sigbottle

3 ms·

1.
by sigbottle 5d ago
> I think people are still not used to non deterministic tools like this, and human perception is absolutely horrible at evaluating trends like this no matter how smart, clever, and experienced you are. The implication is that humans are
2.
by sigbottle 6d ago
Could someone explain to me what the general workflow is now that people are converging to? I haven't really been catching up with the AI ecosystem but I was looking into agent sandboxes and VM's recently and there's a ton
3.
by sigbottle 7d ago
Furthermore, it's not about the current innovation right now - if you sell yourself on a broader mission, your core product can evolve and change with it, and you're more selling yourself as the guy who will make that abstract vis
4.
by sigbottle 7d ago
Well, it's knowing when to push and when to not. You probably have an intuition for, I don't know, abstract algebra objects (I don't know your field of specialty :P), without needing to symbolically manipulate all of it, but
5.
by sigbottle 8d ago
__asm__ __volatile("hlt"); when doing quick and hacky debugging could work
6.
by sigbottle 8d ago
I really want to create a nosology of common generic memes that can be applied to literally anything without context. Saying that the evaluators are just stupid and arbitrarily chasing the fashion of the week instead of the evaluators possi
7.
by sigbottle 8d ago
I really wish there were search harnesses , actually. My LLMs are lazy as hell and seem to want to just report the first thing they find on google. I know they can return truly niche and useful results, but it takes a lot more prompting
8.
by sigbottle 9d ago
aww man. I remember following victor in college. I mean pivots gotta pivot, and this is probably a better one for business, but always thought the interaction combinator framework was cool
9.
by sigbottle 9d ago
Well there are counting arguments to say that a higher level intelligence that somehow achieves say, 100000× brain efficiency of humans, still can't do that much more work than humans, if humans found the "best abstraction"
10.
by sigbottle 10d ago
I do know what a monoid is, but a monad in the category of endofunctors is the scary word for me :sob:
11.
by sigbottle 10d ago
Sorry I changed problems a bit and started talking about me trying to understand matrices lol
12.
by sigbottle 10d ago
It always messes with me: reducing across a specific axis always takes O(whole tensor) time, because there's no difference between "iterate over all dims, then collapse the final one" versus "iterate versus the first dim
13.
by sigbottle 10d ago
Isn't reduce usually used for monoidal operations? Or do people implicitly absue ordering? If the algortihm doesn't work the same forward, backwards, and with a tree scan, it ain't reduce (as a first approximation not IFF)
14.
by sigbottle 10d ago
Honestly part of me feels that way about way too many things in retrospect about my own life and interests.
15.
by sigbottle 11d ago
Upskilling in both linux kernel dev work and also LLM inference systems work. It's kind of hard to just "get into" these though, as they're sufficiently foreign that I'm spending more learning about the "accide
16.
by sigbottle 12d ago
Compression in this modern day and age is so slop. Yes, I'm familiar with keystone results such as Solomonoff induction. It's a direct counterexample to compression - your intensional algorithm can completely outrun reality. I can
17.
by sigbottle 12d ago
Again, I will posit the hypothesis that it's a learned behavior from training. Distinctions , you generally "only pay for" in computational cost, by needing to search twice over an axis you may not need to split. Similariti
18.
by sigbottle 12d ago
It's mixed for me because there are certain things that I clearly think are needed. For example I'm building a custom network architecture and it's to the point where I'm using frontier models to reverse engineer game cl
19.
by sigbottle 12d ago
What's the best way to get into any kind of frontier training or inference? Obviously OpenAI and Anthropic would be great, but I would love to work at inference literally anywhere. LLMs are super cool and I want more systems problems.
20.
by sigbottle 12d ago
Well, we can generalize this a bit to model routing in general. Actually, just AI in general - even when directly interacting with GPT we haven't figured out good protocols yet.
21.
by sigbottle 13d ago
To a first approximation, everything is power. How do we prove alignment? Often times, large regulatory bodies. Who supports that? Does the regulatory body actually have enough power to enforce those standards ? (Hint: look at modern Ameri
22.
by sigbottle 15d ago
For me, I've honestly never been much of an engineering builder - I'm focused on learning for myself. Which is useful - AI has accelerated that a ton - but it's still bottlenecked by, unfortunately, me. I should probably look
23.
by sigbottle 15d ago
May the iterative loop of adding new axes to evaluate on be a natural, healthy progression, instead of needing to frame it as an us-them problem? If you value humans intrinsically, this is necessarily the loop that will converge. I don&#x27
24.
by sigbottle 16d ago
I find the discourse under this blog fascinating. > The big, old ideas about intelligence that ended up basically vindicated were the ideas about how intelligence is about prediction, and prediction is about compression, and compression
25.
by sigbottle 16d ago
In general, a lot of moral invariants that natural selection has rendered as "intuitive" to us are no longer intuitive or possible. These natural brakes are not braking.
26.
by sigbottle 17d ago
At the end of the day, DSP is a lot of math and a lot of hardware. A lot of GNU Radio's code is a monster unification of so many different choices into certain interfaces (or duplication of the same choices) for, presumably performance
27.
by sigbottle 17d ago
Nit: is it any computable function? I thought the requirements were unbounded (in principle) memory and time. (For all intents and purposes given how high dimensional you are and using the "vibes" of computability yes I agree w&#x
28.
by sigbottle 17d ago
Chivalry and honor. Fight them to a duel to the death on horseback and javelin jousting.
29.
by sigbottle 18d ago
> "we demonstrate that LLMs can spontaneously develop novel social biases about artificial demographic groups even when no inherent differences exist" For a while (It's getting better with Astra, but still there), a lot of
30.
by sigbottle 18d ago
Well, according to Terry Tao, there were recent developments (from weeks ago) that made Navier Stokes in principle, solvable. So ignoring time, I say possibly, just because the groundwork was laid. What's impressive is parallelizing it

More ›