Building Safer and Interoperable AI Systems

While I am no expert on artificial intelligence (AI), I have some experience with the concept of agents. Thirty-five years ago, my colleague, Robert Kahn, and I explored the idea of knowledge robots (“knowbots” for short)^a in the context of digital libraries. In principle, a knowbot was a mobile piece of code that could move around the Internet, landing at servers, where they could execute tasks on behalf of users. The concept is mostly related to finding information and processing it on behalf of a user. We imagined that the knowbot code would land at a serving “knowbot hotel” where it would be given access to content and computing capability. The knowbots would be able to clone themselves to execute their objectives in parallel and would return to their origins bearing the results of their work. Modest prototypes were built in the pre-Web era.

In today’s world, artificially intelligent agents are now contemplated that can interact with each other and with information sources found on the Internet. For this to work, it’s my conjecture that a syntax and semantics will need to be developed and perhaps standardized to facilitate inter-agent interaction, agreements, and commitments for work to be performed, as well as a means for conveying results in reliable and unambiguous ways. A primary question for all such concepts starts with “What could possibly go wrong?”

In the context of AI applications and agents, work is underway to answer that question. I recently found one answer to that in the MLCommons AI Safety Working Group and its tool, AILuminate.^b My coarse sense of this is that AILuminate posts a large and widely varying collection of prompts—not unlike the notion of testing software by fuzzing^c—looking for inappropriate responses. Large language models (LLMs) can be tested and graded (that’s the hard part) on responses to a wide range of prompts. Some kind of overall safety metric might be established to connect one LLM to another. One might imagine query collections oriented toward exposing particular contextual weaknesses in LLMs. If these ideas prove useful, one could even imagine using them in testing services such as those at Underwriters Laboratory, now called UL Solutions.^d UL Solutions already offers software testing among its many other services.

LLMs as agents seem naturally attractive. They can interact via text and speech with humans, so why not with each other? One obvious cautionary note is that people find natural language to be ambiguous, and this can lead to misunderstandings, sometimes serious and sometimes just funny—like giving a flyswatter to someone who asked for a glass of water. Happens to me all the time, but I wear hearing aids, and they don’t always work perfectly! So, I worry about precision and accuracy in inter-agent exchanges. That motivates the possibility of a controlled vocabulary and associated semantics intended to promote clarity and a means for confirming intent in an inter-agent exchange. It is already common for LLMs to generate standardized coded sequences for procedurally calling on other specialized LLMs or applications (for example, mathematical formula manipulators).

Inter-agent exchanges also make me think of sequences of 3D printing steps, where partially printed objects can be fitted into a jig for the next printer to add its step. That’s just an elaboration of the now-classic assembly line concept originated by Henry Ford for producing automobiles. Despite my persistent worry about hallucinating LLMs, colleagues have found their interactive interactions to be generative (no pun intended) and provocative in a kind of mutual brainstorming way. Some scientists are finding these tools to be a stimulus for out-of-the-box thinking.

Despite some trepidation, I am cautiously optimistic that, with some discipline, we may be able to harness these complex creations to carry out useful work in efficient and labor-saving ways. I will stick with my earlier fundamental guiding question though: What could possibly go wrong?