arxiv.org/abs/2609.40303 Summary : 🚨 Do AI agents really need complex harnesses? A new study on autonomous ML engineering agents finds that, under equal time budgets and using the same frontier LLM, sophisticated open-source harnesses offer no advantage over a simple coding agent with read, write, and bash access. The takeaway: The model may matter more than the machinery around it. Multi-agent orchestration, retrieval subagents, and elaborate workflows may add complexity without improving performance on current ML engineering benchmarks.

Cognitive computational neuroscience nature.com/articles/s41593-0… 1/ Move beyond brain mapping. Mapping which brain regions activate during a task is not enough to explain cognition. We need computational models that reproduce behavior and explain how neural activity generates perception, learning, decision-making, and action. The goal is to understand the mechanism, not just the location. 2/ Build models that perform cognitive tasks. Allen Newell argued that understanding cognition requires constructing systems that actually perform the functions we want to explain. A theory of decision-making, for example, should predict choices under different conditions—not merely describe them verbally. Simulations help test whether proposed mechanisms work together to produce intelligent behavior. 3/ Integrate three disciplines. Cognitive science explains the computations underlying thought; computational neuroscience investigates how neural circuits implement those computations; AI develops algorithms that perform complex tasks. Each provides different constraints. Combining them could produce models that are simultaneously behaviorally effective, computationally explicit, and biologically grounded. 4/ Neural networks are a bridge between brains and AI. Deep neural networks learn hierarchical representations that can resemble those observed in biological sensory systems. Research has shown that networks trained for object recognition can predict aspects of visual cortical representations. However, predictive similarity does not prove that the brain uses the same algorithms. Models must also be tested against neural and behavioral evidence. 5/ Biological neurons offer more than artificial units. Most conventional artificial neural networks use simplified units that combine inputs and apply nonlinear transformations. Biological neurons exhibit action potentials, dendritic computations, recurrent signaling, oscillations, and complex temporal dynamics. Incorporating these features into task-performing models may reveal computational principles that simpler AI architectures overlook. 6/ Neural decoding is not a complete explanation. Decoding asks what information can be extracted from neural activity. Encoding models predict neural responses; representational similarity analysis compares patterns of activity; connectivity models examine interactions. These methods constrain theories, but they do not necessarily explain how the brain transforms information to solve a task. Mechanistic models must connect representations to computation and behavior. 7/ Bayesian inference offers a theory of perception and reasoning. The brain must interpret noisy sensory information using prior knowledge about the world. Bayesian models formalize this process by combining prior beliefs with incoming evidence. Generative models go further by representing how hidden causes produce observations. They offer a framework for studying uncertainty, causal reasoning, concept formation, and the ability to learn from limited experience. 8/ Learning and control require multiple strategies. Reinforcement learning provides a framework for understanding how actions and feedback shape future behavior. Model-free approaches learn values or policies from experience; model-based approaches use an internal model to evaluate possible outcomes; episodic approaches reuse specific past experiences. Understanding how biological brains combine these strategies could inform more adaptive AI systems. 9/ Combine bottom-up and top-down approaches. Bottom-up neuroscience begins with neurons, circuits, and measured activity. Top-down approaches begin with a cognitive function and ask what computations are required to perform it. Neither is sufficient alone. Strong theories should explain how biological mechanisms support meaningful behavior—and use behavioral requirements to identify which neural mechanisms matter. 10/ The future is an integrated NeuroAI. A major challenge is to connect three levels: computational goals, representations and algorithms, and neurobiological implementation. This requires collaboration across neuroscience, cognitive science, mathematics, and AI, along with shared datasets, models, tasks, and evaluation methods. The ultimate ambition is to build systems that perform intelligent tasks while explaining how brains learn, represent knowledge, and adapt. The big idea: The next leap in AI and neuroscience may come from treating the brain not simply as a source of data, but as a system whose computational principles can be tested through models that connect neural mechanisms to intelligent behavior.

Real-time reinforcement for human-machine interface control cell.com/neuron/fulltext/S08… Summary 1. Real-time reinforcement improves motor performance: Personalized, closed-loop reinforcement feedback enhances force control during human–machine interface tasks. 2. Rapid learning: Fewer than 20 reinforcement trials produce immediate improvements in motor control. 3. Long-term retention: Improvements can persist after training, although retention gains were not observed after short training in chronic stroke patients. 4. Greater benefits with limited sensory feedback: Reinforcement is most effective when visual or somatosensory feedback is reduced. 5. Benefits for stroke patients: Real-time reinforcement improves online force control in chronic stroke patients under conditions of limited visual feedback. 6. Compensates for reduced feedback: Reinforcement helps maintain motor control when sensory information is sparse. 7. Promotes successful action selection: Reinforcement encourages individuals to repeat and exploit motor actions that produce successful outcomes. 8. Translational potential: This strategy could improve rehabilitation and control of human–machine interfaces, particularly for people with motor disabilities.

Programmable design of functional proteins from natural language nature.com/articles/s41586-0…

Tailoring 3D Single-Walled Carbon Nanotubes Anchored to Indium Tin Oxide for Natural Cellular Uptake and Intracellular Sensing pubs.acs.org/nalefd/article/… Objective: Develop a platform for real-time monitoring of intracellular events using vertically aligned single-walled carbon nanotubes (VASWCNTs). Cell model: Used RAW 264.7 mouse macrophages as the eukaryotic cell system. Device fabrication: Cultured cells on indium tin oxide substrates functionalized with VASWCNTs. Direct cellular interface: Demonstrated spontaneous penetration of VASWCNTs into cells without applying external force. Intracellular sensing: Demonstrated detection of methylene blue, a redox-active molecule, within the cellular environment. Key innovation: Established the feasibility of directly interfacing VASWCNTs with living cells for electrochemical sensing. Potential applications: Enables future studies of intracellular redox activity, cellular physiology, and real-time electrochemical monitoring.

Proteome scale drug prioritization from AlphaFold 3 interaction confidence biorxiv.org/content/10.64898… Study objective: Evaluates whether AlphaFold 3 (AF3) protein–ligand interaction confidence scores can support proteome-scale drug repurposing, benchmarking them against BANDOCK+ using the CANDO platform. Comparable performance: Using 2,449 approved drugs and 1,103 matched human proteins, AF3 achieved an nAIA of 10.13 and NDCG of 0.056, slightly outperforming BANDOCK+ (9.93 and 0.054). Best-performing metric: Among four AF3 confidence metrics, global ipTM performed best for downstream drug prioritization, making it the primary metric for subsequent comparisons. Robustness across datasets: AF3 and BANDOCK+ showed broadly similar performance across protein sublibraries selected by experimental structure quality, AF3 model confidence, and random sampling, although BANDOCK+ performed moderately better in the unbiased 200-protein subset. Disease-specific differences: Neither method consistently outperformed the other across disease categories and individual indications. Their relative performance varied depending on the disease, evaluation metric, and protein dataset. Key conclusion and limitations: AF3 confidence scores, particularly ipTM, offer a promising approach for large-scale computational drug repurposing. However, confidence scores are not binding affinities, and comparable ranking performance does not establish interaction validity or clinical efficacy; experimental validation remains essential.

The economics of superintelligence economist.com/leaders/2025/0…

scBaseCount: An AI agent-curated, standardized, auto-updated single-cell data repository Preprint link biorxiv.org/content/10.1101/… cell.com/cell/fulltext/S0092… 1. A massive single-cell gene expression resource: scBaseCount is presented as the largest public repository of uniformly processed 10x Genomics single-cell RNA-seq data, containing over 502 million cells across 27 organisms and 75 tissues. 2. Addresses underutilized public data: The Sequence Read Archive (SRA) contains enormous amounts of sequencing data, but inconsistent metadata makes datasets difficult to discover and compare. scBaseCount aims to unlock this information through automated discovery and standardization. 3. AI agents automate data curation: The researchers developed SRAgent, an AI-powered, hierarchical workflow using large language models (LLMs) and specialized tools to identify datasets, retrieve metadata, and annotate tissues, organisms, and diseases. 4. Substantial improvements in scale and efficiency: SRAgent processed 208,939 SRA experiments and identified 105,343 as 10x Genomics libraries. The study estimates that its automated approach saved more than 500,000 CPU hours compared with brute-force dataset identification. 5. Standardized processing improves comparability: The team developed scRecounter, a scalable Nextflow pipeline that uniformly reprocesses sequencing data using STARsolo. It supports different gene-counting strategies and multi-mapping approaches, reducing analytical inconsistencies across studies. 6. Expanded representation of biological diversity: Compared with resources such as CZ CELLxGENE and the Human Cell Atlas, scBaseCount includes a broader range of organisms, tissues, and experimental contexts. The collection also reveals trends in sequencing technologies, disease studies, and tissue representation. 7. Enables cell-type annotation and cross-species analysis: The researchers applied computational models and reference atlases to annotate cells from humans, mice, zebrafish, and Drosophila. Their analyses identified shared cell populations and generated resources for comparing conserved and divergent cell-type signatures across species. 8. Provides useful non-coding and intronic expression data: The standardized pipeline supports multiple counting strategies, including intronic and exonic reads and non-coding gene expression. These features expand the resource's utility for studying transcriptional dynamics and other aspects of cellular biology. 9. Supports AI foundation models for biology: The study reports that training a cell-embedding model on scBaseCount improves its ability to distinguish diverse cellular phenotypes. The scale and diversity of the resource could help develop models that predict cell states in previously unobserved biological contexts. 10. Establishes a model for autonomous scientific data infrastructure: By combining AI agents, standardized pipelines, and continuous dataset discovery, scBaseCount provides a framework for maintaining large biological databases with less manual curation. The authors have also made SRAgent and scRecounter code available to support further research and development. Key takeaway: scBaseCount illustrates how AI can transform scattered public sequencing datasets into a large, standardized resource for biological discovery and foundation-model development. Its major contribution is not simply the number of cells collected, but the combination of automated curation, uniform processing, cross-species diversity, and AI-ready data.

Advancing NeuroAI through developmental alignment cell.com/neuron/home SUMMARY: 1. Model the development of intelligence, not just its adult state. Current AI models primarily aim to replicate adult-level abilities. NeuroAI should also investigate how intelligence develops from infancy to adulthood to understand the mechanisms that produce sophisticated cognition. 2. Build AI architectures inspired by brain development. The human brain develops hierarchically, with sensory regions maturing before higher-level association areas. AI could adopt this principle by training early layers first and progressively developing higher-level layers, potentially improving robustness and learning efficiency. 3. Use developmental learning as a curriculum. Children encounter information in a natural progression from simple to complex. Training AI with gradually increasing complexity, such as blurry-to-clear images or simple-to-complex visual experiences, may improve generalization, robustness, and data efficiency. 4. Learn from children's limited but meaningful experiences. Unlike AI models trained on enormous datasets, children learn remarkably well from relatively few experiences. Training models on realistic, child-centered sensory data and age-appropriate experiences may help reduce their dependence on massive datasets. 5. Incorporate curiosity-driven learning and exploration. Children actively seek information that is neither too easy nor too difficult to understand. AI models could use predictive learning, information gain, and the ability to select their own experiences to develop more efficient and adaptive learning strategies. 6. Treat developmental milestones as benchmarks for AI. Children's perceptual and cognitive abilities emerge in a characteristic sequence. Evaluating whether AI models acquire similar capacities over training—not merely whether they achieve adult-level performance—could reveal whether their learning mechanisms resemble those of humans. 7. Use NeuroAI to discover the mechanisms of intelligence. By manipulating neural architecture, training experiences, and learning objectives in AI models, researchers can test which factors are necessary or sufficient for particular cognitive abilities. This approach can also help explain brain development, identify mechanisms of dysfunction, and inform future clinical research. 8. Integrate architecture, experience, and learning objectives. These three factors interact throughout development and should not be optimized independently. A developmentally informed approach could produce AI systems that are more biologically plausible, robust, energy-efficient, and capable of generalizing beyond their training data. Central message: To understand and reproduce human intelligence, NeuroAI must study not only what the adult brain can do, but also how the developing brain acquires those abilities. Childhood development offers a powerful blueprint for building more capable and biologically grounded AI systems.

Toward an automated science of the mind cell.com/neuron/fulltext/S08… Summary Artificial intelligence is opening the door to a new era of scientific discovery: the automated science of the mind. Traditionally, cognitive science has relied on a cycle in which researchers design experiments, collect human behavioral data, develop computational models, and refine theories through scientific debate. Although this approach has advanced our understanding of human cognition, it is time-consuming, dependent on human intervention, and often limited by researchers’ existing assumptions about experiments and models. The emerging field of automated cognitive science aims to accelerate this process by integrating AI across four stages of discovery. First, AI can generate and represent diverse experimental paradigms, expanding the range of questions researchers can investigate. Second, cognitive foundation models can simulate human behavior, enabling researchers to explore hypotheses before conducting costly experiments. Third, AI-driven program synthesis can generate interpretable computational models that explain observed behavior and reveal new cognitive mechanisms. Finally, automated systems can close the discovery loop by evaluating competing theories, identifying informative experiments, and proposing the next steps in scientific investigation. However, significant challenges remain. Synthetic behavior may reproduce shortcuts or population-specific biases rather than genuine human cognition. Automatically generated models may be difficult to interpret, sensitive to assumptions, or capable of exploiting evaluation metrics without capturing the underlying mechanisms. Moreover, defining scientific novelty, theoretical importance, and meaningful progress requires careful validation and human oversight. Ultimately, automated cognitive science is not simply about using AI to analyze data faster. It represents a shift toward AI systems that actively generate experiments, simulate behavior, construct theories, and iteratively test new hypotheses. If successful, this approach could transform cognitive science from a largely researcher-driven process into a more scalable, continuous, and data-informed discovery cycle—while keeping empirical evidence, interpretability, and scientific rigor at its core.