tani
banner
ewb.io
tani
@ewb.io
coffee son
compression can be kind of search too or visa versa . i think just everything is mappable to k-sat .
July 7, 2025 at 11:37 AM
cool! how much longer would it take to shrink the error bars 10x?
July 7, 2025 at 7:13 AM
i get your point but my priors are still tilted toward this being an impossible exaggeration. lets see the data like you said!
July 5, 2025 at 7:46 PM
i could buy the denominator being bigger like you say but still i think it would have to be the majority goes well undiagnosed for this to be possible
July 5, 2025 at 6:32 PM
is disease diagnosis <25% accurate in general? how is 4x possible??
July 5, 2025 at 2:55 PM
my guess is that this is a win-win type situation, so while good I’m not bought into their moral imperative.
July 2, 2025 at 5:42 PM
sometimes you have to look really closely to find their detail orientation
July 2, 2025 at 2:15 AM
wonder how yann is feeling about all this
July 1, 2025 at 7:43 AM
it’s dogfooding and makes some sense for that reason. it’s not great all round now, though distinctly useful on some occasions, and has the potential to become better with increased use and internal adoption. they are also building the tools they use.
June 29, 2025 at 9:12 PM
they’re going to need a relevant example to understand. c suite usually not great with abstraction.
June 24, 2025 at 7:50 PM
most consistent account on this site
June 19, 2025 at 9:51 PM
openai.com/index/ai-and...

companies pay for the energy they use so they have a direct incentive to minimize it
AI and efficiency
We’re releasing an analysis showing that since 2012 the amount of compute needed to train a neural net to the same performance on ImageNet classification has been decreasing by a factor of 2 every 16 ...
openai.com
June 19, 2025 at 11:07 AM
incentives are actually aligned since smaller models run faster and users of these models care more and more about latency. these companies are putting a ton of effort into making them more compute efficient which has the benefit of making them more energy efficient.
June 19, 2025 at 10:15 AM
woof
June 14, 2025 at 3:28 AM
bots don’t get jokes
June 12, 2025 at 4:30 PM
one guy forcing the other get in on the other side of the van is some wet bandits shit
June 11, 2025 at 5:50 AM
definitely would be a trigger for godzilla
June 7, 2025 at 12:05 AM
yes of course. the issue is that i engage with the content that upsets me. its human nature, like watching a train wreck. even if someone has no explicit interest in divisive content, when they see it, they engage. a better algorithm would differentiate positive and negative engagement.
June 5, 2025 at 6:59 PM
i get your point but using solely engagement metrics, it would be hard not to promote political content even in a system with plenty of other content. discover feed but without politics sounds appealing to me
June 5, 2025 at 6:41 PM
1 and 2 make sense and echo my experience. a lot of the communities that are presented to me in the discover feed are hostile to outsiders that may not pass their purity test. that’s their prerogative but without obvious alternatives it’s hard to engage safely.
June 5, 2025 at 6:23 PM
say it a third way!! jk jk makes sense

1 definitely suffers from the chicken and egg problem. i’ve seen some celebs join and then stop posting because of the engagement. jimmy fallen famously was dogpiled his first day and quit.

2 could be interesting, any ideas there?
June 5, 2025 at 6:18 PM
what are the leading theories for why people are not sticking?
June 5, 2025 at 6:04 PM
seems to me he just generated this…
June 5, 2025 at 3:25 PM
i think it’s possible rag never was a thing because applying a term to selectively putting stuff in the prompt was silly in the first place
June 4, 2025 at 4:20 AM