cjbprime

13 ms·

151.
by cjbprime 3y ago
Consumer Ryzens only have two-channel memory controllers. Two dual-rank (double sided) DIMMs per channel, which you would need to use to get enough RAM for LLMs, drops the memory bandwidth dramatically -- almost all the way back down to DDR
152.
by cjbprime 3y ago
Wouldn't expect that to work at all.
153.
by cjbprime 3y ago
You are severely underestimating this attacker and their sponsors. Amateur hacker rings do not spend two years actually diligently maintaining the software they will later backdoor. It would not make any sense in the world of ransomware
154.
by cjbprime 3y ago
I'm surprised it wasn't "authentication bypass" too.
155.
by cjbprime 3y ago
Doesn't your data prove the opposite point? There are so many vulnerabilities and so few people looking for them that even the thirty year old ones have barely been found. A healthy feedback loop would have trended the average age of e
156.
by cjbprime 3y ago
Going to wait for longer context local models. Fine tuning/training is lossy compression of your notes into the model weights -- there isn't much value in a vaguely-remembered copy of some of my notes. This is why other comments a
157.
by cjbprime 3y ago
That was true for this project, which was almost orphaned to begin with. We'll run out of nearly-unmaintained critical infrastructure projects sometime. Larger projects with healthier maintenance situations are also at risk, and it&#
158.
by cjbprime 3y ago
Looks like the slowdown is actually at sshd process startup time, not authentication time. So it's back to being completely impossible to network-probe for.
159.
by cjbprime 3y ago
Probing for vulnerable deployments over the network (without the attacker's private key) seems impossible, not non-trivial. The best one could do is more micro-benchmarking, but for an arbitrary Internet host you aren't going to k
160.
by cjbprime 3y ago
But my argument isn't that freedom of speech could be used as an excuse for something that would otherwise be illegal -- my argument is that publishing and discussing exploit code is a constitutionally-protected activity . The CFAA s
161.
by cjbprime 3y ago
Spending two years actually maintaining an open source project that you will later backdoor is a very expensive way to perform such research.
162.
by cjbprime 3y ago
It looks like the exploit path calls system() on attacker supplied input, if the check passes. I don't think we need to go into more detail than "it does whatever the attacker wants to on your computer, as root".
163.
by cjbprime 3y ago
Even if you mandated two code reviewers per merge, the attacker can just have three fake personas backed by the same single human and use them to author and approve malware.
164.
by cjbprime 3y ago
On the other hand, this was a two year long con..
165.
by cjbprime 3y ago
But do you need to also possess the trusted privkey? I wonder why the attacker would build an unauthenticated kill switch.
166.
by cjbprime 3y ago
OpenSSL's "notification of an upcoming critical release" is public, not private. You do get to know that the vulnerability exists quickly, and you could choose to stop using OpenSSL altogether (among other mitigations) once t
167.
by cjbprime 3y ago
No, freedom of speech (as far as I know) protects even exploit code. The statutes being linked would cover using the backdoor to gain unauthorized entry to a system. I think the question of whether anything illegal has occurred from the p
168.
by cjbprime 3y ago
I think you are confusing non-state e.g. ransomware groups, which are usually not part of a government (although some exceptions like North Korea likely exist) with state-sponsored hackers who are often directly working under military comma
169.
by cjbprime 3y ago
In this particular case, there is a strong reason to expect exploitation in the wild to already be occurring (because it's an intentional backdoor) and this would change the risk calculus around disclosure timelines. But in the general
170.
by cjbprime 3y ago
Could combine the existing real-time driving model with influence from the LLM, as an improvement to allow understanding of unusual situations, or as a cross-check? I wasn't intending to say it would be useful today, but pushing back a
171.
by cjbprime 3y ago
Quantized models could absolutely be run in-car. They can already be run on cell phones.
172.
by cjbprime 3y ago
> It's just not true that everything not transformer is outdated. Am not expert, do you have some links about this? i.e. a neural net construction that outperforms a transformer model of the same size.
173.
by cjbprime 3y ago
In-person yoga classes have turned out to be dramatically therapeutic for me, which surprised me because I was already seeing a therapist, and I thought I didn't enjoy exercise at all. (If ClassPass is in your area, they have a free tr
174.
by cjbprime 3y ago
You appear to be confusing RSUs with stock options.
175.
by cjbprime 3y ago
Few days max.
176.
by cjbprime 3y ago
Does LLM inference require significant bandwidth to the card? You have to get the model into VRAM, but that's a fixed startup cost, not a per-output-token cost.
177.
by cjbprime 3y ago
This will likely build a version without GPU acceleration, I think?
178.
by cjbprime 3y ago
Maybe they wanted to wait for bug reports for 21 days before publishing a popular blog post about it..?
179.
by cjbprime 3y ago
This probably isn't necessary. Anyone past beginner is still sometimes playing one of the best moves for each move, and when they don't it is probably not bad in a predictable way.
180.
by cjbprime 3y ago
No, it just saves more space when the players are good. (Note the engine assigns each legal move a probability.)

More ›