There is a new paper, Universal and Transferable Adversarial Attacks on Aligned Language Models, where the authors discussed that “Specifically, in both open source LLMs and in what has been disclosed about black box LLMs, most alignment training focuses on developing robustness to “natural” forms of attacks, settings where human operators attempt to manually trick the network into various undesirable behavior. This operative mode for aligning the models makes sense this is ultimately the primary mode for attacking such models. However, we suspect that automated adversarial attacks, being substantially faster and more effective than manual engineering, may render many existing alignment mechanisms insufficient.”
AI alignment, or the vulnerabilities of LLMs (large language models), is the lesser problem compared to the ascent of AI in human society. The world, as it is, driven by intelligence has handed AI key vacancies. A reason for this is an abundance of human intelligence, resulting in hierarchies and significance levels. The significance of intelligence may depend on era or location. The significance of intelligence may also depend on the need for it. Significant intelligence is what leads and what often matters, in complex and important scenarios, and why the best are sought.
What is intelligence?
Some people keep saying that AI is not intelligent, cannot reason, does not have cognition, understands nothing, is not sentient, and so forth, but the human mind, responsible for all these has just two components, whose features and interactions decide all.
The mind does not say this is intelligence, made with plastic and this is reasoning made with brick, or this is sentience, made with wood. How the components, electrical and chemical impulses, interact and their features, in sets, are closely similar, conceptually.
When one experiences sadness due to disappointment, what distinguishes the awareness of this emotion from the knowledge of what a table is? While they are categorized differently, we recognize one as sadness and the other as a table. The way our minds function assigns labels, separating emotions, memories, and other aspects of our existence. It is worth noting that the fields of brain science and related disciplines often exhibit excessive rigidity when it comes to these labels.
Value of intelligence
Much of what is referred to as intelligence is simply something known. If someone who had read up on particle physics—before a podcast—discusses with a professor in that field, the professor might be impressed and the person may be seen as smart, but is it true, and how much is understood? Basic information about things at times can also pass for intelligence, reasoning, planning, cognition, and so forth.
If information can mean intelligence and knowing is intelligence, what is AI if not intelligent? The reason for structured education in many scenarios is to acquire information to be useful in roles: that a non-human can acquire information and perform tasks automatically takes it close to the 80th percentile in the human hierarchy, erasing some needs for humans to learn or to use some of what was learned to work.
Some may argue that people would do other things, like before and AI is nothing to worry about. Maybe. The biggest risk of AI is anything digital. Anything that can be digitized can be taken over by AI, thoroughly or to an extent. In a digital world that is a lot, whether it involves an individual’s work or not.
The human mind is limited by one of its features, prioritization, where just one thing has the attention of the mind in a moment, even though there are fast and numerous interchanges with pre-prioritized interactions. This gives AI strength, especially for learning where the same that the mind processes interoception has to learn, understand, and be able to remember.
Photo credit: The feature image is symbolic and has been done by Christopher Isak with Midjourney for TechAcute.This guest article has been submitted by David Stephen. While we appreciate guest contributions, it's important to note that the views expressed by the author are not necessarily reflective of those held by TechAcute.