cjbprime

19 ms·

241.
by cjbprime 3y ago
You are simultaneously arguing that it's not possible to accidentally kill everyone by having a switch be in the wrong position, and also that when this actually happens in reality it shouldn't count because "that was pilot e
242.
by cjbprime 3y ago
I feel like you should retract: > Also, there isn't a "if you forget this everyone suffocates" button. (Also note that the pilots did not move the pressurization system out of AUTO; you are blaming them for not discoveri
243.
by cjbprime 3y ago
That sounds lovely, but over here in reality we have: https://en.wikipedia.org/wiki/Helios_Airways_Flight_522 > An investigation into the crash by the Air Accident Investigation and Aviation Safety Board (AAIASB) co
244.
by cjbprime 3y ago
seat you are sitting on gets literally ejected from the airplane "Man, it's unfortunate that there's a problem with the pressurization system in this.. sky."
245.
by cjbprime 3y ago
Yes. I think the aviation community would use something like "explosive uncontrolled depressurization", to distinguish from e.g. someone forgetting to arm the pressurization system.
246.
by cjbprime 3y ago
The lack of nvlink doesn't matter much for LLMs, which can work on distinct layers of the model on completely independent GPUs.
247.
by cjbprime 3y ago
Is it possible that they fine-tuned on the leaderboard test set?
248.
by cjbprime 3y ago
That's especially valuable since it's normally impossible to take even the benchmark percentage increase claims seriously for LLM research, because they're usually reached through training set contamination of the test questi
249.
by cjbprime 3y ago
This looks great! Are you aware of any similar projects?
250.
by cjbprime 3y ago
Some of the recent transformer models can work with audio clips just a few seconds long. I'm sure the final output is less good, but perhaps your friend has audio clips that would work for that from e.g. home movies.
251.
by cjbprime 3y ago
UX looks good, would be great to add ChatGPT's bidirectional voice chat too with e.g. Whisper and Bark and end-of-utterance detection. (That would be a strong differentiator.)
252.
by cjbprime 3y ago
Fair enough. Is there any shared definition of "open source" which permits use restrictions, then?
253.
by cjbprime 3y ago
And not by opensource.org's definition, which prohibits use restrictions. It's not reasonable to act like OP is being idiosyncratic when this fails to meet the protected definition of "open source".
254.
by cjbprime 3y ago
Some infosec and programming channels: Darknet Diaries https://youtube.com/@JackRhysider LaurieWired https://youtube.com/@lauriewired LiveOverflow https://youtube.com/@LiveOverflow fasterth
255.
by cjbprime 3y ago
I guess I don't agree that the analogy doesn't hold. You agree that the training isn't infringing, but you think my analogy is not the reason why. I think it is the reason why. Use without redistribution is not protected by c
256.
by cjbprime 3y ago
This was a lot of words to say that LLMs aren't human. I agree they aren't human. Why do you think that training on GPL code that the model does not redistribute is a violation of copyright law?
257.
by cjbprime 3y ago
Once you think you're at parity, it would also be instructive to try to use differential fuzzing with a compiler (where the fuzzer is synthesizing valid programs to compile) to look for inconsistencies.
258.
by cjbprime 3y ago
TSMC is a Taiwanese company, but it has fabs in both Taiwan and mainland China.
259.
by cjbprime 3y ago
Thanks, I think so too, but the threat model is a bit odd. On a Mac, potentially malicious programs do not normally have access to files in every location (e.g. the prompts to allow a process to access your Documents dir); there is hardware
260.
by cjbprime 3y ago
There are many open questions here. It is unclear that an LLM that has "seen" GPL code is a derived work of that code. Your brain can "see" GPL code without violating the license, even when you later write different code
261.
by cjbprime 3y ago
That's not describing a bad outcome, it's describing how the tool works.
262.
by cjbprime 3y ago
It doesn't have database access to its own training dataset; it only has access to the weights it lossily-compressed that training dataset into.
263.
by cjbprime 3y ago
> Lots of ways for this to go sideways even if the data stays local. Could you name some?
264.
by cjbprime 3y ago
These are not citations. The point is that it does not / can not reliably cite the actual sources it used to prepare an answer.
265.
by cjbprime 3y ago
(I don't think there's any reason to expect GPT-4 to actually know its current cutoff date.)
266.
by cjbprime 3y ago
> GPT4 is absolutely capable of providing links to sources and citations. Do you mean in the Browsing Mode or something? I don't think it is naturally capable of that, both because it is performing lossy compression, and because in
267.
by cjbprime 3y ago
> These questions are either behind a paywall (in the case of the question bank) or published after 2021 and therefore out-of-training data for both LLMs I'm not sure that it's true that GPT-4's web scrape is currently fro
268.
by cjbprime 3y ago
Odd that the article doesn't use the more standard term "parser differential", with "differential fuzzing" as the fuzzing community's method for finding those.
269.
by cjbprime 3y ago
Since the amount of VRAM is low (max 24GB), one question to look into before investing might be whether there's support for chaining multiple cards together.
270.
by cjbprime 3y ago
I don't know whether you're correct, but I'm not sure I followed your intuition for why the LLM could access the second order relationship during inference, and yet it wouldn't appear in the distance threshold on a vecto

More ›