declaredapple

10 ms·

91.
by declaredapple 3y ago
100MB at 1Gbps is ~0.8s to load. This especially sucks for FaaS and CI/CDs. This makes it difficult to use for on-demand use cases that are latency sensitive. That and that's likely 100MB that's going to be competing for memo
92.
by declaredapple 3y ago
I cannot investigate further because I'm at work, but if they use that clearnet url you linked, they aren't really a darknet (since they're clearnet). Underground makes more sense here as they're clearnet but you have to
93.
by declaredapple 3y ago
What use cases are you using it for? I haven't used it a ton, but I've found the UI to be clunkly and somewhat difficult to figure out compared to using the shell with jq
94.
by declaredapple 3y ago
Parent linked the Lmsys chatbot arena which is where humans blindly get the results from two different models, and vote for the response they liked more. So the LLMs are compared by elo. Do you think goodhart's law applies here, since
95.
by declaredapple 3y ago
You can either modify the model weights in a way that doesn't cause any real differences (change a few bits somewhere should be enough), or you could watermark the actual text output. Here's a list of research for watermarking LLM
96.
by declaredapple 3y ago
Just to throw out there - I haven't had any issues with exllama personally, and it's a lot faster last I checked.
97.
by declaredapple 3y ago
Slight knitpick GPT 3.5 came out 2 years after GPT 3 GPT 4 came out 1 year after 3.5
98.
by declaredapple 3y ago
> - They are more focused on AGI, and not letting themselves get side tracked with the LLM race FWIW OpenAI seems to have a corroded definition for AGI that is essentially "[An] AI system generally smarter than humans". They do
99.
by declaredapple 3y ago
> Do you think that the idea of AI poisoning generalizes, though (even if it would require very different algorithms) Yes, these are all different classes of adversial attacks and I don't think any model could resist everything, ass
100.
by declaredapple 3y ago
> It's not like it has to poison all models; it just has to poison the more expensive ones. Did you mean common ones? I'm not 100% sure what you mean by expensive, compute, api provider costs etc. But unless I missed it they di
101.
by declaredapple 3y ago
I don't think I see it mentioned in the paper but what would happen if you used the poisoned images to train the text encoder? Would that make the text encoder itself resistant to at least the previously poisoned images? The images wil
102.
by declaredapple 3y ago
From what I understand the attack requires access to at least a model to attack. The result has the highest success on that model you specifically attacked, but it apparently transfers albeit with a lower success rate. > how can it be su
103.
by declaredapple 3y ago
> Some hidden API that google give them so they can have a better ranking there? I don't know about that second part - but it would make sense that google (and others) may want to use lmsys's arena to benchmark their models. A
104.
by declaredapple 3y ago
> Yes, but you choose to use Unreal because its actually good. Meanwhile Apple gives you no choice. How did you have no choice, did Apple threaten you or something? Android is a competent competitor you could choose instead.
105.
by declaredapple 3y ago
No it's not. It does (did?) this all the time to me generating tests. I know it _couldn't_ know the right answer because I didn't tell it all of the model and serializer schema's and all that. I don't care, because
106.
by declaredapple 3y ago
> Having said all that, I have heard the analogue inputs aren't that great on the ESP32. Yeah they have wide deadzones and just aren't great, they're really only useful for really coarse things, like reading a photoresisto
107.
by declaredapple 3y ago
> [ESP32 LOLIN32 Lite] Analog input maximum voltage 1.0V Where did this number come from? AFAIK all pins should operate at 3.3v, and I see no mention of a 1v maximum for that board, or the esp-chip either.
108.
by declaredapple 3y ago
I would agree with that it's usable as a single datapoint, used with others (other submissions, general performance, etc). However given that we already have professors literally failing people by just pasting and asking chatgpt, I&#x2
109.
by declaredapple 3y ago
I totally understand the need for an "on-rails" model. However I've found that using it as a personal assistant and for some automation tasks it suffers. The fact that it has only gotten worse as judged by humans doesn't
110.
by declaredapple 3y ago
Passkey's don't strictly require 2FA sources. Some hardware keys like yubikey's generally only prove physical presence of the key. And software implementations exist too.
111.
by declaredapple 3y ago
> Yes, but you want alignment even if you don't want censorship. Instruction tuning helps a lot with this but what a lot of people mean is the refusal to do things. You get to chose how "aligned" it is, for some usecases l
112.
by declaredapple 3y ago
Yeah it's mystifying to me that they because the #1 commercial OpenAI competitor and then decided to go all "Helpless and Harmless"
113.
by declaredapple 3y ago
You don't need to be #1 and beat GPT4 if we're going with "a frontrunner". Gemini Pro as available today already puts Google at #4 by company in the chatbot arena.
114.
by declaredapple 3y ago
Their currently available offerings are "meh", but they're definitely top 5 (Currently they're number 4). This leaderboard is the only one most people trust as it uses an ELO system with humans blindly comparing and rati
115.
by declaredapple 3y ago
Don't passkey's just use the hardware key as the only factor (or at least that's how I've seen it implemented)?
116.
by declaredapple 3y ago
This is the difference between security vs archival perspectives. For security you should assume someone recorded it indefinitely. For archival you should assume nobody recorded it including the original creator.
117.
by declaredapple 3y ago
> The trouble is that those warnings should not be necessary. Of course they shouldn't be, but it's a protection against scenerios such as miscommunication, being distracted and forgetting you handed them over, etc. > mantra
118.
by declaredapple 3y ago
Yup. I think many countries (japan already has) will allow IP for training data. They just need to buy time until then.
119.
by declaredapple 3y ago
I'll venture to say the majority of these "open access models" are meant to serve as advertisements of capabilities (either of hardware, research, or techniques) and nothing more. MPT being one of the most obvious example. Ma
120.
by declaredapple 3y ago
I'm speculating they are attempting to avoid controversy about their datasources. That and a possible competitive edge depending on what specific sets/filtering they're using.

More ›