Jump to content

immamac

Administrators
  • Posts

    17817
  • Joined

  • Last visited

  • Days Won

    29

Everything posted by immamac

  1. Thank you for sharing.
  2. Coup when you are already in charge? Interesting.
  3. technically true and very problematic.
  4. works jsut fine, you posted an x link with the link text as surlyx paste as plain text then edit.
  5. I can point surlyx.surly.com there and it'll work cuz the TLD is recognized
  6. I'll work on embedding the videos directly if they are video posts instead of images or text.
  7. Looks pretty fucking real.
  8. Did 20 in a row today. Feeling good about the buildup. Gonna start super setting 20/5 when I feel good with 20 then go to 25/5 then 25/10. I feel dramatically stronger/more stable than I did before I started. My core feels much better and my posture has improved quite a bit already. Unrelated to the pushups I'm gonna start doing lunges and wall sits in preparation for an April ski trip. Don't need the legs of steel I had when I was skiing 100 days a year but not feeling noodle legged after a run would be nice.
  9. BlakeMunroeOTF - 1883964261812355138.mp4
  10. I can serve the images under surlyx.surlyhorns.com if that helps also I've found the easiest way to do this is actually just go to the url and then copy/past the redirect image address it spits out in the address bar tweetimg/12345686.webp or something
  11. can you subscribe for a month? I can just do it. .
  12. what channel is it on i can stream it yes..
  13. I was messing with it you probably just got unlucky on restart
  14. you can run it on your own, on your phone without any data leaving your phone ever. but yes, installing their app is incredibly stupid.
  15. I'm working on a way to self host embed all oembeds with a static loader and rich media. Baby steps.
  16. Not a CDN. I host everything out of my own rack. I can cname a surly subdomain just as easily if you want for the embed hosting itself. That's also not my fucking problem and is actually a stupid fucking way to do firewall configuration.
  17. There's actually a ton of parallels between AI and fusion. Fusion just doesn't talk to you so people aren't as frothy hyped about it. Same story different science. It's a money black hole until it isn't and then you've solved the energy problem or intelligence problem. Same hyperbolic outcome chasing, same insane budgets to make it happen.
  18. Inference has always been a losers game. The cost of training is insane, but the cost of inference is where everyone is getting absolutely fucking destroyed. Until you can inference at scale in a cost effective way, LLMs are a dead end for net positive economic outcomes. Very similar to fusion. Training models at 1/10 the cost is a big deal for sure, but if it still costs the same to inference then who fucking cares? It seems like this model doesn't do anything super special on the inference side from what I've read. It still needs as much hardware to do that as any other model of the same size.
  19. Mostly completely wrong, but to a layperson they would think they knew something. This is why if you feed AI prompts that it spits out it turns into indistinguishable garbage in 3 passes or less.
  20. This is where things get sketchy they run around saying they've got 50k h100s they have h800s they also have been working with instinct at AMD. They are doing all kinds of shit, but meta talks about what they are doing, very specifically. https://about.fb.com/news/2024/04/introducing-our-next-generation-infrastructure-for-ai/ They tell you what they are doing not necessarily how. Deepseek hasn't said what they are doing, but they put out a paper on "how", but it doesn't make a ton if sense because unless you magically figured out a new way to use GPUs there's a certain level of computation that just needs to happen.
  21. https://huggingface.co/deepseek-ai/DeepSeek-R1 With AI there's a few phases of actual computationally intense work. 1) Data prep/loading for training. - This is where you get all the stuff you want to show the AI together, cleaned up and formatted and tagged and whatever. You have to be able to tell an AI what to compare stuff to and when and if something is right or wrong after it spits out whatever it is doing in training. Think of this like decomposing a food dish to its ingredients and how much of each was in the recipe. They "tokenize" the data and assign arbitrary neural net locations to these tokens. Simply put this is a brain from newborn until 10-12 years old where you are mostly in feeding and organizing what is fed to you mode. 2) training itself - this is where the big math stuff happens. It adds parameters and models weights to certain training results and just tries over and over and over and over to spit out something it knows is right by a reward system (that was more right than less right etc) they do this countless amounts of times until there starts to be associations between the neural nodes themselves. So the likelihood of the next token is probably one of the nuerons that has a shitload of connections to the prior token. The one with the most connections is the weighted winner by default, but you can mess with the parameters to get a range. Once you do this enough times you start to get to a place where it's mostly accurate regurgitating prompts that you have fed it. From your big pool of tokenized data. Think of this like having all the ingredients to make every dish you've ever had and also the recipe cards and you just randomly mix them and taste the food then look at the recipe and see how wrong you were (not what you did wrong, just that you weren't right) and then you try again until you get good enough to just recreate any dish you are asked and the recipe is close enough. Simply put this is where you learn how to learn and start to use what you've learned in school. 12-19ish brain where you go to school, play sports learn to drive or do things etc. 2b) reinforcement learning or reasoning etc. This is where you do the things in step 1 but have stuff that wasn't in the training set thrown at you at the same time and you are graded significantly more harshly to the point where you need to not only recreate a dish but also do any modifications or substitutes asked in the prompt and it still be acceptable. Simply put this is where you start to really work on specific problems and use your knowledge to do stuff. Think junior year of college or post journeyman trade moving into graduate or masters programs or into unsupervised work in the trades up to and including taking your PE exam for engineers. 19-25ish of the human brain. After step 2 you can plop that out into what's called a model. This is a saved state of all the neural network and the weights and all that, but without having to store the training set, just the tokens and parameters. The larger the parameters the bigger the model (1B, 7b, 30b, 200b, 875b that you see next to models) you can get pretty awesome results with low parameter models that are well trained without having to go crazy on the next part. 3) inference - this is where you take the model and load it into working memory. So the bigger the model the bigger your working memory set needs to be. And all this working memory needs to be immediately accessible so it can start outputting the answer. This is where you can have a restaurant with anyone in the world asking you to cook anything they've ever eaten that's recorded in history and order it their way and you are expected to immediately cook it in as little time as possible. The faster your memory is and the better your reinforcement learning was the faster and more accurately you can serve this dish. Simply put this is being the engineer or doctor or tradesperson or whatever you functionally expertise at as an adult. This is what you get paid to do with your brain not your body. Human brain 25+ish. Except what we are trying to do with these LLM and agentic models is actually just do this whole thing once and have it be expert level at everything. So instead if making a bot to win iron chef we are trying to make a bot that can win iron chef, master chef, survivor, big brother, call the NFC championship game, engineer a building, fix traffic in Chicago, make a cure for cancer, build the best app in the app store, etc. It can do anything you throw at it and it can do it awesome. Small language models or models trained specifically on a certain topic are far better and you can take an llm and in 2a specialize it and then it can just win iron chef every time. Models right now have 3 styles of release. 1) Commercial/Proprietary - you pay us we give you the license and let you use it according to a specific cost structure. That's whether you pay per token or run it yourself and pay a license fee. 2) Commercial/Open Source - you can download the model and use it on your own computers or you can pay us to use it on our computers for a specific cost structure, unless you make a bunch of money off using it to actually do something and then we get a cut of that (meta llama model) 3) Completely Open Source - you can download it and do whatever you want with it as long as you say it's our model. Tons of people can offer this on any cost structure, but the devekoper usually offers a plan too (deepseek) If you had 8 H100s or MI350X or whatever that had about 800GB of HBM for the GPUs in a cluster you can run deepseek-R1 with no germs and conditions. Download it from the link up top. They have distilled models that make it smaller so you can run it on a laptop all the way through a pretty beefy workstation. Basically they gave the brain away you just need to put it in an android, how much memory your android has is how fancy of a model you can install.
  22. Tell them they are fucked and the orange man in charge is a racist piece of shit. There is no other way to put it, they should be scared that's a feature not a bug. Tell ICE to fuck off if they ever approach you, just like any other LEO, never ever talk to the cops. You can help them by keeping in touch with them when they leave and giving them support when they go back.
  23. The deepseek claim isn't sketchy from what the model does. It's sketchy because no one knows if they actually trained it with less GPUs. The model is as good or better than openai o-1 period. It's better than llama 4. If they did in fact do it with 1/10th the GPUs (complete horseshit) then it's a big moment of accountability for a lot of companies who are overspending on a solvable problem. They have been working with AMD as well, so fo say it's just H100s is silly. It could be that they are using GPUs from both at the same time in a novel fabric architecture, but I'm skeptical until they talk more about what the compute cluster general architecture was to train it. The models are real though and not at all aketch, there's a reason they made them MIT licensed and broadly available. It was to shit on everyone.
×
×
  • Create New...