tarruda

15 ms·

211.
by tarruda 2y ago
From the "Why Open Source AI Is Good for Meta" section, none of the four given reasons seem to justify spending so much money to train these powerful models and give them away for free. > Third, a key difference between Meta an
212.
by tarruda 2y ago
> runs natively on Linux + Mac + Windows + FreeBSD + OpenBSD + NetBSD + BIOS on AMD64 and ARM64 with the best possible performance Is it possible that a single binary will run on both AMD64 and ARM64?
213.
by tarruda 2y ago
It has been a long time since I used stremio, but IIRC it used to have a max cache size for torrent plugins. The default setting was something like 10GB, which means older torrents eventually get deleted after you stream a certain amount of
214.
by tarruda 2y ago
I don't understand the "rewrite X/Y/Z in Rust" trend that has been going for a few years. I'm not familiar with Rust, but I'm almost sure it has a good C interoperability. If a certain piece of software is
215.
by tarruda 2y ago
> Also to annoy people probably. Given how many comments this thread has, I suspect he achieved his goal
216.
by tarruda 2y ago
It is not like strong models are at a point where you can 100% trust their output. It is always necessary to review LLM generated text before using it. I'd rather have a weaker model which I can always rely on being available than a st
217.
by tarruda 2y ago
Interesting tool. I currently use a TODO text file in the root of each project I'm working on, which I update right before I stop working. The lists are hierarchical, so each task I'm focused on has a parent item, which helps me r
218.
by tarruda 2y ago
awesome stuff :D
219.
by tarruda 2y ago
Would be good if companies using AI chatbots had some sort of technical quiz to filter technical from non-technical customers. If they pass the quiz, assume the user knows what they're talking about and that the problem won't be s
220.
by tarruda 2y ago
I recommend reading this paper, as it gives some understanding of the things that are possible with an infected toolchain: https://www.cs.cmu.edu/~rdriley/487/papers/Thompson_1984_Ref... Some modern compiled
221.
by tarruda 2y ago
> Nobody wrote a C++ compiler for this purpose; they wrote a Rust compiler (mrustc) in C++. Edited the comment
222.
by tarruda 2y ago
This might seem funny until you read Ken Thompson's "trusting trust" paper and realize that bootstrapping Rust is a so overwhelming task that someone implemented a Rust compiler in C++ for this purpose: https://git
223.
by tarruda 2y ago
I suggest changing "Remote" to "US Remote", as it is not possible to apply when working outside of the US
224.
by tarruda 2y ago
I wonder if it is possible to have a 100% portable qemu build with this. Would be funny to have a usb stick with qemu + VM that runs on any OS.
225.
by tarruda 2y ago
The author also says this edited model increased perplexity (which as far as I understand, means the quality was lowered)
226.
by tarruda 2y ago
Note that 2 RTX 3060 will probably be significantly slower than RTX4090. Even with RTX 4090, 2 tokens per second is very slow and likely not ideal for most tasks. It is impressive (much faster than previous solutions), but still very slow f
227.
by tarruda 2y ago
Funny thing is that it is trivial to fool these apps by using scripts which interact with the computer/VM automatically, so they will not only scare away good developers, it is likely they are creating incentives for the kind of people
228.
by tarruda 2y ago
Upwork/oDesk has a similar system. In these apps, the hours are only billed when you have activity in front of the computer (moving mouse and pressing keys). I suspect the purpose is the same as platforms like Upwork use it: to ensure
229.
by tarruda 2y ago
Is the source code available somewhere?
230.
by tarruda 2y ago
One thing I'm curious about Tor: What are the incentives for running a node? If there are no monetary incentives, then how does it achieves decentralization? Also, what stops a malicious actor with enough resources (a government) from
231.
by tarruda 2y ago
I have tried on my tablet. The main problem is lack of native touch support, so you have to emulate mouse.
232.
by tarruda 2y ago
There are techniques to extend the context window via fine tuning. The authors claim this method was used to extend Llama 2 to 128k: https://github.com/jquesnelle/yarn
233.
by tarruda 2y ago
> It's quite difficult to continue saying no as they age and this kind of thing becomes socially alienating Being a parent is not supposed to be easy. Just be strong and keep in mind that this will be the best for them.
234.
by tarruda 2y ago
> you buy all kinds of other things for him Not really. My children only gets gifts from me in their birthday and other holidays where it customary to give children presents. I don't think it is healthy for children to get what they
235.
by tarruda 2y ago
> Daisy Greenwell and Clare Fernyhough set up a WhatsApp group to discuss how to stave off their young children’s demands for smartphones IMHO "children's demands" is the problem to solve, as children should not be demandi
236.
by tarruda 2y ago
> When the threshold between annoying and profitable is crossed, Meta's lawyers will start sniping and acquiring users of their IP. I'm curious: given that the model will probably be hosted in a private server, how would meta k
237.
by tarruda 3y ago
One of the most attractive features about Mistral open models is that you can build a product on top of their API, and switch to a self hosted version if the need arises, such as customer requesting to run onprem due to privacy requirements
238.
by tarruda 3y ago
AI + Crypto bull market is a recipe for grabbing tons of VC money
239.
by tarruda 3y ago
You can fine tune without the pre training data too. Mistral models are one example, they never released pre training data and there are many fine tunes.
240.
by tarruda 3y ago
Does this work with integrated Radeon graphics? If so it might be worth getting one of those Ryzen mini PCs to use as a local LLM server.

More ›