s314

8 ms·

1.
Ovid: A pi extension that makes it record proof its features actually work(github.com)
2 points by s314 3mo ago|1 comments
2.
by s314 3mo ago
I built https://github.com/Srinivasa314/ovid , a pi extension that makes it verify the features it builds and record terminal+browser videos onto the PR. The verifications are ordinary code, so re-running them is cheap
3.
Porting Zircon (Fuchsia OS's kernel) C++ code to Rust(cs.opensource.google)
4 points by s314 3mo ago|2 comments
4.
by s314 4mo ago
Using a logit lens (prior art: https://www.lesswrong.com/posts/AcKRB8wDpdaN6v6ru/interpreti... )
5.
by s314 4mo ago
You can actually de-censor an LLM without understanding how it works from a mechanistic perspective. (See R1 1776) So I don't think there'll be effort to "obfuscate"
6.
by s314 4mo ago
It wasn't "a prompt" but several prompts that transformed the raw experimental results to a blog. > hallucinated from the LLMs world knowledge This can't be true because I checked whether the content was consistent wi
7.
by s314 4mo ago
> Steering seems like a circumventable kludge compared to adjusting the training data directly Correct. Steering is used in mechanistic interpretability studies to prove that your model is correct. There are other better ways to "de
8.
What political censorship looks like inside an LLM's weights (Qwen 3.5)(vas-blog.pages.dev)
82 points by s314 4mo ago|44 comments
9.
by s314 4mo ago
Markdown feels too minimal and HTML feels too complex. I think we need something in between
10.
by s314 5mo ago
Expanding this analogy: LLMs like Claude are like V8, and agent harnesses like Claude Code are like Node.js.
11.
by s314 5mo ago
You can think of the LLM like an interpreter.
12.
A text adventure game, all in Markdown(vas-blog.pages.dev)
1 points by s314 5mo ago|3 comments
13.
Crates.io: Malicious crates faster_log and async_println(blog.rust-lang.org)
2 points by s314 1y ago|0 comments