TL;DR: In this post I discuss a recently proposed on-chip compute governance mechanism, a delay based location verification mechanism. Then I demonstrate feasibility and applicability of the mechanism through a few exercises and identify a potential issue with frequent false positives. Finally I propose a potential solution for the identified issue.

Technical governance approaches will be critical for mitigating the risks posed by potentially misaligned AI systems. Even if the alignment problem is solved, technical governance methods will still likely be important for some time to limit misuse or the consequences of AI malfunctions.

Recent AI progress has largely been driven by increases in the amount of computing power used to train new models. Currently the compute used to train notable AI systems doubles every six months. Hence, governing compute could be an effective way to govern AI.

*Figure 1: The amount of compute used to train notable AI models has doubled roughly every six months since 2010. From* *Epoch*

On-chip Mechanism for Location Verification

A subset of compute governance, “on-chip governance mechanisms” could allow a regulator to perform verification or impose restrictions. In this article we explore location verification mechanisms specifically, delay based verification.

Speed of light based verification

In a recent report CNAS (Centre for a New American Security) proposed a location verification mechanism that relies on the hard limit of the speed of light (300 km/ms) and the lower bound on latency from existing communications infrastructure(I believe the maximum available is 200km/ms for pure fiber optic).

This approach has the following main design parameters.

The round trip distance (RTD) between subject chip and export control boundary (B) determines the maximum allowable round trip time (RTT_max), as calculated based on speed of light, for a verification handshake between the trusted server(S) and the subject chip(C).
- RTT_max = RTD(C–>B)/300
Practical latency per distance (L_p) of communication infrastructure determines the maximum possible distance between trusted server and subject chip. The lowest bound on L_p is inverse of the speed of light in fibre optik (1ms/200km).
- RTD_max(S–>C) = RTT_max/ L_p
Substituting for RTT_max
- RTD_max(S–>C) = RTD(C–>B)/ (L_p*300)

Figure 2: *Assuming speed of light, The distance from C to B determines t*

...

Posts

Wikitag Contributions

Comments