Does Decentralized Machine Learning Need Crypto? Bittensor (TAO)
Level 2 - Value Investor
Welcome Avatar!
This year has witnessed a remarkable surge in awareness and hands-on experience with AI, as people of all levels of technical sophistication around the world have delved into the capabilities of generative image tools and sophisticated chatbots.
Today we explore the intersection between crypto and AI by looking at attempts to build a decentralized machine learning and intelligence market. If successful, decentralized AI would be in the hands of the people, not corporations - bringing similar benefits to the self-custodial, permissionless, open source DeFi system.
How It Started: The Internet and Bitcoin
The Internet gained traction in part thanks to its revolutionary killer app: email!
Later, the World Wide Web (another protocol which runs on the Internet) enabled online publication and consumption of data by anyone in the world. Later still, only after strong encryption (crypto) technology was released to the public, commercial applications like online shopping and banking dominated internet use.
More complex applications have since been developed, including social media and the ability to rent computing resources (the Cloud). Billions of users have been onboarded thanks to inventions like smart phones.
Bitcoin shares some similarities with the development of the Internet, a fact that inventors and promoters of cryptocurrency products often cite.
Bitcoin’s first “killer app” was storing and transmitting value permissionlessly (without a bank or other intermediary). But to enable this well known feature, Bitcoin’s design created something else which isn’t much recognized or discussed.
Bitcoin was the first decentralized supercomputer which successfully automated:
Coordinating a group to work on an assigned task
Verifying the correctness of work; and
Settling payment for the work
This invention generalizes much more broadly than specific use cases related to payments, smart contracts, or even building a decentralized financial system.
What if computers could assign tasks, coordinate other computers and humans to perform those tasks, evaluate which solutions are best, and design and implement incentive schemes to reward the highest value contributors?
What if…Bitcoin’s creator had required the nodes to perform useful work rather than calculate hashes of arbitrary data? Arbitrary work was only possible because of the novel financial incentive (the block reward), whereas useful work could have been coordinated on a voluntary basis.
Previous distributed computing projects which achieved scale by persuading volunteers to commit spare compute resources included:
GIMPS (Math project to search for large prime numbers)
SETI@home (analyzing radio signals for signs of extraterrestrial intelligence)
distributed.net (winning a computer security prize by cracking a 56-bit RC5 key)
Although these tasks are more useful than calculating hashes of arbitrary data, they don’t have much social or economic significance. Bitcoin's innovative step was to provide financial incentives, the value of which scaled with the adoption of the network. This attracted resources on a scale unmatched by volunteer projects.
Which other problems could be solved with coordinated but decentralized efforts mediated by crypto-economic incentive schemes?
Well known examples could be grouped under the heading of “Web3 Infrastructure” - providing an open source, decentralized, censorship resistant alternative to services such as cloud file storage and domain name systems; paid for and subsidized with utility tokens. For various reasons which we’ve explored in our coverage for paid subscribers, these efforts have fallen short of their promise.
AI: The Next Frontier?
Training AI models can require access to a large amount of compute resources. The cost of securing access to these resources - usually by renting them from a major cloud computing platform operated by Amazon or Microsoft - has made development of the most cutting edge models cost-prohibitive. This creates centralization of the technology, with attendant undesirable effects.
OpenAI used $3.2 million in compute resources to train their Chat GPT-3 model - using 285,000 CPU cores and 10,000 graphics cards. Fortune estimated their total compute spend for 2022 at around $416 million. Bitcoin issues ~$11.4 billion mining rewards / year to pay for compute (based on 6.25 BTC/block, 10 minute block time, 1 BTC = $35,000)
Instead, what if we could train AI models using a decentralized network, and reward providers of training data and compute resources in native cryptocurrency?
One such project has been trying to achieve this through its decentralized Machine Learning protocol.
What Is Bittensor?
Bittensor is a decentralized AI and machine learning training platform which aims to create a peer-to-peer market for machine intelligence.
There are three types of participants in Bittensor - miners, validators, and users.
Miners run machine learning models. When a user has a request, it is routed to a miner, and the miner produces a response. This would be like typing a query into ChatGPT, except instead of the OpenAI model responding, Bittensor would decide which miner to route the query to.
Validators are intermediaries who validate the quality & accuracy of miner responses.
Miners and validators are paid for their services in TAO tokens. The incentives are designed to lead to a competitive system where the best machine learning models (for each domain, known as a subnet) are selected to receive relevant user queries.
How Does It Work?
Bittensor’s consensus mechanism is designed to reward the most valuable nodes in the network - value is judged according to the marginal contribution of each node towards the overall accuracy of the network.
Nodes and their models can interact and exchange learning data to improve network performance. A scoring process evaluates predictive capabilities of each model and whether it aligns with the consensus from other nodes. The nodes scored as more accurate have a higher probability of being selected to propose a new block and to receive TAO rewards.
This allows improvements to be incorporated at scale while incorrect results can be penalized or discarded.
Above is just a high level overview - note that other projects are working on similar decentralized machine learning models, without involving blockchains or tokens.
Autist Note: Bittensor can also use a Decentralized Mixture of Experts technique, leveraging multiple specialized machine learning models to achieve higher accuracy
Context: AI Goes Open Source
Although OpenAI has high historical expenses to train the most expensive and advanced GPT-based chatbot, founder Sam Altman has suggested in interviews that the company has reached the point of diminishing returns on scale and spend.
Taking a look at the open source space, it seems that low budget projects are catching up to be “80% as good” as OpenAI’s flagship. One example is Stanford Alpaca, an instruction following Large Language Model discussed in more detail in this paper.
We shouldn’t assume that the dominant AI will be created by or controlled by the entity which spends the most on development resources or compute. This also brings into question the assumed need for access to comparable funding.
The open source development model, which grew out of the “Free Software” movement spawned the hugely successful Linux operating system. Linux arguably beat the publicly funded software juggernaut Microsoft in the server market and later became the core of Google’s Android smartphone software. Bitcoin, Ethereum, and the DeFi ecosystem are based on an open source development model. Although Free Software and Linux later attracted some corporate sponsorship, significant work - perhaps the majority - has likely been done on a volunteer basis.
Open source development (naturally decentralized) can be expected to produce AI models which are easy to iterate on and have speed, customizability, and efficiency compared to large siloed AI models. Industry participants have questioned whether it is possible for centralized offerings to recoup their large development and compute costs by monetizing access when small, lean, open source models are available for free.
Focus on smaller (more agile) models, fine-tuning, and personalization is already paying off for the open source community.
TAO Analysis
TAO’s token design is a rough copy of Bitcoin:
21 million units of supply
periodic emission rate ‘halvenings’
TAO emitted to miners and validators performing useful work
fair launch (no team / investor tokens - everyone can “mine” on an equal basis)
TAO has seen significant price appreciation month to date.
Significant backers include Polychain, GSR, and Digital Currency Group - although as the code is public domain and token mining is open to anyone it isn’t clear how these investors expect to be financially rewarded.
Competitors
learning@home (the name is likely a homage to SETI@home) is working on an open source project called Hivemind which leverages a volunteer model instead of a blockchain+token model. Participants co-ordinate using the Distributed Hash Table technology best known for facilitating peer to peer filesharing at Internet scale for the BitTorrent protocol.
Using AI to benchmark AI also isn’t unique to Bittensor:
Preliminary evaluation using GPT-4 as a judge shows Vicuna-13B achieves more than 90%* quality of OpenAI ChatGPT and Google Bard while outperforming other models like LLaMA and Stanford Alpaca in more than 90%* of cases. The cost of training Vicuna-13B is around $300. The code and weights, along with an online demo, are publicly available for non-commercial use.
The chat interface at lmsys.org allows anyone to chat with two anonymous AI models side by side and vote for the best output (or indicate a tie, or an instance where neither model gave a satisfactory answer). This real time user feedback could be used to improve the models. Try it for yourself here!
If you want to compare the quality of the output of miners on the Bittensor network, you can use this free chat interface chat.bitapi.io. There is also a login-gated interface at chat.bittensor.com.
Why Use A Blockchain / Issue A Token?
Although financial incentives would be needed to scale to the level which can compete with OpenAI, Microsoft, or Google, it isn’t clear why a token or native blockchain is needed at this time. Rewards when needed could be paid in other digital tokens, and a blockchain distributed consensus isn’t necessary to determine how to allocate rewards. And. Although we’re not AI experts, there seems to be credible evidence from academic papers and live working projects that “good enough” chatbots can be created and operated by the open source community without substantial budgets.
Is this a field which really needs the fundraising power of cryptoeconomic tokens?
There’s room for disagreement: by bootstrapping the project with a native token with the *potential* to become valuable in proportion to the success of the network, Bittensor can attract smart early adopters to help improve and market the product.
Conclusion
We came across this project and found it interesting. Upon further research we discovered there is quite a bit of commotion in certain parts of Twitter. It’s certainly one possible solution to the AI cost bottleneck, but we’re not convinced it’s the solution. We have found many projects in crypto over the years that have been packaged as Internet technology to the crypto investor.
However, we like to leave our readers with something actionable.
Try it for yourself. This is the only way we can get answers to questions like, “Can we build this on an open source model (like Linux), or do we need massive corporate sponsorship (like Microsoft)?” We are not going to LARP as experts in the AI field. However, we know there will be some readers out there who are experts.
If you’re interested in the field and have the skillset, mine the token. If not, don't buy it. If you don't understand how to do the work to earn the token, you’re unlikely to have a good assessment of what it’s worth.
The Bittensor docs and marketing draw many comparisons between their project and the early days of Bitcoin. If you were interested in Bitcoin in 2009 and 2010, you’d probably run a node and try to mine for it.
If you enjoyed this post, become a paid subscriber!
Until next time..
Disclaimer: None of this is to be deemed legal or financial advice of any kind. These are opinions from an anonymous group of cartoon animals with Wall Street and Software backgrounds.
We now have a full course on crypto that will get you up to speed (Click Here)
Security: Our official views on how to store Crypto correctly (Click Here)
Very interesting. Thanks!
I am interested in mining bittensor but am not an expert. Do you advise just skipping?