Table of Contents

Overview

Technology is rapidly changing, ranging from LLMs devouring programming, viable humanoid robots, and others… Here are some of my predictions for 2031.

Anonymous Agents (21/04/2026)

The main issue with current agentic systems is that you’re reliant on end providers (OpenAI, Anthropic, etc.) to protect the data you send to them by: not sharing it with others, not training or analysing it, and not having the data be shared with the government, etc.

However, do we have to share our data with the end providers to use frontier systems? The answer is quickly becoming no.

Fully Homomorphic Encryption (FHE) allows us to encrypt our data using lattice-based cryptography (which also happens to be post-quantum, meaning there are no known efficient attacks using quantum or classical computers) but also allow us to perform arbitrary computations on the data.

This means that you would be able to locally encrypt your prompt, send it to the provider (who has the benefit of more powerful hardware, a better trained model, etc.), who then processes the response and sends it back to you still encrypted, and only then would you decrypt it. This means that your request (and its response) are never revealed as plaintext to the provider, it remains as ciphertext during the entire journey (true roundtrip encryption).

Okay this sounds amazing, how might this actually be done? The main issue with FHE currently is that the processing time is roughly 10,000x slower than comparable methods. This means that if you normally achieve 100 tokens / second, then an FHE-based system would achieve 100 seconds per token, yes thats right, 1 min 40 seconds per generated output token.

However there has been a program running by DARPA called DARPA DPRIVE which tasked companies, such as Intel, to produce an ASIC device which accelerates the compute primitives required for FHE to be within 10x the speed of native operations for FHE. Intel produced a prototype device called Intel Heracles which has achieved this as of March 2026. This means that by 2031 these prototype devices could realistically be deployed to cloud providers such as AWS, who also so happen to have a partnership with Anthropic, for completely anonymous LLM usage. AWS already provides secure enclaves which serve a similar purpose, although they do not have the absolute guarantee of FHE.

When combined with other technology such as the Tor Network, these anonymous LLMs allow one to have a completely secure and anonymous method of accessing frontier intelligence. You can imagine a network of compute providers taking payment for compute credits via cryptocurrency and providing entirely anonymous frontier open-source (or perhaps closed source?) models.