FergusArgyll

7 ms·

1.
by FergusArgyll 5d ago
It's not always deceptive. Bugs in inference can be very subtle (well, bugs everywhere can be subtle). The providers should be benchmarking their offerings daily
2.
by FergusArgyll 5d ago
The cheapest kind of journalism. Acme corp does x, Scroll twitter, find malcontents, write an article "People are upset about x". I'm outraged at mathematicians. There, write the article.
3.
by FergusArgyll 5d ago
Huh, call of the void is common and it has a name! Great site
4.
by FergusArgyll 6d ago
There are multiple kinds of distillation https://arxiv.org/abs/2106.03310 https://arxiv.org/abs/2207.12106
5.
by FergusArgyll 6d ago
Way way shallower search. A human doesn't consider more than a few candidate moves. But they (we?) build up much better intuitions and heuristics
6.
by FergusArgyll 6d ago
Distilling you don't have the actual model weights of the teacher. All you have are the teachers answers to a lot of questions. You then teach your own smaller model to answer more similarly to the big teacher model. Fine tuning you ha
7.
by FergusArgyll 12d ago
When I studied cybersecurity I always had this nagging feeling "But I can never get past something real " Turns out I can bypass the defenses of a trillion dollar tech company!
8.
by FergusArgyll 12d ago
Ooh I never realized FLT and Code book were the same author. Yes, both great!
9.
by FergusArgyll 14d ago
If that's not art, I don't know what is. https://scosman.github.io/pelicans_riding_bicycles/
10.
by FergusArgyll 16d ago
eww? :r !curl?
11.
by FergusArgyll 17d ago
This was useful for the 1-2 months where models were good enough to use all their newfound tools but not good enough to vibe code them themselves. Now, just install codex (or claude code or whatever) on a vps, tell it to make a way to email
12.
by FergusArgyll 17d ago
Woah, went from > 1 min on git.ffmpeg (difficulty 6) to instant. Thanks! Edit: after some more analysis; Average with extension disabled is 90 seconds, with extension enabled it's .3 seconds
13.
by FergusArgyll 20d ago
Hermes is close
14.
by FergusArgyll 1mo ago
At least AI is not depressed all the time and trying to make everyone lose their money
15.
by FergusArgyll 1mo ago
I think there's still a ton of secret sauce needed for serving them economically
16.
by FergusArgyll 1mo ago
More RLVR. Give it verifiable problems, if it doesn't find a solution move on, if it does, use that as a reward signal.
17.
by FergusArgyll 1mo ago
They shoot through walls!
18.
by FergusArgyll 1mo ago
Would be nice if we can take the bat's sonar encoder, train a projector and fuse it into our NN
19.
by FergusArgyll 1mo ago
Yeah, so it's a spectrum. In general if you want to feel the power of the latest and greatest, you should just not look at the output, give it full access (not on your computer - a server, or old machine you don't use is best) and
20.
by FergusArgyll 2mo ago
> If we represent the operation of the Random Number Generator (RNG) as a time-series graph, the system can be applied to decrypt the bitcoin seed-phrase. Big if true
21.
by FergusArgyll 2mo ago
If you can't find a difference between Haiku & Opus, something's going wrong
22.
by FergusArgyll 2mo ago
Everyone hates regulations until they realize every regulation actually does something. Everyone hates spending until they realize the money goes towards something. None of this will get solved until a crisis
23.
by FergusArgyll 2mo ago
Subliminal Learning: language models transmit behavioral traits via hidden signals in data https://arxiv.org/html/2507.14805v1 "In our main experiments, a “teacher” model with some trait T (such as liking owls or
24.
Is Mythos good at cyber bec it kept hacking Anthropic during training?(lesswrong.com)
2 points by FergusArgyll 2mo ago|1 comments
25.
by FergusArgyll 2mo ago
I had to butcher the title due to length
26.
by FergusArgyll 2mo ago
**reflects** the cost... Not the **literal** cost
27.
by FergusArgyll 2mo ago
Whatever, doesn't matter. The point is a model should be able to exist and be used even if it goes against whoever got 270 electoral college votes
28.
by FergusArgyll 2mo ago
> In our experience, C has proven to be a pleasant, expressive and versatile language for a wide variety of programs. It is easy to learn, and it wears well as one's experience with it grows. The C Programming Language K&R
29.
Synesthesia(en.wikipedia.org)
2 points by FergusArgyll 2mo ago|0 comments
30.
by FergusArgyll 2mo ago
It looks to me very similar to pi - Suleyman's last venture

More ›