AI researchers explore how historical neuron modeling informs modern science

Lessons from a squid nerve cell

In 1951, physiologists Andrew Huxley and Alan Hodgkin sought to understand the 'action potential' of a squid nerve cell. These are the electrical pulses that allow brains to communicate, which form the basis for everything from perception to intelligence. Because the university computer was out of action, Huxley had to solve their equations using a hand-cranked mechanical calculator for over three weeks.

The scientists modeled the cell membrane as an electrical circuit and discovered that they needed to raise one variable to the fourth power to make the model fit their data. At the time, they did not know exactly why this specific power was necessary, though they hypothesized it related to multiple charged particles moving in place. Decades later, researchers discovered a potassium channel protein with four identical pieces, confirming that their 1951 model had correctly predicted a physical reality.

As Anthropic reported, this history highlights a vital point about scientific modeling: simply finding a number that fits the data is not enough. The value of the Hodgkin-Huxley model lies in the fact that simpler versions did not work. To achieve true understanding, a model must be able to make claims about the real-world system and its similar counterparts.

Moving beyond single best fits

When scientists constrain a model, they tune adjustable parameters to match real-world measurements. However, many different combinations of parameters can often produce the same output, much like many different food orders can add up to the same total bill. If a system only provides one successful combination, it hides all other plausible settings that might lead to different predictions.

To solve this, researchers use a statistical method called Bayesian inference, which provides a probability distribution of all working numbers. This method is often too computationally heavy for complex systems, such as networks of neurons or colliding galaxies. In these cases, calculating the likelihood of every possible measurement becomes practically impossible for standard computers.

A newer approach called simulation-based inference (SBI) overcomes this by using sheer computation to reverse-engineer models. The system picks random parameters, runs the model to see the output, and repeats this millions of times. A neural network is then trained on these results to return the range of parameters that could have produced a specific observation, showing where the data is clear and where it is ambiguous.

AI agents as model builders

While machines have traditionally excelled at constraining existing models, AI is now beginning to assist in building the models themselves. The space of possible scientific models is essentially infinite, and human scientists might miss viable options based on their own limited experience. New AI systems can now take a natural-language description of a system and write the code for candidate models.

These LLM-based agents can enter a loop where they propose a model, run it, constrain the parameters, and compare the output to real-world data. If the model fails to replicate the data accurately, the agent revises the form of the equations. This allows the AI to explore a much wider variety of mathematical structures than a human might propose in a single research cycle.

The goal is to treat model building the same way as model constraining by asking for a distribution of plausible models. If every surviving model proposed by the AI shares a specific feature, it suggests the data favors that feature. If the survivors disagree, it points researchers toward the specific areas where more experimental data is needed to break the tie.

The path to scientific insight

The ultimate goal of these AI tools is to move toward what philosophers call 'intelligible' theories. A theory is intelligible if it allows a person to work out its consequences without having to run a simulation. For example, the Hodgkin-Huxley equations allowed scientists to reason out why nerve cells have a voltage threshold and why pulses have a consistent shape.

Currently, these AI systems have mostly been tested on rediscovering models that scientists already know to be true. Evidence that they can solve problems with no known answers is still sparse, but the technology is improving as newer LLMs perform better at research tasks. The researchers believe that as these systems evolve, they will become more capable of identifying the underlying rules of nature.

By providing a distribution of models rather than a single answer, AI can help scientists understand the limitations of their own data. This approach ensures that the machine is not just a calculator providing numbers, but a tool for generating insight. It helps researchers distinguish between models that are merely convenient and those that accurately reflect the physical world.