The case for specialization: Building scientific AI that thinks like a physicist

Large Language Models have changed how we think, work, and do science, but can they truly reason like scientists? At FirstPrinciples, we’re exploring the limits of AI generalization and the promise of specialization through the development of the AI Physicist.

In recent years, Large Language Models (LLMs) have become a disruptive force in the space of AI, and more broadly to society. They have changed the way we access information, the way we think through a problem, and they have provided us with powerful brainstorming assistants accessible anytime, anywhere.

In particular, it is undeniable that LLMs have impacted how science is being done. Any scientist now has access to a personal assistant that, on the surface, reasons like a scientist while having access to the entirety of the web’s resources. Since their first appearance, we’ve seen LLMs grow significantly — modern LLMs with fewer than 10 billion parameters would be considered small. 

AI hallucinations and the illusion of understanding

However, if we look beneath the surface, it becomes clear that large models today do not quite perform like scientists. LLMs have become famously infamous for being prone to hallucinations. In science, though, hallucinations are not inherently bad, as they can resemble the spark of new ideas. Humans come up with new ideas all the time, most just never evolve into great ones. The problem is that, unlike scientists, LLMs are incapable of disclosing when they hallucinate. They present invention as fact, concealed by fluent prose. 

This is the result of two seemingly unrelated behaviours. First, LLMs are being trained on huge datasets that span large, unrelated topics and domains. Upon inference, there is always the chance that the model outputs tokens from a totally unrelated domain. Hence the hallucination. Second, LLMs are trained to be assistants, i.e. pleasing to humans, which incentivizes them to sound confident, always having an answer for everything. The combination of these two behaviours creates models that hallucinate, that are unable to disclose the hallucinations, and even try to conceal them to sound useful.

However, those traits are unfavourable to a scientist as science is built upon transparency. The great Richard Feynman had this wonderful passage in The Pleasure of Finding Things Out:

“We have found of paramount importance that in order to make progress we must recognize the ignorance and leave room for doubt. Scientific knowledge is a body of statements of varying degrees of certainty — some most unsure, some nearly sure, none absolutely certain. Now, we scientists are used to this, and we take it for granted that it is perfectly consistent to be unsure — that it is possible to live and not know.”

At FirstPrinciples, we are building an AI Physicist, so the question is: what makes a physicist? Probably, being a physicist means knowing facts about physics, and probably, it also means being creative at manipulating physics concepts. And what else? The recent trends in Artificial General Intelligence (AGI) is that intelligence will emerge from ever larger models that are exposed to ever larger datasets. However, FirstPrinciples is not concerned with general intelligence. 

Generalization vs. specialization: Introducing artificial intelligence for physics

Instead, we are concerned with “APhyI”, or Artificial Physics Intelligence. We do not know what defines physics intelligence, but we do know what it looks like. The only example of true intelligence we know of today is ourselves. When we look at the academic world, we notice that researchers naturally self-organize around specialization. Experts in different domains collaborate to generate new, original ideas. The same pattern holds true for society in general; specialization is a key feature of our success as a species. Even more broadly, through natural selection, nature evolves specialization as a winning strategy leading to species being adapted to exploit their own niche.

It seems that specialization is a fundamental emergent property of nature, and the question is what is so special about it? To some extent, the answer is that specialization is a strategy of optimization of resources when resources are finite. Just like in nature, AI models are still finite no matter how large they become. Therefore specialization should still hold as a valid winning strategy.

Typically, in today’s paradigm of agent orchestration, specialization is achieved by clever prompt engineering to create multiple agents on top of a single LLM. This approach, though successful, is flawed in a crucial way — each agent will inherit the biases of the underlying LLM, which includes the tendency to hallucinate and to be an overly helpful people pleaser. Instead, at FirstPrinciples, we create specialized models from the ground up that act natively as experts. 

However, the other side of the coin is that, while some specialization seems to be a winning strategy in academic research, it is also true that overspecialization can be a detriment. Instead, intelligence naturally emerges from exposure to a diverse variety of knowledge. Our goal, then, is to find the balance — the sweet spot between generalization and specialization. In our framework, we envision combining specialization with generalization by developing a set of models that are specialized in subdomains, but are orchestrated in an agentic system to facilitate artificial debates. 

Building an AI physicist that scientists can trust

What this gives us is a system that we can trust as a scientist. How we achieve this vision is what we are trying to figure out at FirstPrinciples. Part of this research has to do with intelligent data curation. Is it relevant for a physicist to read the entirety of the academic literature, including social sciences, chemistry, history, etc? There are certainly ideas from different domains that can inform physics in unexpected ways, but probably not all ideas. 

For example, we can imagine that reading philosophy is good for a physicist, as both fields require manipulating logical statements. Listening to music might also help; its rhythmic patterns and rigorous structure mirror the rigor of science and scientific thought. But does a physicist need to read the entirety of Reddit, with the obvious biases that would be introduced into the model? Probably not. 

The other part of the answer lies in developing more focused algorithms for training models. This research centres on imbuing desired qualities, such as curiosity, transparency, honesty, and rigorous reasoning, among others. Some of the techniques that come to mind here are reinforcement learning and adversarial networks. 

The right tool for the right job

But being a good physicist is not just about generating good ideas and being able to reason rigorously. It also means being able to solve differential equations, interpret plots, write papers, and more. For those reasons, part of our journey towards specialization also involves developing highly specialized tools. These serve as dedicated resources to which the model has access in order to perform tasks beyond the scope of language models alone.

At FirstPrinciples, we pursue specialization because, throughout the history of tool making, progress has always come from building the right tool for the right job. Our job is physics, and our tool is built for it.