I think new words is the right approach in general, but golem.
778 karma · joined January 1, 2023
\aside: Adam Smith in The Wealth of Nations has a story about men who worked themselves into the ground, because they were paid by-the-piece (not per hour). They needed to be protected from themselves. (The psychological mechanism wasn't speculated, but I suspect maybe competition rather than greed.)
Here's an elaboration of the point: https://news.ycombinator.com/item?id=37341439
By analogy, if LLMs were allowed in an exam, what harder tasks would be assessed?
But writing enables arrows, lines, crossing-out, small-text notes, circling, variable pressure, colour, etc. Richer than ordering/indenting text. Also higher contrast.
\aside Some exams allow calculators and textbooks, and are really hard. Consider: what would an exam assess, where LLMs were allowed?
Perhaps the experience of symbolic calculators (like mathematica) is informing, since surely, for many years now, students have been able to have it do their homework for them. How do teachers handle that?
Organizing and arguing your own thought would be a good test. I think an assisting tool could still be reasonable - choosing between different organizations. Though it's unclear how to assess for original thought - the only such "essays" I know of are PhD theses.
People said similar about log tables and slide rules. And they were right - something was lost (e.g. the sense of nearby solutions). Yet here we are.
Similar to today, when grades are a proxy for ability, but private tutoring puts an average student in the 2% (Bloom's 2σ problem) for that stage, but doesn't boost the student's general intelligence for the next stage. Also, hard work, self-discipline, and focus will increase grades but not GI. (Of course, stidents do learn that specific stage, necessary for the next stage, so this criticism is just for use as a proxy).
We might say what is really be evaluated is the ability to get good grades (duh) - whether through wealth or work.
The same argument can be applied to LLMs. Using they is an important ability... so let's test that. This is the future. Similar to calculators and open-book exams.
I think today's explosion of LLMs etc will lead (eventually) to actual understanding. Data precedes theory.
You're thinking of nuclear waste? Not representative of today's time capsule, any more than space shuttle software development is representative of software development in general.
Instead of moving goalposts, let's accept that each one succeeded at mechanizing intelligence. Including boolean algebra, writing and arithmetic.
BTW TFA uses "feelings" in two different senses: hunch, inkling, intuition, suspicion, gut instinct vs. emotion (as in music).
In mathematics, there is even the formal term "conjecture" for an unproven theorem.
But now I realize that, in maths research, you don’t just follow logical steps. If you do, you’ll never get anywhere interesting. You have to use your gut instinct and feel your way through something first, and then back it up with logic afterwards.
TBF, it's hard to load proofs on to school students, and since Euclid was dropped, the focus is on technicians and results - not truth. But it can be done: spend the time on building foundations; or omit subjects where you can't. I'd like an axiomatic approach.
Theoretically, the better method could have been worked out from the beginning, but actual progress doesn't seem to happen like that. It may be because a working model enables faster experiments (iteration) to search the space-of-methods faster, and in particular to prune off dead-ends, making the search-space smaller.
Parallelizable methods win (e.g. DL), by using the computation available
One effect is like how Latin used to be the language of professionals, so you'd better learn Latin if you want to access the knowledge base.
...you will anyway (Brooks)
TFA stresses: finding out what you're building; finding the "unknown unknowns" (the things you don't even know you don't know, encountering those problems)
TFA doesn't say it this way, but I think the fun of throwaway-prototyping is your thoughts can focus uninterrupted on the problem, undistracted by secondary issues. You can hack-around tedious parts, now knowing they are there; and you can get absorbed in the genuinely tricky parts, unobscured (even if you don't succeed, you also know they are there). You're not expected to get it all right.
It's turning an unexperienced developer (on this problem) into an experienced one.
PS Somewhat disturbingly, Brooks went back on this according to this interview https://www.computerworld.com/article/2550685/the-grill--fre...
"you should plan to throw one away. You will anyway."
That was the first edition of The Mythical Man-Month. In the second edition, I say that was misguided! You ought to plan to continually iterate on it, not just build it, throw it away and start over. Some of the things I said in 1975 were wrong, and in the second edition, I correct them.