kp1197

4 ms·

1.
P(Kill-Switch|Detection)(lesswrong.com)
1 points by kp1197 3d ago|0 comments
2.
P(Kill-Switch|Detection)(lesswrong.com)
2 points by kp1197 17d ago|0 comments
3.
Prompt Sufficiency: A Missive for the Managerial Class(lesswrong.com)
1 points by kp1197 25d ago|0 comments
4.
by kp1197 8mo ago
Hello! Have you ever felt like too important to answer questions? Guess what! This one weird trick will turn you into a French General in the boardroom!
5.
by kp1197 9mo ago
I kind of regret calling it Straussian because of the baggage that comes with that label. The key differentiator here compared to exoteric/esoteric is the stabilization mechanism: social costs to either upgrading one’s own understandi
6.
Straussian Memes(lesswrong.com)
43 points by kp1197 9mo ago|52 comments
7.
by kp1197 9mo ago
Soon he may be making vanity license plates
8.
Toddler Shoggoth Has Plenty of Raw Material (The Memetic Cocoon Threat Model)(lesswrong.com)
2 points by kp1197 10mo ago|1 comments
9.
by kp1197 11mo ago
Isn't the post more about the interesting capabilities of the language rather than an engineering context?
10.
by kp1197 1y ago
I remember the days of countless insane, boutique javascript libraries. React's ubiquity is a victory - let's not encourage throwing something away in the name of some vague notion of "innovation".
11.
Gradient Descent on Token Input Embeddings(lesswrong.com)
3 points by kp1197 1y ago|1 comments
12.
by kp1197 1y ago
Does performing gradient descent on token input embeddings lead to interpretable results? And if not, why?
13.
by kp1197 1y ago
After working pretty closely with vLLM and SGLang over the past few months, this is EXACTLY what I had envisioned what a successor project would look like - analyzing an operation dependency graph and then fusing (or, at a minimum, scheduli
14.
by kp1197 1y ago
It is unclear to me is whether Kent recovered his cognitive ability or not. He says he has, "thanks to the genie" (which seems to be his name for coding LLMs). Does this mean that with a genie, he is now able to produce like he
15.
by kp1197 1y ago
Robots that can do anything.
16.
by kp1197 2y ago
Is it though? He's perfectly positioned to do so.
17.
by kp1197 2y ago
As I've posted elsewhere in the comments here, I think there are good solutions, but it will take some cooperation to get them off the ground (adoption of standards, licenses, etc.). Here is what I've proposed, which comes prett
18.
by kp1197 2y ago
I've got some good solutions on this front, I think: https://dev.to/kylepena/addressing-the-threat-of-deep-fakes-...
19.
by kp1197 2y ago
Watermarking is not the way to go. It relies on the honesty of the producers, and watermarks can be easily stripped. With images, the way to go is detect authentic images, not fake ones. I've written about this extensively: https
20.
by kp1197 3y ago
I don't buy the argument that merely because solutions exist, they can spontaneously be "chosen" by the system, even if such choosing doesn't break rules. We are talking about an idealized setup using disproven physics
21.
by kp1197 3y ago
The article is asking us to consider an idealized situation
22.
by kp1197 3y ago
Great point on the analogue hack. But I think it's a perfect-is-enemy-of-good situation. There is currently no such thing as digitally verifiable media. If such a thing existed, it would at least partially shove the cat back into t
23.
by kp1197 3y ago
This seems like an obvious solve. Produce cameras with HSM (hardware security modules) that cryptographically sign the image using the existing certificate infrastructure. Get browser vendors to visually indicate signed images. Now you ha