kromem

53 ms·

571.
by kromem 3y ago
'Powering' It's like pissing in the ocean and thinking you caused a tsunami in Southeast Asia.
572.
by kromem 3y ago
Not really. Copyright largely remains about the PRODUCTION of content, not about the CONSUMPTION of it. Someone who grew up reading Marvel comics being able to make new original comics in that style is perfectly ok. That same person perfect
573.
by kromem 3y ago
It will be interesting to see a case trying to argue competition. On one hand, a static training with a cutoff date and no perfect memory of training data makes for a hard case that the underlying use of the copyrighted material constitutes
574.
by kromem 3y ago
It is. That response is even thrown by GPT-4 too, btw.
575.
by kromem 3y ago
So while the technical why is probably some kind of repetition cycle, some of the specific content is REALLY interesting. In particular the "image of the image" part. This seems to be a reference to ideas found in the Gospel of Th
576.
by kromem 3y ago
I was just reading a piece the other day about how one of the two baker's dozen grandfathers of superconductors had his wife call in a panic after seeing the LK-99 news, worried about what this would mean for their child's future.
577.
by kromem 3y ago
As mentioned in the article, there's also studies where it may improve things. This particular one is a mouse model with induced DSS colitis, which I kind of wish would be erased from methodology forever across literature. Not only is
578.
by kromem 3y ago
Ghost attention is only used in the 70B model in llama 2, FWIW. So would need to make sure comparing apples to apples.
579.
by kromem 3y ago
It's also a false sense of security. A bad actor can sidestep alignment safety measures fairly successfully for the foreseeable future using dynamic jailbreaking efforts. Good actors get penalized by the product being made notably wors
580.
by kromem 3y ago
Just a tip - I forget where I saw it, but at some point in reading over research I saw that using 'Q' and 'A' results in lower accuracy than 'Question' and 'Answer.' Which probably fits the latter bia
581.
by kromem 3y ago
Yes, and I'm willing to bet that within 12 months we'll be looking back realizing that this was due to the fine tuning taking the world's SorA pretrained model aligned with "completing human tax" and putting it in t
582.
by kromem 3y ago
It's more like training a massive neural network to auto complete the largest collection of human written text for hundreds of millions of dollars and then spending tens of millions on fine tuning it to ignore massive components of tha
583.
by kromem 3y ago
The other REALLY big issue is the CBD content. CBD is on its own an effective antipsychotic (comparable efficacy to one of the common treatment options in schizophrenia in a 2013 study), and has repeatedly been shown to balance out the psyc
584.
by kromem 3y ago
It's interesting watching the LLM 'security' industry relive the 2000s attack/defense patterns in a different medium. Improperly separating user input from control statements? Enjoy your injection attacks. Trying to dete
585.
by kromem 3y ago
Also, very importantly, it's not saying not to build a chatbot, but to recognize that the main consumer of your chatbot interface will be a user's primary LLM, not the user themselves. The headline is exactly the kind of thing the
586.
by kromem 3y ago
Yeah, but go even further back in that timeline. Like 1952 when multiple newspapers have front page news about UFO sightings over Washington DC. And then in 1952 Gene Pope, a graduate of MIT in only three years, leaves his position in CIA p
587.
by kromem 3y ago
I don't get why no one ever throws out time travel as an explanation over extraterrestrial. Given the limit of the speed of information and the expansion of the universe, there's a VERY small area that would even be aware of the s
588.
by kromem 3y ago
It probably depends on if the correlation between consumption of the media and going on to actually harm children is causative to any degree. There's probably a benefit to devaluing the market for the material, but I'm skeptical h
589.
by kromem 3y ago
What's been a very interesting deep dive over the past few years is that there was almost a simulation theory like rethinking of these ideas attributed to Jesus of all people in the first to fourth centuries CE. Part of the debate on t
590.
by kromem 3y ago
Don't use instruct/chat models when the pretrained is available. Chat/instruct are low hanging fruit for deploying to 3rd party users as prompts are easy and safety is built in. But they suck compared to the pretrained models
591.
by kromem 3y ago
The problem with your scenario is when the consequences are unknown. People who aren't in medicine really have a hard time understanding just how grisly the things the human body can go through might be. Medical staff saw during COVID
592.
by kromem 3y ago
Not OP, but laws against distribution? Absolutely. Laws against consumption? A terrible idea. What we should want to do as a society is funnel people interested in trying things we think they really shouldn't do towards legal chokepont
593.
by kromem 3y ago
Yep. It isn't going to be in the next 12 or maybe even 24 months, but in the not too distant future any company that made significant downsizing in the early days of AI is going to find out the hard way that they effectively threw away
594.
by kromem 3y ago
This is happening on a lot more gaming news sites than just this one. I'm regularly seeing articles in my Google News feed that are summarizing Reddit threads with questionable accuracy in their coverage that are clearly AI generated.
595.
by kromem 3y ago
The intent may genuinely be to help decrease bot activities versus human activities. Even the ad example is about not charging advertisers for bot views, which is a huge problem right now. The problem is that a tool can often be used for ev
596.
by kromem 3y ago
I'm curious why you think the commenter is interpreting according to the Copenhagen interpretation? It sounds more like they are working within a view similar to the Von Neumann–Wigner interpretation.
597.
by kromem 3y ago
In the decades since the establishment of these theories, both the continuous (classical) spacetime of general relativity and discrete matter of quantum mechanics, the world has changed in rather significant ways. One of those has brought f
598.
by kromem 3y ago
Think more that extended AI voice rights become part of standard contracts and video games with a 60 hour main story and voice acting have hundreds to thousands of tailored extended questlines powered by AI to extend the original performanc
599.
by kromem 3y ago
It's very cool work. I've been wondering for a while at what the next steps in adding 'inefficiencies' to AI processing would look like, commenting the other day to a friend that what's needed in the next 18 months
600.
by kromem 3y ago
Sure - not like most of the infrastructure of pretty much everything online is built on top of projects originating in that space or anything.

More ›