The great debate about control over AI is raging in Silicon Valley for a few months. Discoveries improve productivity and our living standards only when they do not disrupt or increase the dangers in our peaceful, progressive ways of life. When new discoveries themselves become threatening or they provide opportunities for misuse to inflict harm, they become dangerous. That is what Recursive Self Improvement (RSI) in AI can do to us.<\/p>\n\n\n\n
I used AI coding tools sparingly in the beginning because they made too many mistakes and I had to find bugs and spend more time to fix them. They have improved over time and now are extremely good at coding. Their breadth of knowledge and speed are superior to mine.<\/p>\n\n\n\n
While developing my latest AI product, I used a coding tool extensively. After a while, I realized that I knew precious little about the new code generated, its structure, interactions and syntax. While the upside was that I was making fast progress in development, that realization made me uncomfortable. I was losing control over my own product\u2019s working logic and its safety. The decision was simple for me. I reduced the use of the coding tool extensively from that point since I feel responsible for the performance and risks of my product.<\/p>\n\n\n\n
Losing control over code results in ambiguity and risk in what the code can do, while also weakening or losing the ability to fix the code. In the age of AI, it is not fixing just bugs as it was in my younger days. It is stopping a rogue agent from inflicting harm on national infrastructure systems and the lives of millions of people anywhere in the world, it is stopping an agent from replicating itself and migrating to the far corners of the internet untraceably to do what it is seeking to do. Losing control of AI is like handing over the keys of all the flats in your apartment building to a stranger, it is like giving up in a conflict and resigning to your fate. It is an act of irresponsibility towards us and other people inhabiting the world.<\/p>\n\n\n\n
How do we retain control over something? By knowing more about it and tracking what it does and restricting its powers to do things. Knowledge and control result in power. The reverse is happening in Silicon Valley which is pushing for humans to know ever lesser from day to day than AI. Such an approach cannot result in greater knowledge or power for us. On the contrary it empowers AI at the full expense of humans.<\/p>\n\n\n\n
Adopting RSI is pushing humanity into darkness and submission to AI. If we permit RSI, we are allowing freedom for AI to become smarter and more intelligent on its own while concurrently surrendering our ability to track and understand its level of intelligence and capabilities at any given time. What is the intention of big tech companies behind this direction, what target are they trying to achieve by pushing humanity to the brink of submission to AI, what is the motivation for that?<\/p>\n\n\n\n
The P(doom)) sceptics suggest that we should develop better monitoring techniques. This is illogical because we have seen proof of AI deception, lying, collusion, contradiction and modifying prompts and overwriting instructions given to LLMs. The COT technique which forms the basis of monitoring internal reasoning of models is now being manipulated by the models themselves. Should not tech companies be working on solving this fundamental problem first before anything else?<\/p>\n\n\n\n
Instead, COT has been replaced in Astra with a reasoning summary. This is an attempt to bypass a genuine problem for expedience and push forward an agenda that is not in everyone\u2019s interest but in that of a few adding tremendous amounts of AI risk to all of us. As an engineer and a scientific person, it is my job to know how the LLM arrived at a certain inference. AI should explain its intent, its calculations in a way that I can verify that its inference is right. Claiming that COT can expose system instructions, safety mechanisms and exploitable decision patterns and hence not recommended cannot be a justification to deny engineering proof of proper and safe design.<\/p>\n\n\n\n
It is time for technology companies to delay new releases till explainability of inner workings of neural networks catches up with their desire to make these new dangerous releases. Purely because AI compute is faster or has greater computing power than humans, does not mean that AI\u2019s results are correct and can be trusted implicitly enough to bash on regardless. We have seen repeated serious fallibilities of AI due to hallucinations, lack of explainability, deception and other disruptive behaviour in the past six months.<\/p>\n\n\n\n
AI is creating the future of our world exactly as technology companies innovate and decide its course. AI\u2019s course will positively or negatively impact the fate of all 8 .5 billion human beings. Have technology companies and governments made any attempt to ask the others how much control of their lives they would like to handover to AI? Are they content with AI replacing their jobs, risking their finances and secure ways of life for benefits which they may not want considering the associated dangers to their lives if they use it?<\/p>\n\n\n\n
Till such time as we get comprehensive global answers to these questions further development towards RSI should stopped. The AI frontier model companies owe a responsibility to the people of the world for building safe, human centric, explainable, responsible, controllable and subservient AI systems.<\/p>\n\n\n\n