The Year Ahead in AI

Safe, Accurate, and Efficient

Dec 20, 2023

It’s been an astonishing year for the AI sector in 2023. I frankly never expected we would see anyone unleash high-risk models to the public in interactive form, much less trained on most of the world’s copyrighted data. Until November of 2022, high-risk self-generating models like LLMs, open-ended, and genetic algorithms were limited to research labs until they could be proven safe. We are a long way from the point of LLMs having the technical potential to be proven safe.

Making LLM chatbots available to the public was premature, representing the most reckless commercialization of high-risk technology I’ve observed in my lifetime (I was born in 1959). It wasn’t just ‘move fast and break things’, which can be helpful in some scenarios, but rather unleashed several types of identifiable catastrophic risks. LLM companies and their enablers failed in upholding the first principle, which transcends technologies and industries: EAI should have governance, ethics, and security built-in from inception.

EAI should have governance, ethics, and security built-in from inception.

To state the obvious, AI principles mean little unless they are executable by the AI system architecture deployed (view our 15 EAI principles, rational and implication of each, and 3-part video interview with Dr. Robert Neilson on KYield’s board).

In an interview with the FT, (viewed on 12/20/2023), the CEO of Accenture, Julie Sweet, suggests most companies aren’t ready for generative AI due to lack of data governance and safety controls:

“Most companies do not have mature data capabilities and if you can’t use your data, you can’t use AI”. — Julie Sweet, CEO of Accenture.

Given the perverse incentives favoring brute force computing with massive waste, it’s unlikely cloud vendors or incumbent chip companies will lead on safe and responsible AI systems, but we’ll likely see much more sober focus in 2024 on safe and responsible AI nonetheless. I expect this trend will increase due in no small part to regulation from the EU AI act, but also demand from very large enterprise customers:

“I need five nines [99.999%] of correctness (accuracy)…I cannot have a hallucination that says: ‘Oh yeah, put widget A connected to widget B’ — and it blows up.” — Craig Martell, CDAI at the DoD (my selection for quote of the year in AI).

I recently published the first in a two-part series of a working paper on safe, responsible, and efficient AI systems. The first portion is focused on the many costs and risks of LLMs. The second expected around the first of the year will focus primarily on why such costs and risks are largely unnecessary, and provide examples of components within an architecture like our KOS as example of a safe, responsible and efficient AI system.

My hope is the AI hype of 2023 dominated by LLM chatbots will rapidly mature in 2014 to a focus on safety, accuracy, and efficiency.