Written by Henry Ang, investor at BACKED VC, in collaboration with co-founder Alex Brunicki
Most of us who stuck around after the collapse of FTX will remember this tweet fondly.
It caused such a big stir because crypto was destined to be our “next big thing” but most of us had no idea what was happening in the world of AI back then. GPT-4 launched in March that same year, showcasing its ability to solve complex math problems, write sophisticated code and excel in many standardised tests. This step up in capabilities from pre-schooler (GPT-2) to smart high schooler (GPT-4) took most of us by surprise, which begs the question: how fast is AI growing and how can we be part of the journey as an investor?
I believe we are still laying the foundations today for AI to go mainstream. Huge capital investments have gone into building specialised GPU clusters to provide the compute and energy required to power the latest cutting edge models. We are training them with as much data as we can get our hands on, fine-tuning them, researching new algorithms to get more efficient and effective with every unit of compute. According to Leopold Aschenbrenner, we can expect AI to be fully capable of performing the role of a drop-in worker by 2027 as long as we continue the current pace of progress over the last 5 years.
So what is the role of blockchains in this age of AI acceleration?
At BACKED, we want to take a first principles approach to identify what these opportunities are in the intersection of AI x Blockchain. This means leaning into the core tenets of blockchains:immutability, transparency and self-custody. While our views are constantly shaped by new information, we share some of our internal stream of consciousness on where we are focusing our efforts today.
Before we dive into the opportunities, I want to first share our mental model for AI:
Resources x Research = Foundational Model x Tools = Applications
Resources are the inputs necessary to run these foundational models, such as compute, energy, bandwidth, data, storage etc. Research includes the work required to pre-train, train, post-train these models, “unhobbling” techniques to improve reasoning and inference capabilities, new algorithms to do all of the above more efficiently and effectively with less resources. The combination of resources and research will lay the bedrock for cutting edge models capable of taking on more intensive workloads.
We then bolt on tools to supercharge these foundational models. This could mean teaching a model how to use the internet, using microsoft office to prepare analysis, or more specific to blockchains — interacting with smart contracts by equipping it with a self-custody wallet. Applications will be the gateway for AI to go mainstream, as drop-in workers that can work alongside us.
Where we see opportunities in Crypto x AI
Decentralised Resources
When it comes to resources, we are less excited about decentralised alternatives for compute, bandwidth, storage for AI applications. In isolation, they are less performant than a centralised GPU cluster, and involve more software complexities when it comes to integrating these separate components. Given the huge amount of interest in decentralised compute today, our view is that the greatest challenge with decentralised compute is the risk of data leakage. This can occur when multiple users share the same physical GPU resources, and memory allocation is not properly cleared before reassignment, leading to unauthorized access to sensitive information from previous computations. Coupled with the performance benefits of centralised GPU clusters that have vertical integration across bandwidth, memory, CPUs etc., decentralised alternatives are not competitive when it comes to AI workloads.
While it can be argued that these privacy challenges can be solved and performance parity can be attained over time, we think the gap is more likely to widen over the next few years given the huge amounts of capital investments made today (e.g. xAI with 100,000 Nvidia H100 GPUs, Inflection AI deploying a 22,000 Nvidia H100 GPU cluster, Meta unveiling two AI computing clusters with ~25,000 H100 GPUs each). Therefore when we look at investing into the intersection of Crypto x AI, we do not see any moat for decentralised compute when it comes to AI workloads, and think that they can be more suited for dynamic jobs that prioritise cost-effectiveness, flexibility. We think that this extends to bandwidth and storage as well, insofar as their usefulness in AI workloads.
Data
Data is where we believe the blockchain properties of immutability and transparency make a ton of sense. We are rapidly approaching a point where the amount of new, high-quality data may be insufficient. Llama3 was trained on 15T tokens and the entire internet (after deduplication) consists of ~30T tokens, i.e. we are almost using all available data. This potential “data wall” could become a major limiting factor unless we innovate on new methods, such as synthetic data generation or dramatically improving how efficiently AI models learn from existing data. We think that the next breakthrough in AI will require automated AI research, where AI is able to run experiments and take their experimental results as inputs for further testing. Such a feedback loop has significant room for error and may result in inaccurate analysis if unsupervised.
This is where blockchains provide a base where quality and provenance of data can be accounted for, fostering greater trust with the data outputs. Areas of interest include teams that work on curating high-quality domain specific data sets for LLM training (e.g. codifying reasoning chains of researchers / scientists), zero knowledge infrastructure that ensure fast and cheap computation integrity, auditability and verifiability infrastructure to map out decision making processes based on the data computed.
Consumer / Enterprise Applications
Another area where we think blockchains can make a difference is in the application layer, specifically through the use of self-custody wallets, and smart contracts. We see smart contracts as a tool to integrate AI models into our work flows by automating complex transactions and processes. We think that the likelihood of AI having access to traditional banking rails is unlikely in the near future, given strict AML and KYC processes (how do we even qualify AI agents?), hence opening the door for delegated self custody wallets, that can unlock financial transactions autonomously.
Imagine this: a restaurant that has an inventory management system that keeps track of the supplies, does a quality vs price comparison on where best to buy the items that are running, before using funds from a delegated wallet (with a spending limit of 5000 USDC) to complete the purchase, while also scheduling a suitable delivery date & time according to the expected foot traffic and availability of the working roster.
We are excited about this huge design space for AI applications, and look forward to chatting with teams that have ideas to integrate smart contracts and wallets. We think that no idea is too crazy, and would love to have a chat with any aspiring founder.
Conclusion
Looking forward, AI progress is unlikely to happen linearly which means we have to remain up to date with what is happening on the ground. If our views resonate with you (or not), we would be keen to have a conversation. If you are a team building in our areas of interest, please get in touch!
