a-dub

16 ms·

301.
by a-dub 2y ago
would be interesting if they had adversarial/null llms attempting the noisy nlp reductions as well. then one could make arguments about the sturdiness of the noisy bit.
302.
by a-dub 2y ago
KDE 6.3 is pretty great these days.
303.
by a-dub 2y ago
to return to specifics, this thread is about the NIH. cutting indirects overnight to 15% with zero warning is vandalism. the united states was a world leader (this means it produced and attracted the best) in this area, hopefully it will re
304.
by a-dub 2y ago
if you have a growth mindset, it certainly is. they're making the united states a less attractive place for the tippy-top of stem research on a world class level. serious people have choices, they can and will choose elsewhere if thing
305.
by a-dub 2y ago
encourage all the laid off federal employees to move to the three special election counties, flip the house next month, use the house majority to push back on the most absurd/draconian moves, push harder at the midterms and try again t
306.
by a-dub 2y ago
if they were honestly trying to improve things, the cuts would be phased in over a few funding cycles at a minimum. it's hard to see surprise cuts as anything more than willful vandalism and a gift to global competitors. time to donate
307.
by a-dub 2y ago
ahh yes, coherent. where i cut my teeth on unix style systems. it supported memory protection on 286 cpus (true segmentation!) but had no ip stack (uucp was supported though). no x or graphics. pretty good documentation in a big doorstop t
308.
by a-dub 2y ago
this is VERY cool!
309.
by a-dub 2y ago
hm. nothing that happens in the fourier domain can touch the phases, which seems like a constraint that would change the behavior of other layers. the default bias of -0.1 with relus and what i would expect to be a flattish spectrum also se
310.
by a-dub 2y ago
i wonder if it would be possible to insert canaries into the negative examples or otherwise detect when the inference trajectory enters the forbidden zone. otherwise fine tuning seems kinda setuid 0-ish. seems hard to harden it against adve
311.
by a-dub 2y ago
1024 capacity, 1 in 1M false positive rate. (false positives fail safe- sending something the client already has), 3.6KB https://hur.st/bloomfilter/?n=1024&p=1.0E-6&m=&k=
312.
by a-dub 2y ago
they did. it was called the moosehead. it replaced the indy and i think was the last of the mips boxen before they became a pc integrator.
313.
by a-dub 2y ago
a content addressed block store with pointers and skiplists for file continuity would be kinda neat.
314.
by a-dub 2y ago
hmm, maybe the client could include a same origin cache state bloom filter in the request. although i suppose it's a solved problem these days.
315.
by a-dub 2y ago
i think it probably varies from workload to workload. reducing handshake time and header compression can have substantial effects. it's a shame server side hunting/push never caught on. that was always one of the more interesting
316.
by a-dub 2y ago
probably more like number of job posts that include at least one programming language for conditioning. (if you want to look at relative trends). distribution over job class as well as an accurate count of actual jobs per post here also pro
317.
by a-dub 2y ago
when i was a child i was tested in this manner. the test was called the "woodcock johnson iv" (har) i remember a big chunk of the test was repeating back increasingly long auditory digit strings in reverse and drawing increasingly
318.
by a-dub 2y ago
that's how you know when you actually have agi, when you have something that you don't have to shovel in every written word known to man to make it work, but rather can seed it with a few dense public domain knowledge compendia an
319.
by a-dub 2y ago
this is correct. "open source" means everything required to recreate from scratch and improve. not "here's a massive binary, an interpreter script and permission."
320.
by a-dub 2y ago
would have been interesting if they had constructed a temporal loading dose response curve which could support a prophylactic protocol for emergency or otherwise high risk hospital admissions.
321.
by a-dub 2y ago
i wonder if we'll ever see good user owned and operated personal "ai agents" or if the technology will just skip the cottage-hipster phase and jump straight to enshitification.
322.
by a-dub 2y ago
that's right with mechanical pencils written on a proper writing surface with plenty of arm/elbow space with real erasers for completely removing noise from mistakes.
323.
by a-dub 2y ago
the rust library exposes a handful of "non_truncating_*" functions that enable error handling. i would expect this to be for drop-in compatibility with old code. amusingly, the python "library" is just a thin wrapper aro
324.
by a-dub 2y ago
lots of relevant interesting reading on the webkit wikipedia page: https://en.wikipedia.org/wiki/WebKit seems it was a quite complicated matter that included clashes between open source and commercial concerns. https:
325.
by a-dub 2y ago
what ever became of khtml? i remember back in the day it had the nicest codebase of all of the browser layout engines.
326.
by a-dub 2y ago
i run nvidia/intel/optimus/wayland in optimus/integrated mode on my laptop. i have anecdotally noticed perceptually salient increased response times when: 1) dpi scaling is in use 2) external displays are used via usb-c
327.
by a-dub 2y ago
fun fact: sqlite is only about 5.5 months younger than windows 2000...
328.
by a-dub 2y ago
i don't really care how many people i respect liked it, i have to be honest, i hated strang's "linear algebra and its applications." there's a strang text on computational science that was much more my speed (less o
329.
by a-dub 2y ago
my st had a 360kb single sided (double density) disk drive. forgot about the "if you don't format with the pc, the pc won't be able to read it" quirk. takes me back!
330.
by a-dub 2y ago
and apparently european 940s did come in a manual five speed, despite the fact that manual was not offered on the 940 in the us. i stand corrected. live and learn.

More ›