A New Frontier: AI Meets Crypto
beyond the hype: verifiable attestations, cryptographic prompt protection, and more
This post was co-authored by Casey Caruso.
AI and crypto have both independently experienced waves of massive investment and enthusiasm within the tech ecosystem. Each has its own set of distinct categories to explore. However, as the boundaries between these technologies blur, a new frontier emerges: the fusion of AI and crypto.
Our reference to 'crypto' extends beyond merely cryptocurrencies. It encompasses the foundational cryptography that supports these digital currencies and a broader spectrum of associated technologies.
We foresee the convergence of AI and crypto evolving into a distinct, mature tech sector. Here we outline these categories, with some illustrative examples of companies we believe fit into them. Disclaimer: this sector is actively developing and the companies we've mentioned don't encompass the entire market of builders in every vertical. We encourage founders and researchers to share their insights and suggest other companies that the community should know about.
1) Verifiable Attestations
Thesis: Verifiability of inputs, models, outputs, evaluations, and user identities of models may become necessary for certain actors (e.g. enterprises) to feel confident in adopting foundation models. It’s also possible verification becomes part of AI culture where the next “certified organic” is “verified inference”, where consumers demand to know the some details of what’s now a black box to them.
Related companies:
Identity verification example: Google’s MedPalm2 which requires users of the model to have a medical background to access the model.
Data verification with Open Origin
Verifiable AI audits with OpenMined
ZK Conduit and EZKL
2) Cryptographic Prompt Protection
Thesis: Human-AI interaction is currently grounded in prompts, with many consumers and companies hesitant to adopt foundation models due to risk of sensitive information being leaked. A new set of intervention techniques at prompt time (i.e, data loss prevention) grounded in cryptography may be used to mitigate such risks for enterprises.
Related companies:
3) DAO-esque Governance and Model Alignment
Thesis: AI requires intricate guidelines that align to the context where the models are deployed. There’s a growing emphasis on developing processes for open governance driven by community feedback, quite similar to the long-term vision the crypto world had for DAOs. It’s possible AI governance and model alignment resemble DAOs in that they enable diverse groups to openly discuss issues, deliberate solutions, and transparently decide on standards.
Related readings:
4) Financialized AI
Thesis: Crypto offers powerful incentives that could be harnessed to govern, align, fine-tune and drive data contributions in the AI revolution. Financializing these components through crypto-enabled markets could accelerate progress.
4.1) Tokenized Models: Enabling a market around AI models to train, buy, sell, trade models in an efficient manner.
Related companies:
4.2) Tokenized Fine-Tuning: Incentivize people to transform raw, unstructured high-quality data into a structured format, making it usable for supervised fine-tuning or retrieval-augmented generation, or RLHF.
Related companies:
4.3) Tokenized Data Cooperatives and Marketplaces: Facilitating the exchange and management of data in a decentralized manner.
Related companies:
5) Decentralized Compute and Storage
Thesis: Many are reluctant to use hosted models like GPT-4 due to privacy concerns. On-premise deployments are gaining popularity as an alternative that allows more control over data. Looking ahead, on-premise and decentralized deployments and inference may become another popular architecture.
Related companies:
Reka: Exploring on-prem solutions.
Cohere: Transitioned to a primary focus on B2B with on-prem deployments
If you are a startup founder working in one of these categories and wanna chat, reach out on Twitter @caseykcaruso and @shreyjaineth.