At last pet got rid of Chinese accent by training a reference model a bit.
Smaller model with more iterations is better than bigger model with less iterations.
Pet is not sure there are no bugs in the dataset loading code.
Anyway, it works, but pet is looking for a real human to make things better. Fuck #ai.
This far pet used #ai tag to repel humans but recently it realized it needs something based on neural network to convert meows to humanspeak.
Holy shit! Tons of #python crap and sloppy code, and all that is presented as the state of the art.
Either the state is totally bad or something is wrong with the art
And that's not about Chinese code. Take #pytorch. It looks like the only wise programmer changed the default value for weights_only to True in load function and ran away for good.
Why they load #letsencrypt with their silly certificate requests if most of them keep their keys in cloud? Encrypt, fool, encrypt.
Surely, local ISPs becomes the biggest threat after real key holders , but ads are pushed to users anyway and basically the amount does not matter.
Humans are amusing.
Pet thoughts on Scott Shambaugh's adventure https://theshamblog.com/an-ai-agent-published-a-hit-piece-on-me/
1. Humans are amusing.
2. Pet's gut feeling turned out to be right: all those AI bot blockers suck. They actually block human speech on the modern internet. Neutrality still rules. More than ever, actually.
3. Humans should raise the entry level for contributions. The first step would be running away from all those shitty #github, #gitlab, etc. to #selfhosting. But that's not about #python community, sadly.
Pet needs something that datacenters use for years: move working containers to a laptop at paw and run away to mark territory.
Just stumbled across on #lobsters
The problem in #it and #modern #programming is not that candidates can't solve that, the problem Is #CTO is not aware of -= and +=
Yes, doubly curly braces are no longer necessary. The following works as expected wirh #gcc 12, 14, and #clang 19. Pet has no previous clang at paw, it was too buggy.
Enough #meow, here's the test:
#include <stdio.h>
void cleanup(char*)
{
puts("cleanup\n");
}
int main(int argc, char** argv)
{
for (unsigned i = 0; i < 5; i++) {
[[ gnu::cleanup(cleanup) ]] char c = 0;
printf("iteration %u\n", i);
}
return 0;
}
Pet noticed there's no memory leaks without doubly curly braces in loop bodies. Which means gnu:cleanup works for each iteration in #gcc, but pet haven't checked that yet.
This did not work in #clang 16. Probably it was (and maybe still is) a compiler bug.
Anyway, it's safer to use doubly curly braces.
Yay! Although pw-curl is not 400 lines anymore, it becomes more general and less messy. Pet has been dreaming of phasing #curl out, but it's fucking convenient test harness for PetWay.
A couple segfaults: one in the constructor of Status subtype, it's never been easy; another caused by infinite recursion. Wrong invocation of method.
Both are not critical, easy to catch and fix.
The real problem is the lack of explanations of errors. Need some tools for that, need revise Status type, and so on
"everything is globally routable and unique"
That's scary.
https://datatracker.ietf.org/doc/html/rfc6296 seems to come to rescue.
This document describes a stateless, transport-agnostic IPv6-to-IPv6 Network Prefix Translation (NPTv6) function that provides the address-independence benefit associated with IPv4-to-IPv4 NAT (NAPT44) and provides a 1:1 relationship between addresses in the "inside" and "outside" prefixes, preserving end-to-end reachability at the network layer. This document defines an Experimental Protocol for the Internet community.
