The Machine That Passed the Medical Exam but Cannot Feel a Handshake

On embodied AI, neuromorphic chips, quantum computing - and why what we have built may be neither artificial nor human.

A human hand and a robotic tactile hand reaching toward each other against a neural network background

I showed a demo to a client last month. The AI answered every question correctly. Summarised a complex report. Suggested next steps.

Then he asked me - does it actually understand what it's saying?

I paused. Honest answer is no.

We built a machine that can pass a medical licensing exam. It cannot feel a handshake. It can describe the warmth of a palm, the firmness of a grip - it has read millions of descriptions of all of it. It has experienced none of it. There is no sensor, no body, no moment of contact anywhere in the system.

At the end of the day current AI is just complex mathematics. Matrix multiplication at scale. We are solving real world problems with mathematical vectors.

Something about how the brain computes is fundamentally different from what we are doing. Not 10x different. Different in kind. And the longer I build these systems, the more that gap bothers me.

What the Brain Actually Does

Your brain runs on 20 watts. Less than a bulb in your fridge. It has 100 trillion synaptic connections. It learns "hot" from one touch of a stove - one time, never forgets. Our best AI models have maybe 1-2% of that connectivity, cost hundreds of millions to train, need a small power plant to run.

The brain integrates touch, temperature, emotion, and memory simultaneously - in the same tissue, with no separation between sensing, computing, and learning. There is no inference phase and training phase. There is just living.

Our largest models don't do any of that. They take text in and put text out. And we have been calling that intelligence.

1. Embodied AI - Closing the Grounding Gap

Close-up of a robotic hand with tactile sensor pads grasping a delicate glass marble, showing precise grip pressure

A child touches a hot stove. That one second wires the word, the pain, the memory, the emotion - all together, permanently. A language model reads 10 billion sentences about heat and still has no idea what hot feels like. It knows the word. It's never had the burn.

Every system I've built has this hole in it.

This is what researchers call the grounding problem. The word is not connected to any real experience. It floats. And no amount of more data fixes it - because the problem is not data, it is the absence of a body.

Meta's DIGIT is a compact tactile fingertip sensor - a silicone pad with a camera behind it that reads deformation at high resolution, cheap enough at around $15 per unit to put on every finger of a robot hand. ReSkin, built with Carnegie Mellon, is a deformable skin with embedded magnetic particles. When it deforms, magnetometers read the change in magnetic field and a model converts that into contact location and force. Robots with ReSkin can grasp grapes without crushing them. That sounds simple until you try to write that control loop yourself.

Google DeepMind's RT-2 takes this further - it treats robot actions as just another output of a language model:

Vision input -> Language model -> Action tokens -> Motor commands

The same architecture that completes your sentences moves an arm. Figure AI is putting this whole stack into humanoid form, with robots learning tasks by watching rather than by hand-coded instructions.

This is more aligned to the way humans understand - not just text. When a robot gets the grip wrong, the world corrects it. Physics, not a loss function. That is the first real crack in the grounding problem.

2. Neuromorphic Chips - Computing the Way Neurons Do

Neuromorphic chip die with sparse spiking neural network visualization overlaid - most neurons dark, a few firing in teal

The second problem is the hardware itself. GPUs compute the way GPUs compute, not the way brains compute. A GPU multiplies dense matrices on a fixed clock, moving data back and forth between memory and compute. It burns energy whether the computation is interesting or not.

Neurons don't do that.

GPU:    tick -> multiply -> tick -> multiply -> (constant energy draw)
Neuron: silence -> silence -> SPIKE -> silence -> SPIKE
                                    (energy only on event)

A neuron fires only when its inputs cross a threshold. No spike, no energy spent.

Intel Loihi 2 and IBM NorthPole are neuromorphic chips built on this principle. They don't do matrix multiplication on a large scale. They fire spikes the way neurons do. IBM NorthPole eliminates the memory and compute separation entirely - memory lives inside the compute fabric, the way synapses work in a brain.

This is much more efficient - like 1000x of the way we do via matrix multiplication for all workloads. First hardware that is aligned on the way a brain actually computes.

I want to be honest here - neuromorphic computing has been five years away for twenty years. The tooling is still immature and training algorithms don't transfer cleanly from deep learning. Nobody has trained anything GPT-class on spikes. But it is the only computing paradigm we have that is even shaped like biology. And if the brain's 20 watts tells us anything, the shape matters.

3. Quantum Computing - What It Can and Cannot Do

A dilution refrigerator cryostat shot from below - cold blue-silver layers receding upward, fragile and precise

Although quantum cannot solve the AI efficiency problem completely - I want to be clear about where it actually helps and where it does not.

Where it genuinely helps: simulating quantum systems. Molecular dynamics, materials discovery, drug chemistry - problems where nature itself is quantum mechanical and classical computers must approximate exponentially complex state spaces. If quantum computing changes AI, it will likely be by transforming the data - letting us simulate molecules we currently cannot - not the training loop itself.

Where it does not help: training transformers. There is no proven exponential quantum speedup for matrix multiplication. Anyone telling you quantum computers will train the next GPT is not being straight with you.

Where we actually are:

Current state:
  Physical qubits  : 105 (Google Willow)
  Coherence time   : microseconds
  Physical qubits per logical qubit: ~1000
  Useful fault-tolerant computation: not yet

Google's Willow chip hit 105 qubits - and the important result was that error rates decreased as the array scaled up. First real evidence that error correction can work at scale. IBM's roadmap pushes toward error-corrected logical qubits over the next few years.

Decoherence is still the killer. Qubits hold their quantum state for microseconds before the environment destroys it. Respect the research. The hype is ahead of the engineering by a long distance.

Neither Artificial nor Human

Put the three together and look at what is coming.

A machine with magnetic skin that feels an object slipping and tightens its grip. A chip that learns from single events through local spikes, drawing milliwatts. Eventually, simulation capabilities that let these systems model the physical world at the molecular level.

What is that exactly?

It is not artificial intelligence - not in the sense the founders of the field meant. The architecture is wrong, the learning is wrong, the energy budget is wrong. And it is obviously not human intelligence either. It is a third thing. Something that thinks differently, fails differently, and is slowly starting to sense the world in ways we cannot predict.

We keep asking the wrong question. When will AI be as good as human?

It was never on that path.

The real question is - when a robot with magnetic skin feels an object slip and fixes its grip, is that feeling? Or just measurement? And do we actually know the difference?

I don't. And I build these systems for a living.

Research links:

If this raised more questions than it answered - that is the point. Reach out on LinkedIn or drop a comment below.