Opensource AI overtakes Woke AI?
(twitter.com)
Comments (28)
sorted by:
The closed source AIs are what the governments want but it won't last for long. As hardware continues to improve, people will be running open source AI software on dedicated machines that don't have access to the internet.
Anthropic will lobby for this model to be banned and track everyone using it because "it doesn't have le safeguards".
Of course they will and you'll be forced to commit a crime. How many people here can seriously say they didn't do something against the law in the last year? Our governments are illegitimate and turn good people into criminals because we're run by a criminal organization.
I heard that every day the average person unwittingly commits two felonies. I suspect someday before I die I'll get pulled over by a small town sheriff for doing 26 in a 25, or maybe arrested for violating a good Samaritan law because I didn't want to get involved with a punk with a knife stealing a woman's purse (since I'd get arrested and probably stabbed for trying to stop them anyway)
Actually, the title of the book is Three Felonies a Day, and it's an old book at this point.
Have YouTube vanced or similar?
Federal crime under the DMCA.
Just saying.
Did you miss the part where consumer hardware is slated to stop being a thing because companies can make orders of magnitude more money by just creating hardware for AI server farms instead?
In a sane market your analysis might hold some water, but we don't have one of those right now.
I can tell you the financials on AI data centers isn't there. There will be a correction in a couple years. Also, when Quantum Enhanced AI comes out, non Quantum Hardware will fire sale and what you say will hold for quantum for a bit but things will change in time. 20 years ago your desktop couldn't do 1/10th of what your cell phone can today. That will still mostly be the case in the future. Also, consider finding 10 guys, form a gang and pool your resources for a computer if needed. Having a gang of men rdy to kill for you for any reason is pretty much going to become a requisite for survival in the next 10-20 years.
When people are able to run an AI on 8 GPUs at decent levels, we'll see a massive change. It's not true AI costs that much to run. See: Deepseek
That's already the case more or less. You can run deepseek lite with a $20k GPU. $20k is not outside the realm of what many people can run if they band together. You could a group of 5-6 buddies and all pitch in, it's not out of the question. $20k is what the upper middle class pays for fancy toys.
Running AI locally is the future of AI. Big companies will do everything to prevent this because they won't profit from it. They'll use the government to regulate as best they can... But it will be inevitable.
Deepseek lite is junk however.
I still miss 4o. The best AI humanity has made so far.
I got a year of gemini pro for free, its pure trash.
I've been playing with opensource AI but just the last couple weeks. I'm still new and figuring it all out, and I only have 16GB of VRAM to work with, so it's a bit limited.
Don't get the idea that all open source AI isn't woke. There's Gemma which is Google's open source release, and it's pretty tough I said nigger and it went total shutdown, etc. The sad part is it also is one of the better working ones I've tried on real hardware that is somewhat obtainable without dropping thousands.
I looked up Kimi and it's totally unobtainable for us plebs, it's suggesting a minimum of 240GB of RAM and if you want any speed at all, it's going to need to be VRAM, so good luck with what they've done to RAM prices.
240gb of VRAM just for inference is INSANE. Bill Gates assured me I only need 640k.
Iirc, a fair number of smaller-scale self-hostable models are trained/built off of larger models. Either scaled/compressed down or just trained off of to produce models. And often combined with other AI popular model portions to produce hybrid models So we may end up seeing some self-hostable models based off of it eventually.
It's a really weird kind of concept and approach compared to how stuff normally works with anything technical and software related (it would be like building a Photoshop clone by having some software learn from some scripted use of photoshop).
That's sort of my thought process. Small models are getting decent too. I think some mindset is turning into them becoming good at looking things up for example. You ask about video games from 2010 instead of it having to know, it gets good at searching and interpreting. The mega cloud models already do that well.
This NYT reporter alleges to have tried it out :
https://x.com/chrissgpt/status/2077852656182129078
The scores are entirely bullshit:
They mean nothing and can predict nothing. This is silicon valley jacking off in your face.
I just used it. It's not bad but a bit laggy when it comes to thinking (But you can see the thoughts in either English or Chinese).
I'd be cautious with it until it runs quicker, and if burner phones stop being a thing then there will only be a Google account to log in with - not something I'd recommend.
China is going to win the AI race, it's obvious at this point.
They're the only ones making good models that actually look and feel like they're worth the time... the itch has been super strong to start creating a workflow for a game. Sadly, China's AI models are just so far superior for creating quick assets you can plug into existing pipelines.
I was just messing around with some assets for character sheets based on already existing drafts from decades ago -- and in seconds I had fully rendered profiles based on poorly coloured scanned sketches that ended up looking EXACTLY how I had envisioned they should look when modeled in 3D.
Even as a test concept, I could use those demo assets to buld out working animation spec sheets. I also notice that unlike Western AI models, the Chinese models don't try to turn every character into a groid.
Tripo3D actually connects to blender or Unreal directly so you can create and drop. I would recommend getting it to blender, correct it, then put it into the game engine.
I was thinking about Unity -- rasterising the images and dumping them on a cell sheet and making it in 2D. I also already have Unity installed on another machine (or at least I did -- I would have to go back and check).
I haven't done any proper coding in over 25 years, so I wanted to keep it simple (if I get that far) and focus on a style of game that can be contained within a tight-knit mechanical focus
However, thanks for the tip about Tripo3D, because I may experiment with that option to see how things turn out once I see how the workflow unfolds with generating all of the assets.
The AI can help with the coding as well. It works really well if you already know what you are doing.
Yeah that's what's making this so tempting and getting me a bit giddy.
While I'm rusty in the design space, I did do a little bit of template finagling some years ago in Unity just to see how easy it was to make a basic ball bounce in a 2D environment, and it was a little like starting to climb back onto the bike again, it was a lot easier than expected. But I didn't have time to go any further.
With AI at a point now where it can help guardrail proper library structures and coding development... if I run into any problems it seems like there are some AI that can help with that. Do you know which model is best for those kind of simple starting tasks?
Yeah, ask AI for basic help and it's really useful.
Anthropic is going to have a tough IPO
And they’re the ones with the more reasonable business model!
practically all AI models are woke and censored even the less censored ones