redox99

31 ms·

661.
by redox99 3y ago
I wouldn't mind having my face deepfaked into a porn or whatever as long as there was a small watermark or something that indicated it was fake. I think enforcement should be focused on disclosing what's AI generated/fake, no
662.
by redox99 3y ago
> So far, his track record in AI research is non-existent. I think he just bet on the wrong horse (RL instead of LLMs), at least in the short term.
663.
by redox99 3y ago
This one is free from VC funding and promises to stay away from it. So it should not suffer from enshittification. See https://github.com/usebruno/bruno/discussions/269
664.
by redox99 3y ago
Disappointing that there isn't a standalone program like Postman. I don't want to install an extension in my browser to handle CORS.
665.
by redox99 3y ago
> Now we’ve transitioned to mostly AI writing the bulk of the code with the senior developers still doing their thing. > We’ve found the code quality coming out of our AI is a lot better than the H1b’s could produce. That sounds weird
666.
by redox99 3y ago
Mastodon has a very heavy moderation against anything that isn't the "SF mentality". And servers that don't moderate like this get defederated. I prefer the more lenient X moderation.
667.
by redox99 3y ago
I'm surprised by your pricing. $1.20/hr for an A100? That's significantly cheaper than something like runpod. Are you offering this service at a loss?
668.
by redox99 3y ago
> 3 billion parameters is a bit large for most smart devices, is it not? Not really. A 3B model quantized to 4 bits should run in any reasonable smartphone (using around 2GB of memory).
669.
by redox99 3y ago
Falcon (from UAE) is also censored
670.
by redox99 3y ago
Legally speaking, what could happen to him?
671.
by redox99 3y ago
You can run a 7B and 13B quantized model on almost any machine on the CPU if you're patient.
672.
by redox99 3y ago
This is still full attention so it should be better in benchmarks, whereas the sliding window attention is more efficient so it allows higher context sizes.
673.
by redox99 3y ago
> None of that has anything to do with "moving fast and breaking things". Take away Musk's welfare and SpaceX and Tesla are out of business within a year. Explain then the reason why SpaceX is more successful than its comp
674.
by redox99 3y ago
The SLS has even more government welfare and it's doomed to commercially fail because of its absurd costs per launch and lack of reusability. So clearly you won't succeed "just" because of government welfare.
675.
by redox99 3y ago
Right, and the SLS doesn't have massive government handouts? Lol
676.
by redox99 3y ago
> Not a big fan of how server-centric the LLM landscape is. That's just not true. You can get ooba[1] running in no time, which is 100% made for desktop usage. There's also koboldcpp and other solution also made for desktop use
677.
by redox99 3y ago
The problem is investors. Investors don't care about a sustainable business. They want infinite growth, and sucking dry both the the customers and the business itself.
678.
by redox99 3y ago
Downloading big games (apple showcased Resident Evil village I believe)
679.
by redox99 3y ago
Nowadays shouldn't it be easy to install solar panels and megapacks, and operate independently from the grid?
680.
by redox99 3y ago
> I have not yet seen it successfully done, and I suspect that updating fractions (~.1%) of the original weights with PEFT methods won’t help. Nitpick, but although when training LoRAs you're only training 1% or less (depending on r
681.
by redox99 3y ago
> I most certainly do, and she has no privacy That sounds awful
682.
by redox99 3y ago
> Funny enough, no matter how big these things get, they never seem to make me a lot more productive. Maybe because the stuff you do isn't bottlenecked by compute? In my case every hardware upgrade resulted in a big productivity imp
683.
by redox99 3y ago
MJ is already better than SDXL if you don't need any of the tools that SDXL has and MJ doesn't. So it isn't that hard to imagine Dalle 3 being better than SDXL.
684.
by redox99 3y ago
You might get 99.999%+ of humans like that. But there's always going to be some people that are immune to it, that didn't catch it, a few crazy blokes that live completely off grid, tribes with no contact, etc.
685.
by redox99 3y ago
You can train a 13B QLoRA (maybe even an 8bit LoRA, I don't remember) with a 3090/4090.
686.
by redox99 3y ago
> to mess around with LLMs You mean training or inference?
687.
by redox99 3y ago
The 24GB of VRAM of the 3090/4090 (instead of 80GB) is a deal breaker for a lot of stuff. Also 3090 only does 2x nvlink (and 4090 has no nvlink). But a lot of people (including me) run 3090/4090 rigs at home, or rent them at runpo
688.
by redox99 3y ago
Almost 100% of their datacenter GPUs are used on Linux. Do do think they run without drivers?
689.
by redox99 3y ago
> But, I didn't want to cover it, because I wasn't sure if that would look suspicious in the case I was in an accident. Now that's some chilling effect
690.
by redox99 3y ago
Cygwin isn't remotely comparable in terms of compatibility. With WSL2 I can train stuff with PyTorch with GPU acceleration without configuring anything. Try doing that with cygwin.

More ›