Andy Kitchen is an Artificial Intelligence, Machine Learning and Neuroscience Researcher. He co-founded Cortical Labs, a research organisation that fuses living brain cells onto computing devices.

---

These are live-blogged notes from a session at NEXT24 in Hamburg. As such, they are prone to error, inaccuracy and lamentable crimes against grammar and syntax. Posts will be improved in the coming days and this note updated.

---

Irreversibility matters. Imagine you were playing the world’s greatest chess player — but you could always take back your moves. Eventually, you’d win. But the world is irreversible. Our lives are a series of irreversible decisions. But our brains allow us to reverse decisions, to think about how things could have been different. And the tension between the two is the source of cognition.

To better understand and harness cognition, Cortical Labs has been growing neural tissue over highly sensitive electrodes. They’ve created closed loops of action and stimulus, with milliseconds response times, proving how sensitive the electrodes are and how closely integrated they are with the neurons. And then they taught these tissue samples to play Pong — and get better at it.

Training the mini-brain

The brain is a cognitive prediction machine. And we’re comfortable when our inputs match our predictions. If you want more predictable inputs, you either build a better mental model of the world, or change the inputs to reduce unpredictability. That’s why we like houses because they make the weather inside more predictable. Because the brain dislikes unpredictability, if you give it unpredictable feedback, it dislikes it, and tries to learn to avoid that. And that’s the idea that they harness to train their mini-brains.

As they’ve improved the system, they can now get feedback loops in a millisecond. And they’re hitting six-month cell lifetimes with their embodied experiments. That means they can run these long experiments and tests.

So, where is this going? Well, towards personalised drug discovery, for one. Embodiment is a key part of the neural process — reproducing that in the lab allows them to trust the effectiveness of drugs that act on the brain. Imagine the impact on, say Alzheimer’s.

Embodied AI

So, they have proved that you can teach a brain to do something outside the body. They’ve learned to do continuous embodiment, so they can run long-term experiments. These are the building blocks of a very different, much more embodied form of AI.

We think a lot about statistical models, like LLMs. Asking if an LLM understands language is like asking if a weather model can get wet. It can’t experience language in the way embodied creatures can.

We’re not an algorithm because we’re embodied humans, and living our life through making a series of irreversible decisions. That’s what separates us from AI — of whatever sort.