Dzmitry Pletnikau

1.4K posts

Dzmitry Pletnikau banner
Dzmitry Pletnikau

Dzmitry Pletnikau

@spring_stream

Curious about how the world works.

Utah, USA 가입일 Kasım 2007
620 팔로잉293 팔로워
Dzmitry Pletnikau
Dzmitry Pletnikau@spring_stream·
I prefer the hammer to not talk back to me, but I found one instance where it's useful: switched to LLM agent for nutrition tracking and instructed it to praise or roast me depending on relation of the meal to target macros. Monkey brain🤝anthropomorphized linear algebra
English
0
0
1
17
Dzmitry Pletnikau
Dzmitry Pletnikau@spring_stream·
@levelsio It sucks too, but differently. I’m switching back to tmux after two months of zellij
English
0
0
1
164
Dzmitry Pletnikau
Dzmitry Pletnikau@spring_stream·
@voooooogel Marimo has that built in, and more. Was a pleasant step up from Jupyter for me
English
0
0
0
7
thebes
thebes@voooooogel·
ooh found a way to connect claude code to a running jupyter notebook so claude can add cells and view the output live. is a bit janky rn but this is a cool way to collab
English
6
1
101
4.2K
Dzmitry Pletnikau 리트윗함
(((ل()(ل() 'yoav))))👾
first reaction: "because these models know how to simplify, not when to do it" and then: "...someone has to TRAIN them to know this and... oh. what a great way to collect labeled data "
Nan Yu@thenanyu

I see things like /simplify and the existence of code review and bug finding AIs. I have to ask, why do these things exist? Why doesn't the coding agent just naturally do these things? I'm sure there's a good answer. Can someone help me understand?

English
4
2
13
4.4K
Dzmitry Pletnikau
Dzmitry Pletnikau@spring_stream·
@gwern @thkostolansky @binarybits Please point me at the research exploring how larger models become more sample efficient. I am interested in understanding that better, and intersection with ICL, SFT, post-training, and "circuits".
English
0
0
0
13
Timothy B. Lee
Timothy B. Lee@binarybits·
Enjoyed this piece. It emphasizes the adversarial aspect, but this is also a good example of the value of tacit knowledge. To model Priya's behavior you have to interact with her over months and build a model of her behavior. Current LLMs don't have a mechanism to do this.
Timothy B. Lee tweet media
English
4
4
64
7.2K
Dzmitry Pletnikau
Dzmitry Pletnikau@spring_stream·
@Cyber_Spock When I say “magical alignment breakthrough” I mean something for example that forever poisons pretraining pool making models incapable of certain capabilities - very “magical”.
English
0
0
1
4
stromqx
stromqx@stromqx·
@spring_stream Doesn't work even with magical alignment because humans are not aligned. It's only stable in equilibrium without an eye of sauron. The many mordors model or something. Someone will build it so multipolarity is the only real solution
English
1
0
1
11
Dzmitry Pletnikau
Dzmitry Pletnikau@spring_stream·
@ch402 I and many I know take this view as the best model of LLM behavior. Your post implies you have only started taking it seriously recently. I am honestly curious what made you not adopt it in late 2022 when it went mainstream?
English
0
0
1
111
Dzmitry Pletnikau
Dzmitry Pletnikau@spring_stream·
@Cyber_Spock Good point. Non proliferation of certain capabilities requires either coordination or secrecy. Or some magical alignment breakthrough.
English
1
0
1
12
stromqx
stromqx@stromqx·
@spring_stream He just has no ways to solve the game theoretic issues with any of his proposals other than "one world government." And fuck that noise
English
1
0
1
17
Dzmitry Pletnikau
Dzmitry Pletnikau@spring_stream·
I want to roll back to Claude Desktop app version before Claude started coding 100% of it, because current version is slower than my web browser with 400 open tabs
English
1
0
1
35
Dzmitry Pletnikau 리트윗함
François Chollet
François Chollet@fchollet·
Sufficiently advanced agentic coding is essentially machine learning: the engineer sets up the optimization goal as well as some constraints on the search space (the spec and its tests), then an optimization process (coding agents) iterates until the goal is reached. The result is a blackbox model (the generated codebase): an artifact that performs the task, that you deploy without ever inspecting its internal logic, just as we ignore individual weights in a neural network. This implies that all classic issues encountered in ML will soon become problems for agentic coding: overfitting to the spec, Clever Hans shortcuts that don't generalize outside the tests, data leakage, concept drift, etc. I would also ask: what will be the Keras of agentic coding? What will be the optimal set of high-level abstractions that allow humans to steer codebase 'training' with minimal cognitive overhead?
English
171
383
3.3K
320.1K
stromqx
stromqx@stromqx·
I mean it's not a bearer asset like Bitcoin but the payments on lighting or tether just work. Obviously theoretically lacks some of the censorship resistance of on chain payments etc. Straightforwardly I don't use stablecoins frequently but if I did this is probably how I would use them. Tether adoption globally is insanely high
stromqx tweet media
English
1
0
1
44
Dzmitry Pletnikau
Dzmitry Pletnikau@spring_stream·
Planning to lift some friends&family from Venmo to self-custody stablecoin for everyday settlement, but all major wallet-apps now have gambling built-in and prominent on the main screen 😭
English
1
0
1
35