Hi @Adrienne I will try to give my input on those.
I’ll leave out my long term vision here, because I believe that we can do much more, but we need to start strong and realistic.
Question 1: What specific problem is POKT Network best placed to solve within AI, and for whom?
Model inference. Running models is not easy, paying for someone else to do it (APIs) is expensive, hosting your own is also expensive as GPUs are almost always needed (for larger models such as LLM or Diffusers).
The ones that would consume this service are developers that need access to APIs, instead of dealing with your own deployment or paying expensive APIs you could just use Pokt and access a series of models to solve many different tasks, all in a single endpoint.
Some examples could be chat bots, summarization, question answering, image generation, aided graphic design, segmentation, sentiment analysis, etc.
Also we can do much much easier tasks such as embeding, while they are cheap to deploy on CPU, I don’t see why we should not also host them (they are also very simple to meassure)
Question 2: What strengths can we draw on to solve this problem effectively for them?
We have a working network and tokenomics that encourages node runners (in this case model runners) to provide services with high quality and is always aiming to maximize their productivity.
If we set the incentives right (aka metrics or ““QoS””), the node runners will be doing the work of getting the best models for each task and the end user (developer) will never have to deal with model updating or changing from one API standard to other.
Node runner have already done this for blockchain, why not also for ML models?
Question 3: What do we need in order to have a solution ready for testing?
We don’t need much to get this up and running in Morse, only be sure on what we will be offering and how we will be testing. The missing piece that we need to build is a credible way to show that our endpoint are not answering trash.
Besides that. we can setup our nodes and gateways and start selling API access to known large language models (LLMs), diffusion models (text-to-iamge) and others today, the network already supports that.
Since we do not try to bring inference into the blockchain or deal with distributed model training (two open problems in my opinion) we can push a real service faster than any other player in the crypto+ai world.
We need partners in the AI world, both to provide inference (spare at a low token sell pressure) and also to guide us on what exactly is needed (a target use case would be wonderful). We can adapt from there, we have the tools and expertise.
Question 4: Who are we competing with in this specific problem space?
From the Crypto+AI world, I don’t believe that there is a real competitor, most projects try to solve other things, like distributed training/inference or GPU computing.
In the AI world we are competing to every other API service out there. But remember, we were also competing against Infura at the beginning, so, competition for Pokt is only temporary
Question 5: What is the customer (from Q1) looking for in a solution?
This is only opinion, but from my experience with model usage:
- Ease of use and migration. When I do a solution including ML and I change the API or the underlying model, it is a pain to re-adapt code for something that should be transparent. This is reflected in the LLM community where self hosting solutions like vLLM try to mimic OpenAI API standard and make it seamless to change the back-end.
- Model specialization. Some tasks you know that can be achieved by cheaper (smaller) models or even be better solved by specialized ones. With an API you are married to a single model (a few options at least) and with your own hardware, well, you run out of GPUs… Pocket could enable you to do model mixing on the fly at no cost.
BONUS ROUND: What could go wrong?
We are truly permissionles, we need to be able to asses that our service is real. As opposed to blockhain nodes, ML models are countless and highly sensible on deployment choices, meaning that trivial tests like majority voting will fail.
We need to find the balance between vagueness and over-specialization. If we only allow a single service per model, we will have >30000 new services with a single node in them if we are too vague the customers wont know how to get what they need.
In my experience, the AI community is not crypto native and wont be lured by simple saying that we are a permissionles service and that we have a token and long live cryptoanarchy. We need to convince them that we provide real value, that we play on their terms and that we speak their language. If we nail that, they wont care if we are crypto, they will stay for the service.