Lovely news to get on the morning of my first day at the UK AI Safety Institute. :)
| web | https://naml.us |
| github | https://github.com/girving |
| location | London |
Geoffrey Irving
@irving
- 482 Followers
- 113 Following
- 465 Posts
Chief Scientist at the UK AI Security Institute (AISI). Previously DeepMind, OpenAI, Google Brain, etc.
Yikes: A backdoor in xz, after 3 years of social engineering and a bunch of sock puppet accounts to gain maintainer access. 😱
This line is particularly interesting. I wonder if oss-fuzz could mitigate this by requiring reports to go to multiple places? Hard to verify independence, though.
…
2023-03-20: Jia Tan updates Google oss-fuzz configuration to send bugs to them.
…
Program to recognize its own checksum:
```
#!/usr/bin/env python3
import sys
import zlib
if len(sys.argv) == 2:
print(('Not me', 'Me.')[zlib.crc32(open(sys.argv[1], 'rb').read()) == 1441602037])
```
Usage: ./self-crc self-crc
(Note: It's supposed to end with a newline.)
@gregeganSF It's even easier if one uses a checksum:
```
#!/usr/bin/env python3
import sys
import zlib
if len(sys.argv) == 2:
print(('Not me', 'Me.')[zlib.crc32(open(sys.argv[1], 'rb').read()) == 1441602037])
```
Usage: ./self-crc self-crc
@[email protected] Purely for curiosity's sake, do you have a guess as to the total number of chicken bits in modern chips, either M-series specifically or generally? Is it <100, 100s, 1000s, etc.?
I have to say I find this situation very strange from a high-level. It doesn't feel like a pattern which will hold for computers and software 1000 years from now.
One of my favorite properties of optimization is that branch-free code is critical to high performance software, but in hardware branches are extremely cheap. So a lot of whole system optimization involves taking something with a horrendous number of cases to branch through (a floating point add, say), pushing it into hardware where the branches are nearly free, then calling it from software in a regular pattern.
χle OrmsbyI love imagining the mathematical needs of whoever decided on the unicode math characters.
"Gosh, I'm just always needing notation for when two lines are sorta-but-not-completely perpendicular. I know! â«¡"
"What if something's not just bigger than or much bigger than but WAY FUCKING BIGGER THAN? Aha: ⫸"
"I'm doing so many specialized contour integrals and hate writing with words so much that I'm going to invent a specific symbol for line integration with rectangular path around a pole: ⨒"
"I don't know whether A is a subset of B or B is a subset of A, but at least one of those statements is true, so... A â«“ B."
"Gosh, I have this element S of the Picard group of a symmetric monoidal category and I want to tensor M with the tensor-inverse of S but don't want to create notation for the tensor inverse. As such, consider M⨸S."
Andrew McNutt
It seems surprising that the best known algorithm for computing π(n) (the number of primes ≤ n) takes time Õ(n^(2/3). I would have naively expected that something Õ(n^(1/2)) is possible, sieving only over numbers k < n^(1/2), but apparently the region [n^(1/2), n] is too messy.