5 ms·

"Our own hope is that, through AI, we can eventually approximate a 1:1 teacher:student ratio for every student in CS50" Would this not be a 0:1 ratio? Not agai

"Our own hope is that, through AI, we can eventually approximate a 1:1 teacher:student ratio for every student in CS50" Would this not be a 0:1 ratio? Not against LLM's for some coding assistance, I use it daily. , but equating a teacher with an LLM solution feels like calling a youtube video a 1 on 1 tutoring session. Well at least a semester at Harvard isn't costing you much ...

Taek 3y ago
An LLM is a million times better than a YouTube video because it can stop and re-explain anything specific that is confusing you, or have discussions about tangential but off-curriculum topics, ask you questions based on its own appraisal of your understanding and gaps, etc. Might not be a full teacher but its definitely an excellent teaching tool
- meroes 3y ago
  You’re not finding the right youtube teachers then. For example Professor Leonard https://www.youtube.com/@ProfessorLeonard https://www.youtube.com/@ProfessorLeonard will get anyone through Calc 1-3 and they will understand it well. I am extremely skeptical of the claim an LLM could do better. Even 20% would be asking a lot.
Paul-Craft 3y ago
I would say that an LLM could count as a "teacher" of sorts, so it would be more than a 0:1 ratio. But, if we're comparing LLM-teacher to human-teacher, I don't think it reaches full 1:1. I dunno, I'd say maybe 0.5:1 or 0.25:1 would be more accurate, given that the thing doesn't even produce correct information a certain percentage of the time? Based on my experience with ChatGPT, it does really well as a "rubber duck." [0] Sometimes, it even gives back useful suggestions. Sometimes, it's just so far off base, I wonder what planet it's getting its advice from. If I were the professor in this case, I think I'd suggest that students use it as a supplement: try LLM-TA first, look into its suggestions, and if it helps, great! If not, then try a few more things, maybe consulting LLM-TA a few more times. If that doesn't get you to working code and an understanding of the problem, then try human-TA. I think this represents something close to the optimal workflow for this tool, given the known limitations of the underlying technology. --- [0]: https://en.wikipedia.org/wiki/Rubber_duck_debugging https://en.wikipedia.org/wiki/Rubber_duck_debugging
- copperx 3y ago
  It sounds like an artificial hurdle to pass to get to the genuine article to reduce support/teaching costs. The only difference is that this hurdle is actually helpful some of the time.
hgsgm 3y ago
Harvards pricing system is such that it never costs much relative to student's family wealth.

Taek 3y ago

An LLM is a million times better than a YouTube video because it can stop and re-explain anything specific that is confusing you, or have discussions about tangential but off-curriculum topics, ask you questions based on its own appraisal of your understanding and gaps, etc. Might not be a full teacher but its definitely an excellent teaching tool

meroes 3y ago
You’re not finding the right youtube teachers then. For example Professor Leonard https://www.youtube.com/@ProfessorLeonard https://www.youtube.com/@ProfessorLeonard will get anyone through Calc 1-3 and they will understand it well. I am extremely skeptical of the claim an LLM could do better. Even 20% would be asking a lot.

Paul-Craft 3y ago

I would say that an LLM could count as a "teacher" of sorts, so it would be more than a 0:1 ratio. But, if we're comparing LLM-teacher to human-teacher, I don't think it reaches full 1:1. I dunno, I'd say maybe 0.5:1 or 0.25:1 would be more accurate, given that the thing doesn't even produce correct information a certain percentage of the time? Based on my experience with ChatGPT, it does really well as a "rubber duck." [0] Sometimes, it even gives back useful suggestions. Sometimes, it's just so far off base, I wonder what planet it's getting its advice from. If I were the professor in this case, I think I'd suggest that students use it as a supplement: try LLM-TA first, look into its suggestions, and if it helps, great! If not, then try a few more things, maybe consulting LLM-TA a few more times. If that doesn't get you to working code and an understanding of the problem, then try human-TA. I think this represents something close to the optimal workflow for this tool, given the known limitations of the underlying technology. --- [0]: https://en.wikipedia.org/wiki/Rubber_duck_debugging https://en.wikipedia.org/wiki/Rubber_duck_debugging

copperx 3y ago
It sounds like an artificial hurdle to pass to get to the genuine article to reduce support/teaching costs. The only difference is that this hurdle is actually helpful some of the time.

hgsgm 3y ago

Harvards pricing system is such that it never costs much relative to student's family wealth.