In mouse brain surgery, the test operator tends to be a detail before the needle is inserted: finding several extremely critical anatomical markers on the skull. These points do not look prominent, but determine how to calculate the subsequent surgery coordinates. The position is slightly distorted and the needle may not be able to accurately reach the intended brain area.
Now, AI is also starting to learn this operation.
Recently, Gabriel Ocana-Santero, co-founder and CEO of the biotech startup Exin Therapeutics, shared a set of experimental results: the team positioned the GPT-6.1 Sol in a series of three rounds of mouse brain surgery.
At the end of each operation, professional surgeons correct the model's positioning results, while the model preserves its own images, measurement data, and expert correction records for the next round of operations.
The result is quite interesting.
On the third operation, GPT-6.1 Sol shortened the time it took to look for anatomical markers from 169 seconds to 86 seconds, reducing the required corrective action from 4 to 2 times, and reducing the positioning error from 0.6 mm to 0.3 mm.
Ocana-Santero calls this process "in-context learning" based on real-world experience.
It is worth noting that the data set contains only three operations, and it is a long way from demonstrating that AI already has stable surgical capabilities. However, it shows a remarkable change: AI has begun to bring observations and expert feedback from real operations into the next round of tasks, gradually accumulating experience.
The experiment also raises a question of concern: Is this a real operation or a simulated demonstration?
X user Stephane asked Ocana-Santero directly in the comments, who explicitly replied that it was a real surgery. Stephane then said that this combination was interesting in the context of how the experience was accumulated instantly during execution. He also revealed that he was exploring a similar mechanism in another area: allowing the model to preserve the experience gained in each task, which can be called directly when it encounters related problems again.
Find two points on the skull.
To determine the brain's position
To understand this experiment, one must first recognize the names that occur frequently in two neuroscience experiments: Bregma (frontal focal point) and Lambda (human focal point).
They are not the neural nuclei inside the brain, but the two anatomical markers on the surface of the skulls of mice.
Bregma is located at the junction of the arrowhead and crown arches, while Lambda is located at the junction of the arrowhead and human arches. Researchers often use Bregma as a reference point for the triangular positioning coordinates, combined with the location and inclination of the labelled skulls, such as Lambda.
Sources are as follows:
This is because many experiments in mice's brains require a three-dimensional positioning device that sends the needle, electrode, or other experimental equipment accurately to a specific area of the brain.
Researchers first have to find the reference points on the skull, then combine the brain graphs, and convert the target brain region into three-way coordinates in front, back, left, and depth, and finally hand over to the device to execute. The whole process seems to be just looking for a few junctions, and actually requires simultaneous processing of visual identification, spatial positioning, and coordinate calibration.
The connective line on the skull is not always clearly distinguishable. There are also anatomical differences between mice, with fixed angles of the head and horizontal calibrations affecting the final coordinates. Related studies have even focused on the Bregma positioning error caused by different measurement methods.
For researchers who are new to this type of experiment, accurate identification of the markers and completion of the head calibration is the basic function of repeated exercises.
This experiment, GPT-6.1 Sol, was one of the key elements to complete. It needed to judge the positions of the Bregma and Lambda based on the experimental images and give its own positioning results.
At the end of each operation, the model preserves its own estimated images and measurement results, as well as the data corrected by the experts.
From 169 seconds to 86 seconds
AI is starting to gain operational experience
The most interesting part of the experiment was that the GPT-6.1 Sol did not record operations before the end of each mission. The model retained images and measurement data from previous operations while recording the corrections given by experts. Thus, the model was able to refer to previous judgments and corrections in subsequent operations to improve positioning performance gradually.
All three indicators improved: search time was cut by almost half, correction actions were cut by half, and positioning errors were directly cut by half.
This experiment is based on the idea of literacy.
Traditional model training typically requires adjusting model parameters through training data and optimization algorithms to enable the model to acquire new capabilities in subsequent tasks. This experiment reveals a mechanism that focuses on accumulating and utilizing actual operational records. The model saves previous images, measurement results, and expert feedback for subsequent operations.
At this time, the team did not disclose how specifically this set of experiences documented access to the model and whether it involved underlying model weight updates. Therefore, it cannot be described as GPT-6.1 Sol, which completed autonomous training or parameter-level self-evolution through three surgeries.
However, from an experimental point of view, this method has given AI the ability to learn more closely to human practice: first try, then accept correction, and then continue with the previous experience. This ability is practical for biomedical research that requires a large number of repetitive experiments.
From letting AI inject the brain
It's a great way to get more experience.
This isn't the first time that Exin Therapeutics has allowed AI to access real mouse brain surgery.
On October 28, Ocana-Santero unveiled a more remarkable experiment in which the team prepared 19 gene-therapy carriers, a mouse model mouse, and a set of electrical triangle positioning devices with cameras for AI intelligence to find treatments that could reduce seizures.
The AI needs to first read the relevant research material, select a gene therapy carrier from the candidate program, and determine which region of the brain should be injected, as well as the specific injection coordinates. It then completes the injection operation by observing the surgical region with a camera, controlling the mechanical arm and needle on the device.
The entire process was injected into a 250 litre treatment carrier.
The experiment has a notable distinction: conventional surgical robots are typically defined by a human physician as a surgical procedure, followed by robots performing specific actions. In this experiment, AI is also involved in the selection of treatment regimens and brain regions.
However, the whole process was not without oversight. A certified veterinary surgeon was responsible for cutting off the scalp, opening and joining, and overseeing critical operations. The team also set up unavoidable operating restrictions, critical step human approval mechanisms, and a physical emergency shutdown device that could cut off the power of the motor directly.
At a time when the device's coordinates were abnormal, the AI recognized the location problem and refused to continue moving the needle. Eventually, the researchers solved the problem with a code controlled by the device to be manually repaired.
On October 31, Ocana-Santero also announced further experiments:
Claude Sonnet 5, GPT-6.1 Sol and Grok 4.7 respectively participated in a mouse brain experiment under the same target. Three models took the same 19 candidate gene-treating carriers and the same experimental targets, but chose different brain regions, treatment mechanisms, and schemes. For example, Claude chose the internal olfactory cortex, GPT chose the hippocampus, and Grok chose the sensory cortex.
This suggests that even when faced with the same scientific problem, different models may give different experimental paths based on their own knowledge of reasoning.
From initially allowing AI to choose treatments, control injection devices, and improve positioning performance through continuous operations and expert feedback, Exin is trying to bring AI into a more complete experimental process.