ICP has Superior Culture
HTML-код
- Опубликовано: 27 сен 2024
- Culture is what defines a country and determines their social standing in the world. Similar to countries, each blockchain has its own culture, and this culture is what will determine their faith in the technical landscape of tomorrow. In this episode, I provide an example of a stellar individual in our community. It is individuals like him that will make Internet Computer the best blockchain network in the world.
Instagram: JamesAllan0
Patreon: / cityxcape
Thanks for the video James and thanks everyone for the insightful comments!
You're one of us Bryan
ICPis the best❤❤❤👍💯
The Dfinity developer forum is great, I'm sure they would be able to answer such questions.
Agreed I will try my luck there
Let me know if you find anything good there Bryan
ICP ❤❤❤!!!!
You couldn't currently run an LLM on ICP. You need 28GB of RAM to run a 7 billion parametres model which is on the smaller side of models that are used in industry. It'll fit on a consumer grade GPU. Something like ChatGPT needs very specialised hardware which is beyond anyone but major companies. The bigger problem is that you really need to run it on a GPU to make it run efficiently so that the calculations run in parallel. Moreover, the calculations in the neural network are highly optimised to run on GPUs which probably wouldn't work on a CPU. There are CPU optimised models too but I struggle to imagine deploying one for use by consumers. It'd take up all of your CPU memory/compute to just run that and it'd be quite slow. They have talked about adding GPU support to nodes though. The AI demo they showed was still very impressive because it shows the power of their nodes.
My goodness! What a sharp answer. Do you mind if I share with the community? Also what is your background? The community might want to know that as well.
@@cxc sure no problem. I’m an AI researcher that works with LLMs. I’m sure I could do better background checks on running LLMs on CPUs but the above is my initial/gut response.
@ln2deep
Given your synopsis, do you still see a strong future for ICP? I'm no tech expert, but everything I've researched and watched on ICP/Dfinity is that this is s slam dunk long term investment.
@@JGR1517 It's definitely the most complex tech out there and it has a great team. Most of the 'work' on AI at the moment in industry is related to legal aspects of AI systems and compliance and the blockchain could help with that in one way or another.
My main AI play is Render because of its very straight forward value proposition and it has extremely impressive collaborations. Having said that, it just tackles the horse power part of the equation. Render tokens are basically GPU fuel while ICP tokens are compute cycles. But ICP is also a 'world computer' which the compute cycles run on. It is mind-blowingly more complex.
My main question with ICP is if it's really worth moving Web 2 apps onto Web 3. It might not always be worth it, but if ICP corners the Web 3 market then that is huge. I don't see why the majority of websites wouldn't run on ICP. It may well come down to compliance and cost. Blockchain is great for compliance issues but the question is whether it can be more cost effective. It's my biggest bag so I sure hope so.
❤ icp
Right now it’s not possible only with cloud infrastructure. But I think it will work in the future.
I see
Great content Bro!
Glad you think so!
thanks for the video😊
You're always welcomed here
great video 💪
Appreciate it!
I am an anarcho-capitalist and hope that you are correct about ICP being a libertarian culture. I am also building on the IC and will test that theory.
Nice to see a fellow anarcho-capitalist. I also view myself as a classical liberal, meaning that I believe in the social contract
Hi James, really cool that situation you have encountered, I am not qualified to fully answer the question but as I recently started studying about A.I , I could understand that the process of collecting data and processing it can be very complex, when you are taking about imagery which needs to be analyzed by the A.I tool in a small scale project the datasets from the LLM and CNNs(Convolutional Neural Network) in a small scale it could reach 1TB easily. So my recommendation and generally what a lot of people do is they develop and train the tool once they have the data set and the tools is starting to work the way they need you can move the project to another environment this approach is much easier then create speculations without know how much it will cost, etc.. however I personally would create this project at home and gather the resources a need to do so later once the project have mature enough then I would make a decision where I would be hosting it, unless you have really good budget etc... prior to do the LLM and CNN data set project is very difficult to calculate cost but if you start to do that in your own environment it will give you a good idea then you can start to speculate more and refine your conclusions. if you intend to train the model in the blockchain we need to remember that the processing power would cost even more.
However I also remember now that I watched a video from someone in the Dfinity team mentioning about A.I as far as I could understand the ICP network is not there yet to support A.I big datasets etc... but I might be wrong... cheers!
@cxc I take you back to Dominic Williams' first AI on-chain video, he states that (at the time) they were running 32-bit which allows for 4GB storage, that they were moving to 64-bit which would allow up to 32 GB storage and leading to LLM's, i'm not sure if that means the 64-bit system would actually allow LLM's or if he meant in future if they moved to 128-bit
An AI researcher has answered the question which I'll share with the community in the next town hall, but according him even the upgrade you talk about would not be sufficient. These LLMs are monsters
From my exp with akt it is possible and the hardware requirements are dense, but stability seems to be one of the current choke points for running LLM via web3, which would bring a hospital lower on the list of appropriate early adopters as system stability is key in healthcare. One of the difficulties in decentralized rental GPU for model training is that the person renting it out to you can decide they need their GPUs back...I've seen customers find work-arounds by using models with parallel algos involving other GPUs of similar density...think 4 sets of H100 providers and if one goes down the algo automatically connects to the second best choice. It would be cool to see ICP testing SLM first with less vital services to explore capabilities
Legend 🤙🏻👍🏻🤙🏻
The 4gb will double or quadruple soon, I think if I heard eight from Dominick on dfinity.
So if nodes are not using GPUs as of now then I wonder will Icp move towards nvidia for there gpus if so nvidia is absolutely killing it. Also they just did a 10/1 split nvidia is going crazy
I own more then you now fingers crossed 😂 that I’m good staked 8 year gang
Everything great, top and then price goes down (: Something else is problem, and nobody knows. Lost more then 40% of value in two months. I’m not sure for happy end.
How to become whale in ICP?
You gotta own 10,000 or more
Nice talk about libertarianism. Which movement within libertarianism do you adhere to?
I run LLMs locally on my desktop and this requires a CPU or GPU. GPUs are much faster in this context and without a GPU the results take painfully slow time. Dominic's demos have shown that it's currently possible to run LLMs on ICP, but among other things, GPUs still need to be added to the network to run the LLMs fast enough and I think this is planned.
In this context, an interview with the Elna AI team may be interesting: ruclips.net/video/LFj3CtqBfXE/видео.html. It may be more efficient for him to continue to build on their project.
Thanks for the insight
Thanks for the lovely answer. I'll be sure to check out the interview you suggested from Blockchain Pill, it seems good.
In regards to the libertarian question, I'm honestly not sure. I think I'm what people call a classical liberal, meaning I believe in the social contract. What about you?
I’m proud to announce I’m now a icp whale just did it 😅
That's great news... Congratulations. I'll share your whaleness to the community in the next Q&A
How many ICP u hold?
@@sjawahar2001 He holds like 12,000
Is not possible to train LLM’s yet, it’s on the new roadmap. Just inference
Here are some general guidelines:
- Model size: 1-10 GB (e.g., BERT-base is around 350 MB, while larger models like RoBERTa-large can be up to 5 GB)
- Memory per node: 16-64 GB (depending on the instance type, e.g., ml.p4d.24xlarge has 64 GB)
- Total memory: 128-512 GB (or more) for distributed training with multiple nodes
Keep in mind that these are rough estimates, and the actual memory requirements may vary depending on the specific use case and model configuration.
Gangculture
(8 year)
Gang Gang!