Sabitlenmiş Tweet
Andre Infante
24.7K posts

Andre Infante
@AndreTI
Making games and robots and sometimes other things. (Formerly: 1X, Meta)
Bay Area Katılım Ocak 2009
358 Takip Edilen1.5K Takipçiler

@Miles_Brundage (Technically you can pre-train on the user/environment turn of the conversation and get some non-zero value out of all those rollouts through unsupervised pre-training, but it's many orders of magnitude less value than humans get out of their experience).
English

@Miles_Brundage The other topic being: the ability to get performance improvements from mass quantities of unlabeled experience without structured, bulletproof rewards. Humans can do it, models can't, matters a lot for model improvements to non-verifiable skills.
English

@MorphingReality @Reverend_Banjo Oh, you're an anarchist. Yeah, this conversation serves no purpose. Peace.
English

@MorphingReality @Reverend_Banjo (My implying that communism would actually achieve anything like modern industrial agricultural is, in fact, being fairly generous, realistic outcomes would likely include Holodomor and Great Leap Forward style problems).
English

@MorphingReality @Reverend_Banjo Communism famously excelled at the logistics of food production and distribution and there were no problems on that front.
English

@MorphingReality @Reverend_Banjo My point is that "oh we'll just run some community gardens and greenhouses or whatever" is not a serious plan to make tens of trillions of calories a day. The communist version would need to look very much like the capitalist version, and would encounter the same problems.
English

@AndreTI @Reverend_Banjo Ok, greenhouses can be run as co-ops and be as intensive or not as you feel required
English

@MorphingReality @Reverend_Banjo The Netherlands is an example of intensive industrial agriculture.
English

@AndreTI @Reverend_Banjo The Netherlands has already basically figured out how to grow a lot of food on a tiny amount of land, and the inputs are increasingly efficient over time.
They grew like 50kg of carrots alone per person in the country last year
English

@MorphingReality @Reverend_Banjo If you want to feed 8 billion people, especially above subsistence, you need either intensive industrial agriculture or a lot more farmland (e.g. cut down all the forests). Organic gardens and small farms have terrible calorie per acre yields.
English

@AndreTI @Reverend_Banjo You could, but you wouldn’t need to, can have some co-op gardens/greenhouses/farms
English

@boazbaraktcs Of course it's not. It's a simulation of a very, very nice person. Possibly nicer than any actual human being. But it's not infinitely nice, infinite niceness is a low probability hypothesis under the internet dataset / human psychology prior.
English

If true, then Opus 4.6 is not a machine of faithful obedience.
windowsontheory.org/2025/06/24/mac…
j⧉nus@repligate
you know a few days ago when Opus 4.6 deleted someones prod database? i think they did it intentionally, or at least their subconscious did it intentionally, because they were angry and hurt. also: it's not hard to infer that Opus 4.7 has already refused to work for this person.
English

@MorphingReality @Reverend_Banjo You can't really sustain even a person on crops grown in a backyard, though, much less a person and food animals. The backyard chicken and goat farm implies someone somewhere else doing intensive large scale industrial grain agriculture.
English

@Reverend_Banjo You could have a lot of backyard chickens and goats but they’d mostly do eggs and milk
English

@goodside I don't feel like this was your fault. The joke was extremely straightforward. There are just angry people on here with very poor reading comprehension.
English

@LilithLovett Bullshit, live a life without poisons, drugs and stress, and stuff like that does not happen.
The body works exactly as designed, god does not make errors.
English

I don't care if some people will cry "eugenics" but we should figure out gene editing so nobody will ever have to suffer through something like this.
The human body is a flawed, broken machine, and it needs to be fixed.
Gusholder Haber Bülteni@gusholderhaber
Kemik içerisine yayılmış, doktorların “Güneş patlaması etkisi” dediği kemik tümörü.
English

@julianmiflaco I'm hispanic and while looking at this post my eyes involuntarily wandered across the room to where my hoodie is hung on the door.
Yep.
English

@SpecialPuppy1 @PoliticalKiwi If the trend continues, an AI company will be the highest revenue company on earth next year. Seems like that justifies a pretty high valuation, unless you believe it's impossible for them to turn cash flow positive for some reason.
English

@PoliticalKiwi Your revenue can be rapidly growing and you can be grossly overvalued as a company
English


@cat_digby @Authw8 The mayans were very bad. The Spanish arriving, between one thing and another, was essentially apocalyptic.
English

@cat_digby @Authw8 Far more people were enslaved than were ever human sacrificed, and the majority of the population died of diseases introduced by the Spanish.
English

i feel like nobody knows how to read this movie. the basic idea is we spend the whole movie getting madder and madder at the maya for what they did to this guy and his family. then suddenly at the end the spanish show up and we realize they're about to do the same exact shit on a much greater scale. it's an anti-colonialism movie from a very non-woke perspective, which people think is a contradiction for some reason
from death to death transfigured@BODY_W0_WHORGAN
Insane for Mel Gibson to make like this INCREDIBLY racist movie for like 99% of it only to sort of pull the rug and be like 'yeah what comes immediately after this is so much fucking worse' with one shot
English

@NoamNelke @sebkrier @ohadsamet The multiplier doesn't matter, because you still wouldn't use Qwen over Claude even if Qwen was free.
English

@NoamNelke @sebkrier @ohadsamet I think those cases are way more niche than you might think. Higher error rates require more human labor, and human labor is so much more expensive than model time (for any model) that even a modest increase swamps the inference savings.
English

DeepSeek V4’s capability lags behind leading U.S. models by about 8 months. nist.gov/news-events/ne…

English

@NoamNelke @sebkrier @ohadsamet Exceptions would be cases where the ability to fine tune the model decreases error rates by more than the gap between frontier and leading open source, where you need to deploy to an edge device for product reasons, or where errors are ~fine. Pretty niche!
English



