Misan
@misanchan.bsky.social
What will be, unburdened by what will have been.
Assume that the students can use an LLM in a test. What are the questions to ask in that test?
I don’t know the answer, but it’s part of the educator’s job to figure that out.
I don’t know the answer, but it’s part of the educator’s job to figure that out.
November 23, 2024 at 2:27 PM
Assume that the students can use an LLM in a test. What are the questions to ask in that test?
I don’t know the answer, but it’s part of the educator’s job to figure that out.
I don’t know the answer, but it’s part of the educator’s job to figure that out.
Trust will not work, I can see that. The real solution is to shape the rewards correctly (which is hard).
You wouldn’t ask a student to compute 49206 x 19581 by hand anymore, because there’s a calculator for that. In the same sense, the challenge for educators is to keep up with the tech.
You wouldn’t ask a student to compute 49206 x 19581 by hand anymore, because there’s a calculator for that. In the same sense, the challenge for educators is to keep up with the tech.
November 23, 2024 at 2:27 PM
Trust will not work, I can see that. The real solution is to shape the rewards correctly (which is hard).
You wouldn’t ask a student to compute 49206 x 19581 by hand anymore, because there’s a calculator for that. In the same sense, the challenge for educators is to keep up with the tech.
You wouldn’t ask a student to compute 49206 x 19581 by hand anymore, because there’s a calculator for that. In the same sense, the challenge for educators is to keep up with the tech.
Claude is a better teacher than any other teacher I’ve had, but you have to ask it to teach you. That takes discipline, and discipline comes with awareness.
Saying “No LLMs in education” is throwing the baby out with the bathwater.
🧵 3/3
Saying “No LLMs in education” is throwing the baby out with the bathwater.
🧵 3/3
November 23, 2024 at 2:12 PM
Claude is a better teacher than any other teacher I’ve had, but you have to ask it to teach you. That takes discipline, and discipline comes with awareness.
Saying “No LLMs in education” is throwing the baby out with the bathwater.
🧵 3/3
Saying “No LLMs in education” is throwing the baby out with the bathwater.
🧵 3/3
Re: the second facet — you can use LLMs to solve a problem, or you can use LLMs to learn how to solve a problem. The first can be problematic, but the second is a massive unlock.
🧵 2/3
🧵 2/3
November 23, 2024 at 2:12 PM
Re: the second facet — you can use LLMs to solve a problem, or you can use LLMs to learn how to solve a problem. The first can be problematic, but the second is a massive unlock.
🧵 2/3
🧵 2/3
There are two facets to the calculator analogy: one is complexity of implementation (obvious) and the other is how much sense it makes if implementation is magically taken care of (less obvious).
🧵 1/3
🧵 1/3
November 23, 2024 at 2:12 PM
There are two facets to the calculator analogy: one is complexity of implementation (obvious) and the other is how much sense it makes if implementation is magically taken care of (less obvious).
🧵 1/3
🧵 1/3
This is like asking Texas Instruments to install anti-cheating software on their calculators.
November 23, 2024 at 3:37 AM
This is like asking Texas Instruments to install anti-cheating software on their calculators.
We’ve had some time to build the cognitive muscles to filter search results, and we’ll need some more to build different muscles for dealing with LLMs. But at the end of the day, markets will deal with the laggards and luddites.
November 23, 2024 at 3:35 AM
We’ve had some time to build the cognitive muscles to filter search results, and we’ll need some more to build different muscles for dealing with LLMs. But at the end of the day, markets will deal with the laggards and luddites.
Ads and product placements, I’m afraid.
November 23, 2024 at 3:15 AM
Ads and product placements, I’m afraid.
It isn’t, though I imagine it’s possible to develop an intuition about how fast an architecture could be if implemented well.
November 21, 2024 at 9:39 AM
It isn’t, though I imagine it’s possible to develop an intuition about how fast an architecture could be if implemented well.
Loved that post. + I sense a fun connection to the fact that neural nets learn the lower frequencies in data before higher frequencies (ie they autoregress in spectral domain).
November 19, 2024 at 4:21 PM
Loved that post. + I sense a fun connection to the fact that neural nets learn the lower frequencies in data before higher frequencies (ie they autoregress in spectral domain).
Generalist systems are hard to benchmark, we’ve seen this for LLMs. Agents (including humans) have an even larger action space, so I’m less optimistic about benchmarks. But that can be a blessing in disguise — nothing to Goodhart against.
November 19, 2024 at 3:09 PM
Generalist systems are hard to benchmark, we’ve seen this for LLMs. Agents (including humans) have an even larger action space, so I’m less optimistic about benchmarks. But that can be a blessing in disguise — nothing to Goodhart against.
BFL is the rare German shop not full of shit, please don’t steal that from us.
November 19, 2024 at 2:57 PM
BFL is the rare German shop not full of shit, please don’t steal that from us.
Only gods control their own destinies.
November 16, 2024 at 10:03 PM
Only gods control their own destinies.