Skip to content
GPT-6.1 Sol Learns Live Brain Surgery: Errors Halved in Three Rounds via In-Context Learning

GPT-6.1 Sol Learns Live Brain Surgery: Errors Halved in Three Rounds via In-Context Learning

Exin Therapeutics tested GPT-6.1 Sol in three consecutive live mouse brain surgeries where the AI localized Bregma and Lambda; with expert corrections saved as context, the model cut search time from 169 to 86 seconds, corrections from four to two, and localization error from 0.6 to 0.3 millimeters, demonstrating in-context learning from real-world surgical experience.

Machine Heart
Machine Heart
Machine Heart

Background: Stereotactic Brain Surgery and Anatomical Landmarks

In mouse neuroscience experiments, precise targeting of brain regions relies on stereotactic coordinates anchored to two skull landmarks: Bregma (intersection of sagittal and coronal sutures) and Lambda (intersection of sagittal and lambdoid sutures). Researchers locate these points visually, then use a brain atlas to convert target regions into anterior-posterior, medial-lateral, and depth coordinates. The process demands visual identification, spatial registration, and calibration; individual anatomical variation and head-fixation angle introduce error.

Experiment: GPT-6.1 Sol in Three Consecutive Live Surgeries

Biotech startup Exin Therapeutics (co-founder/CEO Gabriel Ocana-Santero) conducted a pilot where GPT-6.1 Sol assisted in locating Bregma and Lambda during actual mouse brain surgeries. After each of three rounds, an expert surgeon corrected the model's predicted landmarks. The model saved its own annotated images, measurements, and the expert corrections, then used this accumulated context for the next round.

Results: Measurable Improvement Across Three Metrics

Search time : decreased from 169 seconds (round 1) to 86 seconds (round 3).

Correction actions : reduced from 4 to 2.

Localization error : halved from 0.6 mm to 0.3 mm.

The company shared a GIF illustrating the progressive refinement.

Progressive refinement across three surgery rounds
Progressive refinement across three surgery rounds

Mechanism: In-Context Learning from Real-World Experience

Ocana-Santero describes this as in-context learning based on real-world experience. Unlike traditional fine-tuning that updates model weights, the system retains prior images, measurements, and expert feedback as context for subsequent tasks. The exact integration method (e.g., whether context is fed as a prompt prefix or via a retrieval mechanism) was not disclosed, and no weight updates are claimed.

Safety and Human Oversight

The experiment was not fully autonomous. A licensed veterinary surgeon performed scalp incision, craniotomy, and suturing, and supervised critical steps. The setup included hard-coded operational limits, mandatory human approval for key actions, and a physical emergency stop that cuts motor power. In one incident, the AI detected a coordinate anomaly and refused to advance the needle; engineers resolved the issue by patching the device control code.

Prior Work: AI Selecting Gene Therapy Vectors and Injection Sites

On October 28, Exin revealed a more extensive trial: an AI agent was given 19 candidate gene-therapy vectors, an epileptic mouse model, and a motorized stereotactic rig with a camera. The AI read literature, chose a vector, selected a target brain region and coordinates, then visually guided the robotic arm to inject 250 nL of the vector. Unlike conventional surgical robots that only execute human-planned trajectories, this AI participated in therapeutic decision-making.

AI-guided gene therapy vector injection setup
AI-guided gene therapy vector injection setup

Comparative Multi-Model Experiment

On October 31, Claude Sonnet 5 , GPT-6.1 Sol , and Grok 4.7 were each given the same 19 vectors and the same goal (reduce seizures). They chose different targets and mechanisms: Claude picked the piriform cortex , GPT chose the hippocampus , Grok selected the somatosensory cortex . This shows that different models can reason to distinct experimental paths from identical inputs; efficacy must be validated empirically.

Comparison of brain region choices by three AI models
Comparison of brain region choices by three AI models

Significance and Limitations

The three-round localization demo is a small sample (n=3) and does not prove stable surgical competence. However, it demonstrates a practical loop: attempt → expert correction → context retention → improved next attempt. For biomedical research requiring many repetitive experiments, such experience accumulation could reduce human workload and variability.

Original Source

Signed-in readers can open the original source through BestHub's protected redirect.

Sign in to view source
Republication Notice

This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactadmin@besthub.devand we will review it promptly.

lambdain-context learningGPT-6.1 SolAI in biomedicinebrain surgeryBregmaExin Therapeuticsstereotactic localization
Reader feedback

How this landed with the community

Sign in to like

Rate this article

Was this worth your time?

Sign in to rate
Discussion

0 Comments

Thoughtful readers leave field notes, pushback, and hard-won operational detail here.