Turns out when you call for ending H1Bs,
On a platform run by thousands of H1Bs,
Suddenly accounts start getting mass censored
🄳🄾🄾🄼🄿🄾🅂🅃🄸🄽🄶
On a platform run by thousands of H1Bs,
Suddenly accounts start getting mass censored
🄳🄾🄾🄼🄿🄾🅂🅃🄸🄽🄶
🤬5👀5👏1💯1
Smarter people -> Better grades, within a major
Smarter people -> More legit major
More legit majors -> lower grades, across majors
= Simpson’s paradox situation
🄳🄾🄾🄼🄿🄾🅂🅃🄸🄽🄶
Smarter people -> More legit major
More legit majors -> lower grades, across majors
= Simpson’s paradox situation
🄳🄾🄾🄼🄿🄾🅂🅃🄸🄽🄶
💯8👀5
AI coding, so close and so far
Weirdest part is many of the hugest problems are extremely fixable
…But OpenAI refuses to fix them
E.g. extremely avoidable but huge problems like:
+ AI’s bad habit of silently swallowing all errors, opposite of fail-fast, one of the ways the AI often cheats
+ AI’s bad habit of terribly vague variable names, leading to it getting completely mixed up about what the units and types the variables contain, just like a human would
+ AI’s bad habit of constantly telling itself that code is “correct” or “fixed” without testing at all whether this is true — and then the dumb thing stubbornly believes its own claims of something being “correct”, leading to a huge fight with the AI to fix things it insists that are not broken
+ Total refusal to properly test code — total failure to understand what the general form of proper code testing even should be, to limit potential cheating, which combined with its bad habit of silently swallowing errors, leads it to often completely cheating on tests
+ Very little in the way of consistent coding pattern habits, which hugely helps in knowing how code should be or should work even when you can’t see it - which is critical to handle context size challenges - and the lack of which leads to the pieces of larger codebases the AI creates not working together at all, despite that they easily could if consistent coding patterns were used
^ Probably could list dozens more huge ones like these.
All would be trivial for OpenAI to fix… but they don’t
Did everyone but the low-tier H1Bs who never did any real coding on their own check out of OpenAI?
Would fix these myself and turn it into a product, if many people were willing to pay, though guessing that’s not hugely likely
Someday
🄳🄾🄾🄼🄿🄾🅂🅃🄸🄽🄶
Weirdest part is many of the hugest problems are extremely fixable
…But OpenAI refuses to fix them
E.g. extremely avoidable but huge problems like:
+ AI’s bad habit of silently swallowing all errors, opposite of fail-fast, one of the ways the AI often cheats
+ AI’s bad habit of terribly vague variable names, leading to it getting completely mixed up about what the units and types the variables contain, just like a human would
+ AI’s bad habit of constantly telling itself that code is “correct” or “fixed” without testing at all whether this is true — and then the dumb thing stubbornly believes its own claims of something being “correct”, leading to a huge fight with the AI to fix things it insists that are not broken
+ Total refusal to properly test code — total failure to understand what the general form of proper code testing even should be, to limit potential cheating, which combined with its bad habit of silently swallowing errors, leads it to often completely cheating on tests
+ Very little in the way of consistent coding pattern habits, which hugely helps in knowing how code should be or should work even when you can’t see it - which is critical to handle context size challenges - and the lack of which leads to the pieces of larger codebases the AI creates not working together at all, despite that they easily could if consistent coding patterns were used
^ Probably could list dozens more huge ones like these.
All would be trivial for OpenAI to fix… but they don’t
Did everyone but the low-tier H1Bs who never did any real coding on their own check out of OpenAI?
Would fix these myself and turn it into a product, if many people were willing to pay, though guessing that’s not hugely likely
Someday
🄳🄾🄾🄼🄿🄾🅂🅃🄸🄽🄶
💯8😎2