1/ Verifiable inference enables proving the correct model was used and that inputs/outputs were not tampered with. Why is this useful, who's working on it, and what does the future hold? Summary below 🧵 ✍️ Written by @hammyx_ equilibrium.co/writing/state…

2/ Inference (calling an existing AI model with a prompt) has two main variations depending on where the compute happens: Local and Outsourced Outsourced inference can further be split into trusted and untrusted providers, depending on whether they have a reputation at stake.

3/ Verifiable inference enables proving integrity during the inference process, more specifically: 1. The correct model and weights were used 2. Inputs and outputs were not tampered with This brings two main benefits - increased traceability/auditability and enhanced trust:

4/ Increased traceability is particularly useful for compliance and legal purposes, while enhanced trust enables outsourcing inference to untrusted parties. However, the relative importance of these benefits differs depending on where the compute takes place:

5/ In a blockchain context, verifiable inference is mainly relevant for: - Increasing expressivity of onchain applications - Enforcing onchain agents and agentic networks - Enabling decentralized inference networks

6/ Approaches to verifiable inference include ZKP, TEE, hashing, OPML, and random sampling. These offer different tradeoffs wrt cost, performance, strength of guarantees, usability... The post covers each approach, including main use cases, benefits, drawbacks & open problems

7/ We've also included a detailed industry overview with project summaries and links to further reading. As we can see, there is a growing number of teams working on different parts of the problem space. Several of them are also collaborating with each other 🙌

8/ Future Directions 🧭 We expect demand for verifiable inference to grow by several orders of magnitude from today, primarily due to three main drivers: 1. Trust in big tech weakening 2. Broader integration of AI into critical workflows 3. Expanding onchain use cases