Assessing the Promises and Perils of Scaled Language Models

Discussions regarding Artificial Intelligence are often surrounded by a lot of uncertainty. Large language models like ChatGPT, Bard, and Claude are well known, but hardly really understood by the typical user. They’re used as tools for writing essays, planning lessons, solving equations, or telling a story. But the fear of AI “taking over” the world is always present in the back of people’s minds because, quite frankly, it’s not impossible. In the latest Dwarkesh Patel podcast episode, Patel interviews Anthropic CEO Dario Amodei about the future of AI, the risks, the rules, and the possibilities that can come from the controlled and uncontrolled growth of LLMs.

Throughout the podcast episode, Amodei explains that most of the potential pros and cons of training an LLM derive from how much a model is scaled. Scaling refers to the ability of AI algorithms, data, models, and architecture to operate at a needed size, speed, and complexity. So, according to Amodei, if a model is continually scaled at a rate in which two to three years from now, it could be dangerous, certain precautionary measures need to be put in place. Currently, LLMs understand language, but in order to accurately replicate the behaviors of human beings, years of training and testing are required. Intelligence, however, isn’t “a spectrum.” There are a number of different kinds of skills and expertise, and there are some things that current models can’t do today, like mirror a generally well-educated person in casual conversation, while there are other things that current models excel at, like biology. It all depends on the kinds of questions and topics one presents to a model, and it depends on the kind of model one is using. Anthropic’s Claude, for example, has a particular set of skills that perhaps Google’s Bard doesn’t, and vice versa. When talking about the intelligence level of Claude in contrast to a human being, Amodei refers to the current version of Claude as “an intern in most areas,” but that if the laws of scaling continue at this trajectory, he “suspect[s] that the rising tide will lift all the boats.” 

So, what does this mean for businesses? And most importantly, what does this mean for society in the near future? Well, there isn’t really a clear-cut answer for either of these questions. Amodei suspects that when AI passes a certain threshold, it’ll be “weird and different than we expect.” More and more money is going into the scaling and research of AI models, which means that as models improve, they will become more lucrative. This does raise questions, however, about the ethics and safety of scaling LLMs.

There are many key factors that go into the improvement of LLMs, but the mechanistic interpretability of models is what both excites and worries Amodei. Mechanistic interpretability is the study of taking trained language models and analyzing their data to reverse engineer the algorithms learned by the model. This means that, after reaching a certain point of intelligence, the mechanistic interpretability of a model can alert data trainers of possible problems that arise from a powerful AI, but not necessarily provide specific solutions to those problems. As models reach a higher scale, there is a higher probability of “very serious danger.” Amodei clarifies that as of right now, models don’t have that level of autonomy yet, but it isn’t that far ahead. Unless there is government intervention that slows down scaling, or a bigger emphasis on AI alignment, the possibility of “serious danger” is only a few years away. Alignment is when researchers aim to steer AI systems towards humans’ intended goals and ethical principles. Although meant to keep language models in check, Amodei warns that handing a powerful LLM to a government can lead to misuse of AI alignment, but he also knows that without some sort of government regulation, AI will quickly get out of hand. 

Powerful models are an inevitability. If such a model were to, in Amodei’s words, “wreak havoc,” there would be no way to stop it. A model with that much power may not exist now, it may not exist in three years, but it will exist at some point. What worries Amodei the most is how little grasp society seems to have on the simplest form of AI models now, and what that means for human beings when models scale past the threshold that separates powerful models from simple ones. Cybersecurity needs to be at the forefront of research and development, otherwise misuse, bioterrorism, and security breaches will become bigger problems.

All of this sounds… scary. One hopes that listening to a podcast with one of the creators of Claude would put our mind at ease, but unfortunately that is not the reality. Amodei does emphasize how simple language models are right now, and that the “threat” of losing control rests on whether or not regulation is enforced within the next few years. Of course, as AI progresses, there will be more improvements in the collection and analysis of data, the creation of new technologies, and greater efficiency in many industries. The effects of scaled language models are already being felt today, so one can imagine how much progress will be made in the next ten years. The uncertainty of AI is undoubtedly its most worrying aspect,but in the meantime, it’s best to learn as much as possible about AI systems, because the future is no longer just an idea, it’s happening right now.