have them, you learn faster if the feedback is always good than if the advisor sometimes makes mistakes.

Fast

The faster you get the feedback for something you have done, the fresher are the thought patterns you used in your mind, the better you can update.

Example Applications:

• If you learn for an exam by calculating through a full practice exam, and then take out the practice exam solution and see what you did wrong, this is inefficient. Much better to stop after every small exercise and check whether you got it right or what mistake you made.
• If you want to learn making calibrated estimates even in cases where motivated reasoning might apply, you can make an estimate on whether your proposed solution is correct briefly before you look up the solution.
• If there are people more experienced than you (in a particular subfield), get feedback from them as often as you can manage (e.g. often write quick drafts to share).
  • Perhaps even try to convince them to watch you work and give you feedback for what kind of mental moves might be better and why.^[1]
    • E.g. for learning programming, learning to solve math problems, learning to write emails well, ...

Rich

When someone tells you “you screwed up”, this is less useful than when someone tells you “you screwed up, here’s what you should’ve done instead …” (assuming accurate feedback). Supervised learning is faster than reinforcement learning.

Example Applications:

• After you solved a problem, review what the key insights were, and if possible also how you found them, and also where you wasted time on unproductive angles. Ask yourself how you could’ve succeeded faster. What were the key properties of the thinking avenues which were or were not useful? (Related [LW · GW].)
• (Advanced: The more precise you know what thought patterns you use and ought to have used, the faster you can learn. (Related [LW · GW].))

(Recall reminder for reader: You have the option to pause and quickly summarize what you learned from reading this post so far, to better consolidate your understanding.) 

Exercise

1. Pick a skill you currently want to learn/improve-upon.^[2]
2. Clarify the mechanism of how you currently get feedback to improve that skill.
3. Optimize your feedback. Answer the following questions. You might want to initially relax constraints by e.g. asking "If I had unlimited money for learning this, how could I get more accurate/fast/rich feedback?". (If you feel like you've already maxed out in a dimension, feel free to skip this dimension.)
   1. How could you get more accurate feedback?
   2. How could you get faster feedback?
   3. How could you get richer feedback?
4. Clarify your plan for how to learn that skill.^[3]
5. (Optional:) Post your notes from this exercise in the comments.
6. (If you can afford the time, please share feedback in the comments about how this exercise worked for you and about what parts you found difficult.)

1. ^{^}

   I recommend telling your tutor to just say thoughts that are relatively close to your current skill level, so you sorta could've thought of that. Your tutor shouldn't just go ahead and present the advanced solution which you didn't have a chance to find.
   Btw, I didn't test whether this kind of tutoring works well, but I'd guess it does, at least if you have a decent tutor.

2. ^{^}

   If you don't have one, change your life so you always have at least one.

3. ^{^}

   Bonus points if you apply the murphyjitsu technique [LW · GW] (shorter post [LW · GW]) to optimize your plan.

0 comments

Comments sorted by top scores.