moyix

10 ms·

61.
CTF Writeup: Abusing select() to factor RSA(threadreaderapp.com)
135 points by moyix 3y ago|26 comments
62.
by moyix 3y ago
From what I can see this captures screenshots? A nice feature of asciinema is that you can select text from the recordings during playback or programmatically extract it from the log files.
63.
by moyix 3y ago
I recently started using this for recording all my terminal sessions (without upload). It's pretty neat to be able to go back and replay any session I had. I also threw together a small tool to grep through the output, although it&#x27
64.
by moyix 3y ago
IIRC the name refers more to cleaning staff at a hotel, who would have access to your laptop while you’re out of the room.
65.
by moyix 3y ago
Yes. Inpainting lets you erase part of an image and have the AI model fill it in based on your prompt.
66.
by moyix 3y ago
Thanks, I missed that! It looks like the previous discussion didn't touch on the Cavium news, though.
67.
by moyix 3y ago
ARM just licenses the ISA and provides some reference designs. Individual manufacturers can (and often do) add their own extensions and design the actual chips.
68.
Snowden leak: Cavium networking hardware may contain NSA backdoor(twitter.com)
1205 points by moyix 3y ago|591 comments
69.
by moyix 3y ago
You'd be surprised. In dense cities like NYC/Brooklyn fiber availability varies from block to block, I assume because adding new infra is so expensive. We recently moved about five blocks away and lost access to fiber :(
70.
by moyix 3y ago
I believe there are somewhat standard tools for scanning memory dumps for cryptographic material, which have been around since the cold boot attack era. And I can imagine attackers opportunistically looking for crash dumps with that in mind
71.
by moyix 3y ago
I was under the impression that most modern (past few years) SoCs like Exynos, Qualcomm, Apple silicon all had IOMMU support. Sometimes it’s misconfigured to be too permissive but that’s getting better. Qualcomm SMMU: https://www
72.
by moyix 3y ago
I think you're misremembering here; as far as I know (and as far as I can tell from searching just now) MS has never sued ReactOS. There was a claim made back in 2006 on the mailing list that a portion of syscall.S was copied, and this
73.
by moyix 3y ago
The lowest temperature isn't always the one that results in working code! This was shown in the original Codex paper: > When evaluating pass@k, it is important to optimize sampling temperature for the particular value of k. In Figur
74.
by moyix 3y ago
I'm sure you're right but the only note attached to the author list is in the opposite direction – Tom B Brown has an asterisk with "Work done while at OpenAI".
75.
by moyix 3y ago
Seems marginal – half of those authors, including the lead author, are from OpenAI!
76.
by moyix 3y ago
RLHF is from Google? The reference I know of is OpenAI: https://arxiv.org/pdf/2009.01325.pdf CLIP also seems novel?
77.
by moyix 3y ago
It is voluntary. The convention is used in BIG-bench [1] and mentioned in their README here: > All task files (including README.md and task.json) contain a "canary" string, which should not be edited. This is to prevent benchm
78.
by moyix 3y ago
I genuinely don't see how this viewpoint is compatible with the grokking results.
79.
by moyix 3y ago
From what I remember they check in at a few intervals after the suggestion is made and use string matching to check how much of the Copilot-written code remains.
80.
by moyix 3y ago
The number is 40%, and it's 40% of code written by Copilot users . It's also just for Python: > In files where it’s enabled, nearly 40% of code is being written by GitHub Copilot in popular coding languages, like Python—and we
81.
by moyix 4y ago
We found that Codex was pretty bad at Verilog and so we fine-tuned our own models on Verilog code, starting from the SalesForce CodeGen model. The paper will be at DATE: https://arxiv.org/abs/2212.11140 The largest mod
82.
by moyix 4y ago
Not an informed/sota practitioner, but isn't this just a standard property of high dimensional spaces? https://en.wikipedia.org/wiki/Random_projection > The core idea behind random projection is given in t
83.
by moyix 4y ago
This seems like a silly objection to me, although I am one of the authors :) Yes, Codex has been trained on lots of linked lists. It has also been trained on lots of bad and insecure versions of linked lists! And when we did our earlier
84.
by moyix 4y ago
There’s a group of folks on 4chan doing this on gaming class hardware (4080s etc). They have a doc here: https://rentry.org/llama-tard-v2
85.
by moyix 4y ago
Maybe they meant to link to something like this: https://github.com/ggerganov/whisper.cpp/pull/540 Web demo: https://whisper.ggerganov.com/
86.
by moyix 4y ago
Even if we found a solution to inference-time adversarial attacks tomorrow, backdoor attacks would still be possible, which makes them pretty different IMO.
87.
by moyix 4y ago
Adversarial attacks are inference-time, backdoors are training time. This paper isn't the first to propose the idea of backdooring DNNs (I believe our paper [1], concurrently with a couple others [2,3], did that). But it makes a big st
88.
by moyix 4y ago
Here’s a more directly comparable example: https://arxiv.org/abs/2210.11610
89.
by moyix 4y ago
Many of these models are already being trained in FP16, and FP8 seems likely now that the H100s support it.
90.
by moyix 4y ago
That thread makes a bunch of assumptions that seem a bit dubious to me. We've known since Chinchilla that you don't need 175B parameters to get GPT-3 quality – a 70B model can outperform GPT3 [1]. And his numbers assume the model

More ›