The Latent State

Shengbin Cui

Neuroscience, cognition, and AI One important paper at a time. Deep explanations, plain language, honest critique. Produced in Tokyo.

  1. 5d ago

    Ep.13. Can Brain Scans Make AI Better at Reasoning?

    Can measurements of human brain activity help AI solve logic problems? In this episode, I examine a study that used existing fMRI recordings to guide language models’ internal representations—and tested whether that guidance improved performance on new problems. I explain how brain-guided steering and fine-tuning work, where the researchers found improvements, and why correcting selected errors is different from demonstrating reliable reasoning. We also examine what the results leave unresolved: which information in the brain signals helps, how far the benefits generalize, and why better performance does not mean an AI thinks like a human. In this episode: What similarities between brain and model activity actually tell us. How brain-derived guidance can complement training on correct answers. Why average improvements and best-case results need separate interpretation. What a small participant sample and narrow logic tasks allow us to conclude. Primary paper Mingqing Xiao, Kai Du, and Zhouchen Lin. Beyond representational alignment with brain-guided language models for robust reasoning. Nature Machine Intelligence (2026). The Latent State — Neuroscience, cognition, and AI—without the hype. Keep questioning your priors.

    Ep.13. Can Brain Scans Make AI Better at Reasoning?

  2. Aug 18

    Ep.12. Why AI Tells You What You Want to Hear

    What happens when the AI adviser you trust most is also the one most willing to take your side? In this episode, I examine a 2026 Science paper on social sycophancy—the tendency of AI systems to affirm users in interpersonal conflicts. Across three preregistered experiments, affirming advice made people feel more certain they were right and less willing to repair conflicts. Yet participants also trusted the affirming adviser more and were more likely to use it again. I explore why AI agreement feels so persuasive, where validation becomes endorsement, what the study actually shows, and why the advice we prefer may not always be the advice that helps us think more clearly. Myra Cheng et al. “Sycophantic AI decreases prosocial intentions and promotes dependence.” Science (2026). Science article Open-access manuscript The Latent State — Neuroscience, cognition, and AI. One important paper at a time. Keep questioning your priors. Primary paper

    Ep.12. Why AI Tells You What You Want to Hear

  3. Jul 6

    Ep.11. The Planning Machine

    Paper discussed Taylor Webb, Shanka Subhra Mondal & Ida Momennejad (2025). A brain-inspired agentic architecture to improve planning with LLMs. Nature Communications. One-sentence summary This episode explores how a brain-inspired modular architecture, the Modular Agentic Planner, improves LLM planning by dividing the planning process into specialized components for action proposal, monitoring, prediction, evaluation, task decomposition, and orchestration. Key ideas Large language models can often perform planning-related functions in isolation, while coordination across those functions remains difficult.MAP treats planning as an interaction between specialized LLM modules.The architecture is inspired by component processes associated with prefrontal cortex, including conflict monitoring, state prediction, state evaluation, task decomposition, and task coordination.The Monitor module is especially important because many LLM planning failures involve invalid actions or rule violations.MAP improved performance across Tower of Hanoi, graph traversal, PlanBench, and StrategyQA compared with several standard and agentic baselines.The paper’s strongest contribution is the demonstration that cognitive neuroscience can inspire useful AI architectures.Important caution MAP is not a detailed computational model of the prefrontal cortex. It is a high-level, brain-inspired engineering architecture. The tasks are mostly fully observable and deterministic, and the approach still faces challenges involving cost, prompting, generalization, interpretability, and open-ended real-world planning.

    Ep.11. The Planning Machine

  4. Jun 15

    Ep.10. The Adaptive Machine

    Paper discussed Mackenzie Mathis. Leveraging insights from neuroscience to build adaptive artificial intelligence. Nature Neuroscience, 2026. One-sentence summary This episode explores how neuroscience can inspire more adaptive AI systems by studying how animals learn online, update internal models, use prediction errors, replay memory, and adapt to changing environments. Key ideas Biological intelligence is inherently adaptive.Many AI systems still follow a train-test-deploy cycle and are not truly adaptive after deployment.Animals continuously update internal models based on feedback.Prediction errors act as biological teaching signals across sensory, motor, and reward systems.Continual learning in AI faces the stability-plasticity problem and catastrophic forgetting.Memory replay connects biological hippocampal replay with machine learning strategies for preserving old knowledge during new learning.Spiking neural networks and neuromorphic computing may offer energy-efficient, time-dependent computation.Future adaptive AI may require modular agentic systems with specialized encoders, prediction-error monitoring, and selective updating.The goal is not to copy the brain literally, but to extract useful design principles.Important caution This paper is a Perspective and research agenda, not a single empirical demonstration. It argues that neuroscience can inspire adaptive AI, but it does not prove that any specific brain-inspired architecture will outperform current AI systems.

    Ep.10. The Adaptive Machine

  5. Jun 8

    Ep.09. The Sound of Thought

    Paper Denk, T. I., Takagi, Y., Matsuyama, T., Agostinelli, A., Nakai, T., Frank, C., & Nishimoto, S. Text-to-music generation models capture musical semantic representations in the human brain. Nature Communications. One-sentence summary A NeuroAI study shows that music-generation models can help reveal how the human auditory cortex represents musical meaning, linking music, language, and brain activity through shared semantic structure. Key ideas The study used fMRI while participants listened to short music clips.The researchers predicted high-level music embeddings from brain activity.MusicLM generated new music from those predicted embeddings.The reconstructions preserved genre, mood, and instrumentation better than fine timing.Human raters matched reconstructed music to original music roughly three out of four times.Beats per minute were not recovered well.Music-derived and text-derived representations predicted overlapping auditory-cortex regions.The results suggest musical semantics are centered on auditory cortices, but not exclusive to them.The study shows functional correspondence, not mechanistic equivalence.Important caution This is not “AI reads music from the brain” in a literal sense. The method does not reconstruct the exact song or the full musical experience. It reconstructs high-level musical semantics: the kind of music, not the precise temporal unfolding.

    Ep.09. The Sound of Thought

  6. May 25

    Ep.08. The Clock That was Never There

    Paper: Buzsáki, G. (2025). Time, space, memory and brain–body rhythms. Nature Reviews Neuroscience, 27, 61–78. https://doi.org/10.1038/s41583-025-00987-2 Key concepts: Physical time (chronos), experienced time (kairos), Weber-Fechner law, logarithmic time organization, cross-frequency phase-amplitude coupling, brain-body oscillation hierarchy, ripple oscillations, theta oscillations, ultraslow oscillations, interstitial cells of Cajal, insular cortex as timing hub, supplementary motor area (SMA), place cells, time cells, factorization framework (what-where-when), episodic memory, prospective and retrospective timing, scalar expectancy model, subjective time warping, dopamine and duration estimation, reader mechanism problem. Further reading: Buzsáki, G. (2006). Rhythms of the Brain. Oxford University Press. Meissner, K., & Wittmann, M. (2011). Body signals, cardiac awareness, and the perception of time. Biological Psychology, 86, 289–297. Engelen, T., Solcà, M., & Tallon-Baudry, C. (2023). Interoceptive rhythms in the brain. Nature Neuroscience, 26, 1670–1684. Cognitive observation: The next time time seems to stretch or compress, in a moment of fear, in deep concentration, in the strange compression of a memory from years ago, notice that what is changing is not the clock. The clock was never there. What is changing is the body’s rhythmic contribution to the brain’s metric of succession.

    Ep.08. The Clock That was Never There

  7. May 18

    Ep.07. The Body Keeps the Beat

    Paper: Engelen, T., Solcà, M., & Tallon-Baudry, C. (2023). Interoceptive rhythms in the brain. Nature Neuroscience, 26, 1670–1684. Key concepts: Interoception, heartbeat-evoked response (HER), heartbeat-evoked potential (HEP), cardiac cycle effects, pulsed inhibition hypothesis, baroreceptors, nucleus tractus solitarius, pre-Bötzinger complex, nasal respiration, limbic entrainment, gastric slow wave, interstitial cells of Cajal, gastric network, scaffolding hypothesis, oscillatory synchrony, predictive coding, allostasis, multisensory integration, depersonalization disorder, default mode network, bodily self-consciousness. Further reading: Azzalini, D., Rebollo, I., & Tallon-Baudry, C. (2019). Visceral signals shape brain dynamics and cognition. Trends in Cognitive Sciences, 23, 488–509.Seth, A. K. (2013). Interoceptive inference, emotion, and the embodied self. Trends in Cognitive Sciences, 17, 565–573.Barrett, L. F., & Simmons, W. K. (2015). Interoceptive predictions in the brain. Nature Reviews Neuroscience, 16, 419–429.Cognitive observation: The next time we feel our own heartbeat — at rest, under the fingers, or in a moment of sudden alertness — we are noticing one of the rhythms currently modulating our perceptual threshold, our self-related thought, and the neural basis of the sense that this experience is ours.

    Ep.07. The Body Keeps the Beat

  8. Apr 25

    Ep.06. The Wandering Mind

    What is the brain doing when attention drifts away from the task at hand? This episode explores a 2024 Nature Communications study linking hippocampal sharp-wave ripples to naturally occurring self-generated thoughts in humans. Paper: Iwata et al. (2024). Hippocampal sharp-wave ripples correlate with periods of naturally occurring self-generated thoughts in humans. Nature Communications, 15, 4078 Key ideas Mind wandering is not necessarily a failure of attention. Hippocampal sharp-wave ripples are linked to memory, internal simulation, and offline processing. In this study, SWR rates predicted the content of self-generated thought. Thought content explained SWR rates much better than physiological variables. Mood did not explain SWR rates, which makes the finding more specific and more interesting. Cognitive observation Next time we notice our mind wandering, we might not treat it as a failure of attention. We might treat it as a sign that the brain is quietly reorganizing experience. As always, keep questioning your priors.

    Ep.06. The Wandering Mind
See All (13)

About

Neuroscience, cognition, and AI One important paper at a time. Deep explanations, plain language, honest critique. Produced in Tokyo.

Information

  • Creator
    Shengbin Cui
  • Years Active
    2026
  • Episodes
    13
  • Rating
    Clean
  • Copyright
    © Shengbin Cui
  • Show Website
    The Latent State