Future Trends in Cognitive Computing and Neuroscience

Fields move when something new becomes measurable, buildable, or affordable. Three frontiers are open at once right now, and each one is a different answer to the same question: what should a cognitive system be made of?

Decoding language from motor cortex

A brain–computer interface — a system that reads neural activity and converts it into a control signal — used to be judged on whether a participant could move a cursor. The target now is unconstrained speech. Electrodes in the part of motor cortex that controls the mouth, tongue, and larynx pick up the attempt to speak; a recurrent network converts that activity into a stream of phoneme probabilities; a language model turns the stream into words.

Virtual brain twins

A virtual brain twin is a whole-brain simulation built from one individual's scans. Structural imaging supplies the regions and the white-matter connections between them; each region gets a mathematical model of its local population dynamics; then the free parameters are estimated by asking which settings would have produced that person's recorded activity. The result is a model you can perturb — stimulate a region, remove one, change excitability — and watch what follows.

The furthest-developed case is the virtual epileptic patient, used to locate the network responsible for seizures before surgery. It ran as a clinical trial across hundreds of patients with drug-resistant epilepsy, and the approach is now being extended to psychiatric conditions, where the target is predicting response to a drug rather than a surgical margin.

Wetware: computing with living neurons

The third frontier abandons silicon. Human stem cells are coaxed into brain organoids — spheroids of neurons and supporting cells a few hundred micrometres across — which are then placed on a multi-electrode array, a chip whose electrodes both record spikes and deliver stimulation. Nutrient medium flows underneath continuously; cameras and environmental sensors watch for trouble. Experiments are scripted in Python and run remotely.

• Organoids survive on the array for weeks to months, against hours in early attempts.
• One platform has run over a thousand organoids and accumulated more than 18 terabytes of recordings.
• Reward is delivered chemically: dopamine is released from a light-sensitive molecular cage by a pulse of ultraviolet light.
• Access is remote and shared, so groups without a wet lab can run electrophysiology.

The reason anyone bothers is the same reason this page keeps returning to: a human brain runs on roughly twenty watts. But the honest summary of what these systems have computed is thin. The best-known result had a culture of cortical cells improve at a simulated game of Pong, and even the researchers building the platforms have publicly warned that terms like "organoid intelligence" are running well ahead of the evidence. Firing patterns drift day to day, so a training procedure that worked yesterday may not transfer.

What these three trends share, and where they split

Hardware decides which algorithms are practical, and the algorithms that are practical decide which theories of cognition get tested. The current constraint is energy, and the current response is to stop computing things that have not changed.

Neuromorphic hardware reaches billion-neuron scale

A conventional neural network evaluates every unit on every pass, in lockstep, and shuttles weights back and forth from memory. Neuromorphic hardware does neither. Units are spiking — they stay silent until an accumulated charge crosses a threshold, then emit a one-bit pulse — memory sits next to compute, and nothing happens where nothing has changed.

Why the savings are conditional

Event-driven efficiency is only realised when the input is genuinely sparse in time — continuous audio, event cameras that report changed pixels instead of whole frames, robot control loops. Dense matrix multiplication, which is what today's large language models are, maps poorly onto this hardware. None of these systems has shipped as a commercial product; deployments are research partnerships, and the binding constraint is software rather than silicon.

The arithmetic behind the claim is simple enough to check. If a layer fans out to a fixed number of downstream units, a dense pass costs one operation per unit per step regardless of activity, while an event-driven pass costs one only per spike. Sparsity is the whole story.

import numpy as np
rng = np.random.default_rng(0)

def cost(spikes, fanout=128):
    dense = spikes.size * fanout
    event = int(spikes.sum()) * fanout
    return dense, event

for rate in (0.5, 0.05, 0.005):
    spikes = rng.random((1000, 256)) < rate
    dense, event = cost(spikes)
    print(f'rate {rate:<6} dense {dense:>9} event {event:>8} saving {dense / max(event, 1):.0f}x')

The memory problem in artificial agents

The other innovation reshaping cognitive systems is not hardware at all. A transformer-based model stores what it knows in fixed weights, set during training and frozen afterwards. Training on something new disturbs what was already there — catastrophic forgetting — so deployed systems do not learn from the work they do. Everything they retain across sessions has to be written down outside the model and fed back in.

None of the three is solved, and the framing matters more than the leaderboard: all of them are attempts to give an artificial system the two properties biological cognition has by default — memory that accumulates and a model of the world that supports imagining an action before taking it. Notice that reinforcement learning already borrowed the first of these from the hippocampus in the form of replay buffers; the borrowing is not finished.

Where hardware and cognition meet, and where they don't

Forecasting badly is easy: pick the most exciting demonstration and extend the line. Forecasting usefully means separating three questions that get conflated — is the underlying claim supported, is anyone using it for real work, and what observation would show it is going nowhere. A direction with no answer to the third question is not a research programme.

Energy is the organising constraint

Read the whole page through this number and the trends line up. Neuromorphic chips, organoid processors, sparse routing, and on-device inference are four different attacks on one gap. Whether the gap is closable by better engineering or reflects something specific to biological substrate is, at present, genuinely unknown — which is precisely why the substrate experiments are worth running even though their computational results are unimpressive.

Reading claims in a field with money in it

What the next decade is likely to settle

Tools & Tutorials

• Lava — end-to-end neuromorphic tutorial — build and run a spiking classifier in the open-source framework that targets Loihi 2, starting from processes and connections rather than layers, with a simulation backend so no special hardware is needed.
• FinalSpark live organoid feed — a continuously updating view of electrical activity recorded from human brain organoids sitting on multi-electrode arrays; useful for seeing how irregular real spontaneous activity looks.
• Brain-Score — an interactive leaderboard scoring vision models on how well their internal representations predict recorded neural activity and primate behaviour, with the benchmarks and submission pipeline open.
• The Virtual Brain on EBRAINS — the simulator behind virtual brain twins, with hosted notebooks for building a whole-brain network model from a connectome and comparing simulated to recorded dynamics.
• Open Neuromorphic — a community-maintained guide to hardware, software frameworks, and datasets in the field, including talk recordings and a comparison of the available chips.

Further reading

• Kudithipudi, D., Schuman, C., Vineyard, C. M., et al. (2025). Neuromorphic Computing at Scale. Nature, 637, 801–812. — a twenty-three-author roadmap covering what scaling actually requires: architecture, benchmarks, tooling, and the ecosystem gaps that have held the field back.
• Wang, H. E., Triebkorn, P., Breyton, M., et al. (2024). Virtual Brain Twins: From Basic Neuroscience to Clinical Use. National Science Review, 11(5). — the standard reference on personalisation, model inversion, and applications across epilepsy, Alzheimer's, Parkinson's, multiple sclerosis, and psychiatry.
• Kunz, E. M., Abramovich Krasa, B., Kamdar, F., et al. (2025). Inner Speech in Motor Cortex and Implications for Speech Neuroprostheses. Cell, 188(17). — shows imagined sentences are decodable in real time across four participants, and proposes a keyword-gated safeguard for mental privacy.
• Card, N. S., Singer-Clark, T., Peracha, H., et al. (2026). Long-Term Independent Use of an Intracortical Brain–Computer Interface. Nature Medicine, 32, 2504–2510. — nearly two years of unsupervised at-home speech and cursor control, and the clearest account of what stability rather than peak accuracy demands.
• Smirnova, L., Caffo, B. S., Gracias, D. H., et al. (2023). Organoid Intelligence: The New Frontier in Biocomputing. Frontiers in Science, 1:1017235. — the agenda-setting paper for wetware computing, including the ethical framework its authors argue has to be built alongside the technology.
• Molteni, M. (2025). Brain Organoid Pioneers Fear Inflated Claims About Biocomputing Could Backfire. STAT. — reporting from a field meeting where the researchers themselves push back on the marketing vocabulary; a useful calibration against every other source on this topic.