Accepting “the bitter lesson” and embracing the brain’s complexity

Large-scale AI offers transformative potential for neuroscience, but realizing this vision hinges on overcoming significant data challenges—ranging from access and standardization to infrastructure and incentives. A fundamental issue is that academic reward systems prioritize novel discoveries over the meticulous work of curating and sharing high-quality datasets. Without clear incentives, much of neuroscience’s vast and complex data remain siloed, underutilized or inconsistently formatted, making it difficult to integrate them across studies. The lack of widely adopted data-sharing standards further complicates efforts to build foundation AI models, which rely on uniting diverse datasets from multiple sources.

To fully leverage AI in neuroscience, several key issues must be addressed. First, data accessibility requires dedicated outreach, training and resources to ensure that available datasets are not just published but also easily discoverable and usable by the broader community. Without structured efforts to promote and maintain shared repositories, valuable data risks being overlooked. Second, harmonizing diverse datasets is critical for scaling up AI models. Standardizing data formats and metadata across institutions reduces the barriers to integration and allows researchers to build more robust and generalizable AI systems. Global efforts to develop and implement these standards are essential to prevent fragmentation and inefficiencies in large-scale projects. Third, the complexity and volume of neuroscience data demand specialized expertise in high-performance computing, multimodal analysis and advanced visualization techniques. Managing and processing these large datasets requires sustained investment in infrastructure and personnel capable of handling the technical challenges associated with AI-driven research.

Beyond technical solutions, structural and cultural shifts are needed to support long-term data stewardship. Funding agencies and institutions should prioritize resources for sustained data curation, provide career pathways for experts in scientific data management and establish policies that recognize and reward data-sharing contributions. (To read more about the role that research software engineers can play in academic neuroscience, see “Neuroscience needs a career path for software engineers.”) Promoting a collaborative approach to infrastructure and standardization will be key to unlocking the full potential of AI in neuroscience.

Understanding the brain is one of society’s most pressing challenges, particularly as populations age and extending brain health becomes increasingly urgent. The brain’s complexity, however, poses significant hurdles. Advances in neuroscience have enabled us to record activity from more neurons than ever before and directly link it to behavior, but we urgently need tools that can help us make sense of these rapidly accumulating data.

The possibility of the research community collaborating on building foundation models for neuroscience is a thrilling prospect. Such models could transform the field, with applications ranging from translating insights across species and individuals to unlocking the full potential of brain-machine interfaces and personalized medicine. Beyond this, foundation models could significantly accelerate experimental research. For instance, they could be used to predict the outcomes of experiments, enabling researchers to identify the most informative experiments before conducting them. This would save time and streamline the discovery process.

But significant challenges remain. Do we have enough high-quality, diverse data to train such models? Can we successfully integrate the multitude of neural data modalities in a way that enhances model generalizability? Addressing these questions is critical to building models that are robust and useful across different contexts.

Moreover, neuroscience models are not just tools for prediction; they also enable us to draw key mechanistic insights. Beyond prediction, a foundation model could even be used to uncover fundamental principles of brain function. By combining interpretability techniques on foundation models with reduced mechanistic models that capture the essence of observed phenomena, we could bridge the gap between data-driven predictions and mechanistic understanding. With this dual capability—prediction and mechanistic insight—foundation models could become cornerstones in neuroscience for both research and applications.

One of the most profound experiences in my life as a neuroscientist was following receptive field mapping experiments by Kevan Martin in the basement of the Institute of Neuroinformatics in Zurich. As he moved bright stripes in front of the animal’s eyes, I could hear the activity of a neuron—it was active only when the stripe was in the right place, the neuron’s receptive field. Out of these kinds of experiments came a model of how the visual system works; a set of neurons that extracts lines, a subsequent set that would extract curves and, after a few steps, a set of neurons that would recognize objects. This idea of the brain as a collection of simple specialized neurons grouped into brain areas was deeply influential for the field and for me, giving rise to both the leading theories in neuroscience and, incidentally, the development of the field of artificial intelligence (AI).

Over the years, however, we found that rather than specialized neurons cleanly extracting human-understandable features, neurons are entangled in a web of multi-purpose, context-dependent processing. The most wonderful vision of simplicity in neuronal coding, the bedrock of the logic of neuroscience, has died.

This growing realization of complexity has forced us to reconsider how we make sense of neural data—enter AI. AI is surprisingly good at describing high-dimensional, heterogeneous neural data and solving engineering problems, such as decoding speech in people who have lost the ability to talk. In the search for understanding, we now turn to AI—not just as a tool, but as a framework.

But is it a framework for understanding? Unlike classical mechanistic models, AI models often lack interpretable structure, offering predictive power but little conceptual clarity. Some hope to salvage simplicity by using AI models to extract interpretable models, an approach known as “distillation.” The problem is that such simple explanations are generally not good at describing real-world data. And this brings us to the inevitable conclusion: Simplicity is dead. As we move toward AI as a guiding principle, we must ask: What if the brain’s solutions are not merely complex but fundamentally incomprehensible?

Training an AI model is like sculpting: We begin with a raw slab of material and chip away at it to reveal the underlying truths of neuroscience. For large-scale models with many degrees of freedom, this slab is enormous—like a multidimensional block of marble. Experiments, observations, datasets and theories act as instructions to carve away the inconsistent bits; the remaining marble is defined by the constraints. As advocated by Dyer and Richards, training large-scale AI models for neuroscience will require unprecedented investment and collaboration. Neuroscience surely needs such commitment to advance, but what keeps me awake at night is the ill-posed nature of these problems: Do we have enough constraints to uncover the hidden beauty shaped by evolution?

Modern AI models are optimized for their predictive performance. If the goal of modeling is practical—such as designing a mechanical limb—then a clever model that generalizes and predicts observations suffices. In this case, having 10,000 equally well-performing but fundamentally different solutions doesn’t matter. But if the goal is to uncover the mechanisms of brain function—creating an interpretable representation of its biological implementation—we may need far more constraints even if the predictive performance is high.

Recent AI successes in science are often used to bypass a complex unsolved problem with a black box for the sake of downstream applied science and engineering. But fundamental science, specifically neuroscience in this case, demands more. Achieving deeper understanding requires not only embracing the complexity of the brain but also collecting new, highly informative data that can further sculpt our models. It may even necessitate a fundamentally different class of AI models. Instead of relying on the “marble” of current architectures—powerful but superficial—we may need the “granite” slab of models that are harder to sculpt but capable of expressing the recurrent dynamical mechanisms we seek to discover.

Dyer and Richards’ call to action deserves a strong endorsement and careful consideration. Their argument for embracing the brain’s inherent complexity resonates deeply with our understanding of the brain as a profoundly interactional system. The challenge before us is determining how best to implement this vision.

I propose that we amplify their embrace of complexity even further. As Dyer and Richards cogently argue, we are dealing with a non-stationary, non-linear, stochastic, high-dimensional biological system. But this raises critical considerations regarding their proposal to leverage AI as the path forward.

If we accept that the brain operates in a fundamentally context-dependent manner—perhaps even radically so—then capturing meaningful principles requires sampling from an extraordinarily vast space of possible states and computations. This presents both conceptual and practical challenges for any large-scale modeling approach.

A primary consideration is that current laboratory paradigms inadequately capture the rich dynamics observed in naturalistic environments. Though Dyer and Richards acknowledge this, I would push further: We need data from animals freely engaging in species-typical social interactions, both cooperative and competitive, across multiple ecological contexts. Moreover, the environments themselves must be sufficiently complex to allow meaningful bidirectional animal-environment interactions that mirror real-world dynamics.

A second critical consideration concerns the interpretation of multi-species datasets. In a nonstationary, nonlinear, stochastic, high-dimensional system, we should expect significant species differences in both organization and function. The architectural variations across rodent and primate brains at micro-, meso- and macro-scales likely produce meaningful functional divergences. This does not diminish the value of comparative work but rather demands we explicitly account for these differences in our theoretical frameworks and modeling approaches.

If neuroscience embraces Dyer and Richards’ vision while thoughtfully addressing these challenges, we may be entering an extraordinarily productive era for understanding brain function and behavior.

Leveraging AI for understanding neural function is rapidly becoming an integral part of computational and systems neuroscience. Gone are the days when experimentalists stared lovingly at each hard-earned single neuron trace. Neural data has become unwieldy and complex, and modeling these data demands the development of computational approaches that scale effectively. Large-scale AI may just provide the solution to many researchers’ problems. But the transition to large-scale AI in neuroscience is not without challenges, as highlighted by Dyer and Richards, including data accessibility and standardization for neuroscience-specific models, as well as data engineering and computational resources to achieve these. Additionally, in a field as vast and diverse as neuroscience—with upwards of 25,000 people attending the Society for Neuroscience annual meeting in a typical year—achieving consensus on the key questions best suited for large-scale AI remains a challenge.

Neuroscience has long relied on tailored models for a handful of experiments at a time (often in paired experimental-computational collaborations). The advent of team science has started to shift computation into a central role, bridging collaborations between larger groups of experimentalists in a hub-and-spoke model to address broader questions more efficiently. Large-scale AI has the potential to accelerate this process, but not all research areas will benefit equally. Many questions at the frontier of neuroscience rely on novel recording techniques or unique experimental paradigms, which inherently produce small, noisy datasets that may not be well suited for pre-trained AI models.

Unlike text-based AI, where sheer data volume alone drives model success, neuroscience may benefit from structural biological knowledge. To model neural systems effectively in “small-data” regimes, AI models have the potential to integrate anatomical and physiological constraints, such as known connectivity between brain regions and principles of neural dynamics. A hybrid approach—combining large-scale AI with neuroscientific knowledge—may be crucial for achieving interpretable and biologically meaningful results. Lastly, these models could operate in a closed loop with experimentalists, guiding neuromodulation and validation experiments that advance both scientific discovery and health-care applications.

We stand at the threshold of a new era in AI in which large-scale models are becoming more accessible and deeply integrated into everyday life. Fostering a diversity of computational approaches for neuroscience, including investing in the use of large-scale AI, is crucial to ensure that we do not overlook the next major neuroscientific breakthrough.

Neurons are traditionally interrogated in the context of discrete tasks, such as responses to visual stimuli, in which the choices and stimuli are limited in number. This tight control of stimulus and response enables interpretation of the neural recordings in the task context. But neurons can participate in many tasks in many different ways, so interpretations derived from a single task can be misleading. We can now simultaneously record from tens of thousands of neurons in many brain areas, but neuroscientists are still using the same old single-task paradigms. One solution is to train animals on many different tasks. But training a monkey, for example, takes weeks to months for each task. Another solution is to expand the complexity of the task over longer time intervals, bringing it closer to natural behaviors. In the real world, natural behaviors are primarily self-generated and interactive. This is especially true for social behaviors. Studying such self-generated, continuous behaviors is much more difficult than studying tightly constrained, reflexive tasks. For example, thinking is a self-generated behavior.

What if a large language model (LLM) were trained on massive brain recordings during natural conditions and accompanying behavior, including body and eye tracking, video, sound and other modalities? LLMs are self-supervised and can be trained by predicting missing data segments across data streams. This would not be scientifically useful from the traditional experimental perspective but would revolutionize how brains are studied. But once the sensory input stops, the activity in an LLM terminates; there is no self-generated activity. Another limitation is the amount of needed data, not all of which can be collected from an individual person.

A good place to explore alternative approaches is to start with a small brain. Flies have about 100,000 neurons that can be recorded optically as light flashes from fluorescent dyes sensitive to neural signals while monitoring behavior. Ralph Greenspan and I collected near-whole-brain recordings from fruit flies during self-generated walking. Gerald Pao at the Okinawa Institute of Science and Technology recently used a new method based on dynamical systems theory to download these recordings into a generative model. He used convergent cross mapping (CCM), which extracts causal relationships between recorded neurons and behavior. This method produces a reduced graphical model that captures the low-dimensional brain trajectories that control behaviors.

Only a few minutes of simultaneous neural and behavioral recordings were needed to train a CCM walking fly model. The model’s spontaneous behaviors and neural activity were indistinguishable from those observed in the walking fly. Because much fewer data are needed, many behaviors from an individual brain can be downloaded into a single generative model. We are applying CCM to human functional MRI data collected during a virtual-reality driving game (see figure). Will the generative model be able to drive and deliver packages in the virtual town? Will we eventually be able to eavesdrop on self-generated human thinking? Hello world!

As Sutton describes, large “general” models in AI, such as LLMs, are trained on trillions of words—orders of magnitude more than a human—before they are proficient at tasks such as writing. This suggests that AI training procedures and learning rules are quite different from human learning. If we can figure out some of the learning rules and paradigms in the brain, we may be able to reduce the training time and size of these models. More importantly, these insights may help us understand how learning is disrupted in neurological disorders. Figuring out neural learning rules is hard, though. The brain is complex, and neural activity is high-dimensional; responses to visual stimuli in the mouse and primate cortex span hundreds of dimensions, for example. These many dimensions will only be well-sampled in datasets that are acquired over long time periods and across many neurons. Further, to infer learning rules, we need to collect these datasets across the course of learning. From these datasets, we can then fit models across learning and investigate how the model weights change with learning. If we fit a very large model, though, as proposed by Dyer and Richards, there will be many weights that can change, making it challenging to identify learning rules. A large black-box model is less likely to provide insight into how a neural computation is performed. These medium-to-large models may be necessary as starting points, as they are often easier to fit and perform well, but then the size of these models can be reduced through weight pruning, sparsity penalties or other strategies. Smaller models will be easier to visualize and may correspond better to brain physiology, thereby generating testable hypotheses about learning rules and neural circuit architecture. Our ultimate goal in neuroscience is to understand how the brain functions, not to build the best predictive model, and thus we may need to explore many strategies beyond building large foundation-like models.

Dyer and Richards make a compelling case that neuroscience must embrace the brain’s full complexity through large-scale approaches. Their call for shared infrastructure, compute resources and professional engineering teams to handle vast neural datasets is timely and important. But their emphasis on using AI to analyze neural data leaves a critical gap: the fundamental goal of understanding how brains compute.

The authors correctly identify that neural activity is context-dependent, nonlinear and high-dimensional. Traditional reductionist approaches, though valuable for basic mechanisms, cannot capture the emergent computations that arise from networks of neurons. But invoking AI as a solution risks treating it as a black box—an analysis tool that finds patterns useful in applications, but without providing insight into the underlying principles of neural computation.

A more productive approach is to use these artificial neural networks as interpretable models of brain function. By training these networks on neural data and to perform the same tasks as biological circuits, we can create simplified yet functional analogs of neural systems. The key next step: reverse engineering these trained networks to understand how they solve computational problems. This process of model-building and systematic analysis can generate testable hypotheses about how real brains compute.

The authors invoke Sutton’s “bitter lesson”—that scaled-up general methods tend to outperform cleverly engineered solutions in AI. This principle certainly applies to modeling neural data: Larger networks trained on more data will yield better predictions than will hand-crafted models. But understanding how neural circuits compute requires a fundamentally different approach. As of yet, we have no optimization metric for scientific insight. The path to understanding runs through human creativity and rigorous analysis of our large-scale models, not just scaling them up. We need theoretical frameworks for understanding network dynamics, methods for identifying computational motifs, and ways to connect artificial network solutions to biological mechanisms.