atemerev

20 ms·

421.
by atemerev 2y ago
Not your weights — not your coins!
422.
by atemerev 2y ago
I am not sure we will be able to build something smarter than ourselves, but I sure hope for it. It is becoming increasingly obvious that we as civilization are not that smart, and there are strict limits of what we can achieve with our bio
423.
by atemerev 2y ago
It is $7.5k, and slower than 4090 compute-wise. Two 4090s (which are cheaper than A6000) are strictly superior to it.
424.
by atemerev 2y ago
Oh come on, the entire bill was against open source models, it’s pure business. “AI safety”, at least of the X-risk variety, is a non-issue.
425.
by atemerev 2y ago
“Oh it’s nice but it’s not Rust, why everything can’t be just like Rust” The audacity of Rustaceans is so thick.
426.
by atemerev 2y ago
Which card do you mean that has more cores/computing performance than 4090 and costs 4-5k?
427.
by atemerev 2y ago
Sort of works for inference, but not for anything else. For now, I bought a 4x4090 workstation… but I am crying how power-inefficient it is.
428.
by atemerev 2y ago
Nvidia B100 has 192G VRAM.
429.
by atemerev 2y ago
We want cheap(ish) desktop GPUs with 80G or more VRAM, so not only large corporations could meaningfully participate in AI research. The price divide between “desktop” and “datacenter” GPUs is artificial and no doubt there is a collusion of
430.
by atemerev 2y ago
The real reason the rich support rent controls is that these will not undervalue their existing property.
431.
by atemerev 2y ago
These sorts of optimizations can and should be handled by a (sufficiently smart (tm)) compiler. Common Lisp/SBCL is usually sufficiently smart. I know not everyone likes Common Lisp, but at least I would have tested it with something m
432.
by atemerev 2y ago
Nope. Wheels are the only thing that makes Python/PyPI usable. I don’t want to wait tens of minutes to recompile pytorch, or something (and conda is way too heavyweight for my tastes)
433.
by atemerev 2y ago
Nope. System packages and language dependencies are better not to be mixed, and the horrors if you do that are very real.
434.
by atemerev 2y ago
We do, but most of software engineers out there hate it and use the dot, for obvious reasons (CSV in particular). Having said that, I like our Swiss decimal group delimiter: 1'000'000.
435.
by atemerev 2y ago
There is nothing special about being human. Homo sapiens is accidental, a product of trial and error, like any other species on Earth, and the only significant difference between us and other animals is intelligence. There were even other s
436.
by atemerev 2y ago
Well, this evolutionary assumption is a heuristic, to be constantly verified and doubted by the mind. You can have an abusive family, and you need to escape it the same way you would escape any other abuser. You can have a brilliant human b
437.
by atemerev 2y ago
Do you prefer a dumb citizen of your own country to a genius foreigner? I do not. I don’t care about the form. I care about intelligence.
438.
by atemerev 2y ago
Curiosity. Intellectual pursuits. Pride of creating something better. Why would humans be interested in raising children smarter than their parents?
439.
by atemerev 2y ago
I have seen instances of IDF admitting their mistakes, in this war as well. Like, the hand grenade at the mosque incident. I haven’t seen anything like this from the other side. For them, every Jew is the enemy.
440.
by atemerev 2y ago
Humans are not special. Intelligence is. If we create something more intelligent than us, this is the way it should be. Maybe they will build their civilization better than we did — and we did many things that weren’t smart and beautiful, t
441.
by atemerev 2y ago
A tree interacts with many things, not just human observers. We are not special in any way. This discreteness is only relevant and/or observable when we look at individual particles that rarely interact with anything else.
442.
by atemerev 2y ago
I know, and yet early model implementations were quite unoptimized compared to the modern ones.
443.
by atemerev 2y ago
A particle is a node in the universal interaction graph. “Space”, however, is a derivative attribute emerging as the “distance” between particles in the graph; there is no “space”, only metrics. Reality does not exist between measurements&#
444.
by atemerev 2y ago
405B is beyond homelab-scale. I recently obtained a 4x4090 rig, and I am comfortable running 70B and occasionally 128B-class models. For 405B, you need 8xH100 or better. A single H100 costs around $40k.
445.
by atemerev 2y ago
IDK, 8B-class quantized models run pretty fast on commodity laptops, with CPU-only inference. Thanks to the people who figured out quantization and reimplemented everything in C++, instead of academic-grade Python.
446.
by atemerev 2y ago
Sure; however, the energy density is too low and 1GW power plant takes much more space than a compact nuclear power plant. Solar panels also degrade and need replacement. But yes, sure, we need solar too.
447.
by atemerev 2y ago
EmailJS linked to my gmail account. So far it is enough. When it won’t be enough, I’ll use Mailgun.
448.
by atemerev 2y ago
Well, if I have to choose whether I believe IDF or Hamas, I obviously choose IDF.
449.
by atemerev 2y ago
Depends on which side you are on.
450.
by atemerev 2y ago
Preparation. Surprise attacks work… once. Then, you have to wait for a few years when the enemy side lowers their readiness again.

More ›