cjbprime

11 ms·

121.
by cjbprime 2y ago
Specifically, it's not mentioned on the video but it's likely they wanted to replace the screen because it's an expensive part that they can then "refurbish" into another device.
122.
by cjbprime 2y ago
There are some ARM SBCs with e.g. 32GB RAM and an NPU for under $300, such as the Orange Pi 5 Plus, but I'm guessing refurbished Apple Silicon hardware is the best answer for the price.
123.
by cjbprime 2y ago
How would they game them? As I understood it, this leaderboard is one where you vote on which output is best without knowing which LLM prepared it.
124.
by cjbprime 2y ago
Oh cool! But at the cost of twice the VRAM and only having 1/8th of the context, I suppose?
125.
by cjbprime 2y ago
I think the parent is pointing out that if the signature has to be both syntactically and cryptographically valid, then this would defeat the fuzzer for obvious reasons. But I don't think it does, for this vulnerability. The signatu
126.
by cjbprime 2y ago
(For a model of GPT-4's size, it could also be 8 nodes with several GPUs each, each node comprising a single expert.)
127.
by cjbprime 2y ago
Yes, but I think it's standard to do inference at q8, not fp16.
128.
by cjbprime 2y ago
In general you can swap B for GB (and use the q8 quantization), so 8GB VRAM can probably just about work.
129.
by cjbprime 2y ago
The rumor is that it's a mixture of experts model, which can't be compared directly on parameter count like this because most weights are unused by most inference passes. (So, it's possible that 400B non-MoE is the same appr
130.
by cjbprime 2y ago
Fair enough, although it means we don't know whether a 1.8T MoE GPT-4 will have a "size advantage" over Llama 3 400B.
131.
by cjbprime 2y ago
Where? I only see comparisons to Mistral 7B and Mistral Medium, which are totally different models.
132.
by cjbprime 2y ago
(You can't compare parameter count with a mixture of experts model, which is what the 1.8T rumor says that GPT-4 is.)
133.
by cjbprime 2y ago
You're ignoring geohot, who is a credible source (is an active researcher himself, is very well-connected) and gave more details (MoE with 8 experts, when no-one else was doing production MoE yet) than the Twitter spam.
134.
by cjbprime 2y ago
It's a very plausible rumor, but it is misleading in this context, because the rumor also states that it's a mixture of experts model with 8 experts, suggesting that most (perhaps as many as 7/8) of those weights are unused b
135.
by cjbprime 2y ago
Has anyone prepared a comparison to Mixtral 8x22B? (Life sure moves fast.)
136.
by cjbprime 2y ago
It's fine to have LLM skepticism as a default, but here it's not justified. Google is showing here that the LLM-written harnesses improve massively on the harnesses in oss-fuzz that were written over many years by the combined su
137.
by cjbprime 2y ago
These links are a little different to the GP comment, though. Both of these cases (which I agree show LLMs being an excellent choice for improving fuzzing coverage) are static analysis, going from the project source code to a new harness.
138.
by cjbprime 2y ago
I'm not sure the post (from 2022) is/was correct. I've looked into it too, and I expect this was reachable by the existing x509 fuzzer. There's a fallacy in assuming that a fuzzer will solve for all reachable code pa
139.
by cjbprime 2y ago
Not quite, you don't save memory, only compute.
140.
by cjbprime 2y ago
No, it's unrelated to quantization, they just weren't using the instruct model.
141.
by cjbprime 2y ago
It's odd to use the word "voice" to describe a conversational text model.
142.
by cjbprime 2y ago
It's hard to imagine what could happen instead. Even with a model with infinite context, where we imagine you could supply e.g. your entire email archive with each message in order to ask questions about one email, the inference time
143.
by cjbprime 2y ago
Doesn't nvlink work natively on 3090s? I thought it was only removed (and here re-enabled) in 4090.
144.
by cjbprime 2y ago
Literally every serious C/C++ project has shipped memory unsafety vulnerabilities. We have discovered, as the global community of programmers, that humans are not smart enough to write C code without doing that. It is time to blame the
145.
by cjbprime 2y ago
It's still a company, still making and selling products, and I think he's still pretty heavily involved in it.
146.
by cjbprime 2y ago
I don't think P2P is very relevant for inference. It's important for training. Inference can just be sharded across GPUs without sharing memory between them directly.
147.
by cjbprime 2y ago
What do you think mind control is ? Think President Trump but without the self-defeating flaws, with an ability to stick to plans, and most importantly the ability to pay personal attention to each follower to further increase the level of
148.
by cjbprime 2y ago
It doesn't sound like you gave serious thought to the arguments. The AGI doesn't need to hack robots. It has superhuman persuasion, by definition; it can "hack" (enough of) the humans to achieve its goals.
149.
by cjbprime 3y ago
Reading Hamlet is harder than writing your own, not very good, short story about something that you’re motivated to share. Reading Hamlet such that you fully understand its internal structures and design and could propose changes to it that
150.
by cjbprime 3y ago
I have some really bad news for you about OpenSSL.

More ›