killerstorm

7 ms·

91.
by killerstorm 3mo ago
Well, modern medicine + economy + social pressure resulted in RADICAL change in fitness function for human population. It's very, very different. So it's quite likely that modern population is not fit according to old criteria. &g
92.
by killerstorm 3mo ago
> A computer genetic algorithm run for a billion generations doesn't lead to anything anywhere near the the complexity of a human. What?... Our computers can't simulate anything similar to a real world. You're comparing a
93.
by killerstorm 4mo ago
Yeah, this seems to be more like a concept piece. Just something to talk about, not really useful
94.
by killerstorm 4mo ago
Well, I mean you mixed up "fine-tuning" and "reinforcement learning" a bit when describing these options. Regarding the value of these options, SFT communicates more information to the model being trained, but there&#x27
95.
by killerstorm 4mo ago
Fun fact: no dev containers in Cursor.
96.
by killerstorm 4mo ago
I'm sorry, but you got the terminology exactly backwards. Training on the answer is called supervised fine-tuning. Just for the sake of clarity: 0. Full distillation uses logits of the teacher model - that's much more information
97.
by killerstorm 4mo ago
I'm actually surprised with how bad is Google's voice recognition - with all their immense compute resources, R&D, owning Android, etc, what prevents them from having SotA voice recognition?.. Even "it's not a priori
98.
by killerstorm 4mo ago
I think the situation aetherspawn is "highest rank in a low-rank lobby", e.g. an experienced player playing together with newbies. That might be a person who's just enjoying executing a specific strategy. True top-tier player
99.
by killerstorm 4mo ago
There's not much hostility in Beyond All Reason outside of certain "newbie" lobbies. Occasionally somebody has a conflict, but people who take game seriously avoid offensive words as they don't want to get banned.
100.
by killerstorm 4mo ago
I tried playing SC2 for a few years but controlling units was too difficult: the only thing I could do is to send entire army to fight. No such problem with BAR, at least in early game: quite feasible to control each unit even if you don&#x
101.
by killerstorm 4mo ago
Sorry, but it's just not true. Pretty much the only case I've seen where a person is kicked out is either griefing (deliberately playing bad) or refuse to listen to advice. If you don't know basics there's plenty of oppo
102.
by killerstorm 4mo ago
What's unrealistic about fission? France can do it. If best minds in Europe join forces they can figure how to do it even cheaper/safer/faster. The missing part is political will. If anybody with unreasonable objection will b
103.
by killerstorm 4mo ago
Well, EU is convinced it's a real problem, but the only solution they are considering is to reduce emissions largely via degrowth. OTOH to actually solve the problem we'd need geoengineering and cheap energy e.g. from fission. Jus
104.
by killerstorm 4mo ago
It's not only expensive, it's also wasteful - there's no value in using an obsolete model. Open source AI manifesto demand that "Opensource AI should remain ... economically viable". That's just wishful thinkin
105.
by killerstorm 4mo ago
Perhaps some radical MoE where you download _exactly_ the components you need as you need them. Currently MoE is switched usually on per-token per-layer basis, so you need all weights locally. But e.g. Apple made one which pre-selects all e
106.
by killerstorm 4mo ago
Cars were quite desirable in Soviet Union, where industry was not allowed to advertise. You had to get into a queue to buy a car, the state was not interested to make them in a quantity to satisfy the demand. Very few people actually _neede
107.
by killerstorm 4mo ago
My view on this is based on situation in Ukraine: Ukraine definitely didn't have any car lobby at least until 1990 as Soviet Union was heavily investing into public transportation and did not profit from car sales. Still, general opini
108.
by killerstorm 4mo ago
Perhaps more like nuclear power? Potentially very beneficial but also dangerous. JFYI for-profit companies make pretty OK nuclear power plants
109.
by killerstorm 4mo ago
We just need to make it on blockchain so it is immutable. <ducks>
110.
by killerstorm 4mo ago
I think it's fundamentally not useful as long as there are other open source model releases. E.g. suppose you make SotA model at a particular size via decentralized training. Amazing. In a month Qwen/Deepseek/etc release a ne
111.
by killerstorm 4mo ago
Objectively we care about the amount of information in the signal, not air movement. Air movement is just a medium which conveys information. It's not just "subjective perception", it's the meaning of the process.
112.
by killerstorm 4mo ago
I don't think the pressure of the auto lobby is really the reason. People feel cars are more convenient and more prestigious than riding on a bus. Car lobby certainly accelerated the process, but car users were the main driving force.
113.
by killerstorm 4mo ago
Yeah... I've got downgraded to Opus 4.8 in a purely theoretical discussion of a secure permission model for agent tool calls. So classifier is very broad, indeed
114.
by killerstorm 4mo ago
Hmm, looks like by "hardware hackathons" he really means "making software for more resource-constrained platforms". Nothing wrong with that, but I don't think that's a right name for it. Also adding some solder
115.
by killerstorm 4mo ago
It's silent specifically for ML topics. For biology and cybersec (like, finding bugs) it will show a clear warning.
116.
by killerstorm 4mo ago
Yeah, I'm sure Anthropic loves people switching to Codex. Brilliant marketing.
117.
by killerstorm 4mo ago
I'm sure there are people in Russia, China, ... who don't feel like they're living in an authoritarian country.
118.
by killerstorm 4mo ago
"Malicious use" means spam, propaganda bots, etc. It's nice to give people who work on spam filters some heads-up.
119.
by killerstorm 4mo ago
Presumably from internal all-hands presentation in Google: “Now we must double every 6 months… the next 1000x in 4–5 years.” reported by CNBC in November 2025, attributed to Amin Vahdat, Google Cloud VP / AI infrastructure lead. Sundar
120.
by killerstorm 4mo ago
I'm not sure these things are about "big hit of dopamine". It's more about keeping user's attention on screen. And e.g. tiktok repeatedly shows minimally interesting videos, keeping viewer in expectation: how does t

More ›