Two developments in June 2025 pushed at the boundary between artificial and biological intelligence from opposite directions. A team at the Chinese Academy of Sciences reported that large language models spontaneously organize everyday objects along conceptual dimensions that closely resemble those used by people, and an Australian company began selling a computer whose processing is done by living human neurons.
Object concepts inside language models
The study, published in Nature Machine Intelligence and announced by the Chinese Academy of Sciences on June 11, was led by researchers at the academy’s Institute of Automation. The team wanted to know whether language models hold something like a concept of an object, as opposed to merely reproducing statistical associations between words and pixels.
To test this they borrowed a method from cognitive psychology called the odd-one-out task. The models, including OpenAI’s ChatGPT-3.5 and Google’s Gemini Pro Vision, were shown descriptions or images of three objects at a time and asked which did not belong. Across millions of such judgments covering nearly 2,000 natural objects, the researchers derived the underlying dimensions the models were using to make their choices.
According to the paper, the models’ choices could be captured by 66 interpretable dimensions, such as whether an object is animate, edible, man-made or associated with a particular setting. Those dimensions aligned closely with the ones that emerge when humans perform the same task, and the multimodal model’s representations also correlated with activity patterns in brain regions that specialize in processing categories such as faces, places and bodies.
The researchers are careful about what this does and does not show. The models group objects by meaning and context, not only by visual features such as color and shape, which the team describes as a human-like conceptual organization. But the paper does not claim that the models understand objects as people do, and the authors note that the representations are not identical to human ones. Emotional and metaphorical dimensions of understanding were outside the study’s scope.
A computer made of neurons
At the same time, Melbourne-based Cortical Labs has been shipping the CL1, which it describes as the first commercially available biological computer. Announced in March 2025 at a price of about $35,000, the device houses roughly 800,000 lab-grown human neurons on a silicon electrode array that both reads their activity and stimulates them, with signaling at sub-millisecond speeds. A life-support system keeps the cells alive for months.
The company’s earlier work, published in 2022, showed that a culture of neurons on such an array could learn to play a simplified version of the video game Pong. Cortical Labs says the CL1 is intended for research and drug-testing laboratories that want to observe how real neurons process information and learn, including in models of neurological conditions such as epilepsy. For customers who do not want to maintain the hardware, the company also offers remote access to its devices under a model it calls “Wetware-as-a-Service.”
Cortical Labs argues that biological neurons are far more energy-efficient learners than conventional deep learning systems. That claim is plausible in principle, since the brain performs its work on roughly 20 watts, but head-to-head comparisons with silicon AI on practical tasks have not yet been published, and the CL1’s capabilities so far have been demonstrated on simple problems.
What the two results have in common
The two projects approach the same question from opposite ends: one asks whether silicon systems trained on human data come to represent the world as humans do, the other asks whether human cells can be made to compute like a machine. Neither answers the larger questions that follow, about whether machine systems could ever have genuine awareness or grasp cultural and emotional nuance. Researchers in both fields describe the work as early, and the next steps as incremental: more detailed comparisons between model and brain representations, and more capable tasks for cultured neurons to learn.
