I am surprised this book is not mentioned. Talks about how optimization doesn't lead to optimal goals in complex environments (e.g. life).
https://www.amazon.com/Why-Greatness-Cannot-Planned-Objectiv...
The Lesson to Unlearn
111–120 of 588 posts
Re: The Lesson to Unlearn
#112My wife and I are contemplating unschooling our kids. The longer our son is in primary school, the more we see the deterioration of his willingness to learn anything school-related. He hates homework (as do most kids), and this is even more painful to see given that it has been shown that homework is essentially useless for learning performance. He still likes to write and do math, but as long as he is in school, we…
> Gates, ... - all were not in a conventional school
wikipedia said he went to a regular private school then went to Harvard which sounds pretty conventional. Yeah he didn't finish, but worry about that later.
Re: The Lesson to Unlearn
#113My wife and I are contemplating unschooling our kids. The longer our son is in primary school, the more we see the deterioration of his willingness to learn anything school-related. He hates homework (as do most kids), and this is even more painful to see given that it has been shown that homework is essentially useless for learning performance. He still likes to write and do math, but as long as he is in school, we…
We thought the same way but it’s difficult to replace the social component. Still it’s hard to see how bad the school system is and force our kids to participate. It could be so much better.
Not these days. There are so many home school groups that it's very easy to replace the social component.
Re: The Lesson to Unlearn
#114Forget the grade. Impress your profs. Let them learn from you.
For example, you could work on analog computers (look it up). The early AI systems were all analog, using motors on "volume knobs" to adjust the weights.
Analog is MUCH faster than digital. You could probably train an analog neural net "at movie speeds" because it all happens in parallel. Analog weights are trivial (capacitors), backprop is just differentiation which analog does trivially. Function inversion just involves putting the function in the feedback loop of an op-amp. A board with 100x100 mini-caps would be dirt cheap and would have 10,000 "weights".
Besides being "up to speed on AI NNs" you would also be breaking new ground (well, really old ground but nobody remembers anything before last year).
Imagine talking the department into an "analog AI lab" in cooperation with the EE dept and you get to be "the person". Heck, now you've nearly "joined the department" rather than being a student.
Think of college as a big company that lets you do ANYTHING. Profs have to prove themselves to get tensure. Students have to "get good grades" and "learn to hack the FAANG interviews so they can be a 'good cultural fit (aka grade hackers)". But YOU are at college as a "free agent" and can use all the resources to excel at what you love. Imagine having 4 years at a company with the resources of a college that lets you do ANYTHING.
You will never have it this good again. Forget the grade.
Re: The Lesson to Unlearn
#115> In theory you shouldn't have to prepare for a test in a class any more than you have to prepare for a blood test.
The key here is the “in theory” part. It is very hard to design exams that actually test your understanding of some topic as opposed to your memorization skills.
As somebody who recently got his bachelor’s CS degree (currently working on my master’s degree), and who also puts more emphasis on actually learning rather that getting good grades, I can tell you that understanding a subject might be enough to pass, but is seldom enough to get good grades.
Re: The Lesson to Unlearn
#116You need a metric to measure performance. It's hard to get such a metric right. You need this metric because the government institutionalized schooling and need to have aggregate data on performance.
However, for the student, it'd be better if they'd focus on the intersection of what they're interested in and what is societally relevant and not to be graded at all.
I know a friend of mine who went to a school for 1 year where he didn't get graded, all he learned was soft business skills. I can clearly see that he grew a lot from it. But I can also see that it is tough for anyone else to see that he learned anything at all. Partially this is because he isn't graded and the behaviors he picked up are now natural to him and it just looks like it was always there.
Re: The Lesson to Unlearn
#117I'm glad he made the leap from school tests to funding tests, as it seemed to be a thinly veiled analogy from the start. I'm disappointed, though, that he didn't take the final step and admit that money itself is also a poor, hackable test. Funding, growth, usage -- these are all still one level removed from something worthwhile or beneficial. Cigarettes have amazing usage numbers even in 2019. I'm waiting for the en…
How about DHH? He is the closest to this I know.
Re: The Lesson to Unlearn
#118This is a great essay. But, I find it odd that YC's famous "cohorts" are invented by PG. Cohorts work because startups in that batch compare against each other, which in itself is a grading system. Getting into YC, topping the batch and raising money are inevitable symptoms of "cohort" thinking. We could argue YC uses cohorts to make startups compete for meaningful metrics instead of vanity metrics. But still, YC is based on this innate human nature to compete and get social acceptance.
Re: The Lesson to Unlearn
#119Are you in EECS? Most big data centers use FPGAs "on the wire" to do things like encryption on the fly or compression so that they get "wire speedups" without burning CPU cycles. So FPGA solutions that "live on the wire" and can be "plugged into the LAN port" are a really useful area you can hack in your dorm room. I worked in the security area. One of the key problems is "exfiltrations" where someone tries to copy v…
Wrong thread?
or a very overtrained neural net machine commenter
Re: The Lesson to Unlearn
#120Yes, most tests are bad, but there is a quick advice that improves them immensely: never ask questions that are not at least Level 3 in Bloom's taxonomy of knowledge: https://blog.testdome.com/blooms-taxonomy/ I cofounded a technical screening startup, and despite our efforts to educate our customers that enter their questions on our platform, 95% of their custom questions are bad questions. They tend to ask trivia q…
can anybody comment on the difference between "understanding" and "applying" in that taxonomy. The examples provided don't seem very good to me. is it the difference between understanding a language and not speaking it, perhaps? following a conversation, but not being able to participate in it?
I have learned Bloom's taxonomy with comprehension instead of understanding. So if in an algorithm's class, you were taught that algorithm X was exponential time, then the exam question at the comprehension level can be, "describe why does it take exponential time to run algorithm X".
An applying level question could be, "Find the running time of algorithm Y (which was never discussed in class)."