About 20 years ago, larry abbott The principles and foundations of the combined field of theoretical neuroscience were explained from the following perspectives. neuron title”The rise of theoretical neuroscience” One thought that stuck with me was that rigorous models, expressed in equations rather than words, could be formulated, investigated, and often rejected at a pace that no experimental program could match. Abbott argued that the equations “enforce models to be accurate, complete, and self-consistent,” and that this precision and speed, combined, can act as an intelligent filter to screen out possibilities before expensive experiments begin. Abbott drew a picture of a modeler rapidly iterating through phases of model exploration and rejection.
In reality, this doesn’t happen at all. Converting word models to equations, converting equations to code, and integrating models and data all represent engineering work that is invisible to the scientific story. These costs create bottlenecks in determining what models are explored, adopted, and published, ultimately slowing down the pace of theory exploration, which Abbott noted is a core benefit of theory.
Agenttic coding frameworks—systems that allow you to write, debug, and integrate code through natural language interactions—are rapidly changing this landscape. Not by replacing the human work of coming up with models and weighing their merits, but by eliminating the engineering scaffolding that previously limited which theories could be fully explored, which models could be tested, and how quickly they could be rejected. Theoretical neuroscientists can now specify models in conversation, outline hypotheses, explain data, outline inference steps, and produce working code in days rather than months. This is exactly what the field wants.
How will this reshape the landscape of theoretical and computational neuroscience? Below I outline four main directions in which I think seismic shifts will occur. To find out how other experts view this change, I asked six neuroscientists to weigh in.
F
First, agent models remove technical hurdles, making models that previously required collaboration with quantitative neuroscientists more accessible to experimenters. In previous days, I would spend months adapting established but “beginner-level and above” models for collaborations, wrangling data, implementing model variants, and debugging convergence. Most experimental labs were unable to implement these models because the technical barriers were just beyond their limits. Instead, I relied on my collaborators to guide me. This is often a mutually beneficial process, but one that does not require modelers to live at the forefront of technological innovation. Agent coding frameworks eliminate that friction. An experimenter can ask an artificial intelligence (AI) agent, “I have this data. Can I fit a hierarchical generalized linear model to this?” Within an afternoon, create a suitable model and accompanying instructions. Reducing this small but real technical barrier puts these models directly into the hands of experimental researchers and parallelizes processes that were previously bottlenecks.
Second, theorists will finally be able to accelerate their research work at the pace that Abbott envisioned. Abbott argued that theorists can “formulate, explore, and often reject models” faster than experimentalists can test them. However, that advantage was always tempered by the time required to convert word models to equations and equations to code. In the former era, potential ideas were left in notebooks simply because the cost of implementation wasn’t worth the return, assuming most ideas were dead ends. Alternatively, this friction has led researchers to build simple models that can be quickly analyzed and coded. They are less likely to develop and explore complex models that require more engineering overhead. Sometimes this simplicity yielded deeper insights, but sometimes it failed to capture important biological realities. In the age of agent frameworks, theorists can sketch bold new theoretical ideas (modified recurrent neural networks, new learning rules, different population coding schemes) and immediately implement them, test them on synthetic data, and refine them. The theoretical frontier expands significantly, allowing exploration of more and more complex models from seed to fruit. Many ideas will fail, and that’s a good thing. Abbott’s vision of rapidly exploring and rejecting models is finally possible.
Third, and most disorienting, is what agent systems can detect on their own. Evolutionary algorithms guided by large-scale language models (LLMs). Google’s AlphaEvolveallows you to explore model space in ways that human intuition doesn’t naturally arrive at, and on a scale beyond the reach of even the most caffeinated graduate student. We may not think of or have infinite bandwidth to try out certain combinations of nonlinearities or certain connections between neural populations, even if the precursors of those combinations are sitting on a shelf waiting to be combined. By sheer violence AI systems can explore these combinationsunder construction Millions of potential models It includes long-established ideas and unexplored new directions.
History provides insight into the potential pitfalls of rapid technology-driven scientific discovery. spectrometerFor example, the 1860s introduced a wave of new hypothetical chemical elements, including both actual discoveries, such as helium, and phantoms, such as: coronium—a hypothetical element rooted in real-life observations that required decades of theoretical research in physics and chemistry to resolve. With AI-guided search, some discoveries become scientifically meaningful. Others will be artifacts of the search process. We need to develop new skills to understand models that we have not directly built and continue to develop first-principles-based models for separating helium and coronium.
Fourth, agent coding frameworks enable more mathematically sophisticated models. Some of the most highly regarded and mathematically rigorous models in computational neuroscience, such as mean-field theory from statistical physics, employ advanced techniques that require years of intensive training to master. As we move to “,”prove the atmosphere”, the outlook for equation-based AI-assisted theoretical research is changing. Systems that can work with symbolic mathematics and executable code simultaneously open up new research directions with mathematically sophisticated models by supporting researchers who are very accustomed to computer terminals but not blackboards.
It is worth recognizing that this acceleration and expansion of theoretical possibilities comes with real risks, especially for trainees. For theorists, it’s not just friction that transforms ideas into equations and then into code. It is a form of disciplined thinking. As theorists build models one by one, they develop an intimacy with the models that forces understanding. Slow, painstaking construction makes modelers aware of gaps in their own reasoning, and vague intuitions confront reality. Automating this struggle risks creating a world in which theorists are unable to deeply understand their models. For trainees, this is especially dangerous. Advisors face the heavy responsibility of determining which struggles are worth preserving to protect the conditions in which learning takes place.
But there are still more serious risks. It is precisely this struggle that often evokes real theoretical insights. By spending nights staring at unexplained behavior in models, theorists make conceptual leaps that lead to real discoveries. If we lose the battle, we may lose not only understanding but also inspiration itself. The field, while prolific, is in danger of becoming shallow, generating models faster than we can generate insights. as Tim Requas recently discussed About AI-assisted writing, “If struggling to articulate an idea is part of the process of understanding it, tools that avoid that struggle can degrade the quality of thinking that is most important to actual discovery.”
Agenttic coding frameworks help realize Abbott’s vision. That means the freedom to explore and reject ideas quickly, the removal of technical barriers that reduce model adoption, and the ability for more researchers to explore complex, mathematically rigorous models. The north star of living through these times is remembering that most ideas are bound to fail and that to cultivate and maintain this discretion, we may need to periodically step away from the AI and return to the proven resources of our own minds.
