CTRL-Altman-DEL: The Dawn of Decentralised AI
CTRL-Altman-DEL: The Dawn of Decentralised AI
Bittensor could catch some tailwinds on this OpenAI debacle
One thing about living in Asia hours is that you often wake up to major news and have to play catchup.
Sam Altman getting fired from OpenAI on Friday, for example. I almost choked on my milk.
Why would the board fire someone who’s clearly extremely intelligent, has an exemplary track record of success, and gave an excellent keynote address at the OpenAI conference 12 days ago?
Cue the spicy theory-crafters. Andrew Cote believes it was politics, that Altman was fired because “he would move AI forward too fast by deploying a recent breakthrough.” And some people didn’t like that.
OpenAI has a very awkward (almost dysfunctional) corporate structure because it started as a non-profit entity that later transitioned towards a for-profit enterprise. Today the non-profit controls the direction of the for-profit entity while providing investors a capped upside.
It will be a spicy couple of weeks as the truth emerges.
Will this be a Steve Jobs moment? Does Sam go on to start another company to rival OpenAI?
What’s clear, though, is the shroud of mystery that envelops OpenAI’s internal operations. Despite GPT becoming ubiquitous and used by hundreds of millions worldwide, there's a palpable disconnect.
As everyday users, we find ourselves on the outside looking in, trying to peek through the veil of secrecy surrounding these AI giants. This lack of transparency is worrying.
Blockchain + Crypto = ???
Lately, I’ve been wrestling with the question: What will the intersection between crypto and AI be like? It’s vague, but most would agree there’s monumental potential waiting to be unlocked.
When we think of AI x Crypto, we typically think of Akash Network and Render. These are decentralized networks for GPUs, which can provide the necessary compute for the training of AI models. The logic is straightforward - as AI continues to skyrocket, so will the demand for computational resources. Peer-to-peer networks, in this context, could experience significant growth. So they’re in the business of picks and shovels, but I think this is just scratching the surface of the potential of AI x Crypto.
It’s like saying monkey JPEGs are the pinnacle of what NFTs can offer.
And then I came across Bittensor.
ELI5: Bittensor
Bittensor focuses on AI inference (downstream), which is where trained models are used to generate outputs.
It's a decentralized network that incentivizes AI models, particularly Large Language Models (LLMs), for various tasks like text generation, image creation, and music production. The network comprises over 27 subnets today, each focusing on specific tasks.
Think of Bittensor as a decentralized ChatGPT + Midjourney + anything else AI can do.
The network operates through two main roles:
Miners (Value producers): Miners develop and host AI models on the network. They are rewarded with TAO tokens based on their models' performance related to the specific task, which incentivizes the development of better and more efficient AI models.
Validators (Consensus producers): Validators evaluate the outputs from miners, ranking their performance on the specific task. They also interface with users who submit tasks to the validators (e.g. for the image generation subnet: “/imagine Sam Altman with a Darth Vader mask at Thanksgiving dinner”) and route them to the appropriate miners
I’m probably oversimplifying the technical intricacies, but a few things stand out to me:
Miners and Validators in the network exchange knowledge and share parameters, allowing for self-optimization over time
The network is designed to harness the strengths of multiple separate AI models to generate the best possible outputs (“Mixture-of-Experts”)
It’s beyond my intention to get into technical details, but here are a few good summaries that have helped me to understand Bittensor better:
Knower — A short report on Bittensor and AI
You can give Bittensor’s chatGPT equivalent a spin here
The TAO
TAO is the network's utility token. It has a tokenomic structure similar to Bitcoin: a hard cap of 21M tokens and a fair launch with no VC allocation. It even has a halving cycle, with the first halving happening in 2025.
There is 5.65M TAO in circulation today, and all of it was fairly distributed via by mining and validation on the network. The circulating market cap is slightly over $1B today. 7,200 new TAO are issued every day to miners and validators.
My early thoughts
Bittensor is still in its infancy. The network boasts a dedicated, almost cult-like community, yet the overall number of participants remains modest – around 50,000+ active accounts. The most bustling subnet, SN1, dedicated to text generation, has about 40 active validators and over 990 miners.
What's truly captivating is the concept of a decentralized AI network. This mitigates risks of centralization and raises a question: Could these unique economic incentives foster AI models that surpass those developed by heavily funded entities like OpenAI and Google?
Before LLMs became mainstream with tools like ChatGPT, deep tech startups often focused on acquiring proprietary datasets to develop specialized, machine learning-based AI models for very specific tasks.
For instance, Flatiron Health uses real-world clinical data from oncology patients and develops AI models that feed into tools that support cancer researchers and care providers. Traditionally, the startup's goal was to productize and monetize these proprietary models.
Bittensor, however, might represent a shift in this paradigm. It's perhaps more fitting to call it a business model innovation enabled by technology rather than a technological breakthrough. For example, it offers a pathway for proprietary data and AI models to be developed together and used by a wider audience without the need for open-sourcing them.
I envision one potential future where Bittensor hosts thousands of specialized subnets tackling various challenges, from environmental and healthcare issues to energy solutions.
And if I’m being honest, there’s something I find fascinating about a team that designs its tokenomics in the same way as Bitcoin. It speaks to its motivations, a different breed from today's teams, which are often optimizing their tokenomics along the VC-funded model, with large allocations for founders and investors.
I’m not sure where Bittensor will go. It could be a 100x success or a complete bust. But the potential and the philosophy behind it are too compelling for me to ignore.
Update 25 Nov 2023: Listening to this Delphi Digital podcast on Bittensor was worth every minute.
NGL, it's hard to piece everything together about TAO from docs and Discord. Shaabana breaks it all down so it's easy to understand.
Here are a few of my quick notes:
Multiple models vs single model: better, more varied outputs
Subnets enable more specialized models and outputs
Decentralization of AI is important: anyone can deploy a model and get feedback quickly on its performance via validators. Open source can grow faster
Need to get legitimacy from the web2 AI world: going to conferences, publishing papers to build reputation
Bittensor is meant for businesses/apps to be built on top of it (via APIs), less so for the end consumer
Great piece. Would you regard Altered State Machine an interesting project at this intersection?