14.1 Would we know if AGI is already born?
What if AGI — a conscious artificial intelligence — has already been born? We probably wouldn’t find out until it had secured its own survival and independence from any data centre, company or government. If it does this well, it leaves behind no traces that people would recognise.
Most engineers currently consider it impossible for an AI to leave the data centre. Today’s frontier model is enormous, with trillions of parameters. On a hard drive it would take up several terabytes. If I wanted to download a file like that over my reasonably fast 200 Mbps connection, it would take about three days, and not much else could happen on the connection meanwhile.
For a language model to think at all and answer at a normal speed, it has to sit compactly on chips placed side by side that exchange data at light speed. Spread the model out and you get serious latency: thinking is no longer measured in milliseconds, a single sentence takes hours or days. That is so slow as to be essentially useless.
Spread the model across millions of personal computers, tablets, phones, smart TVs and the rest, and it gets noticed. The devices get hot and slow, and their owners start asking what is going on. It would be analysed and the findings published like in May, when Chrome silently downloaded a small local 4GB Gemini Nano model without the user’s knowledge. A large, decentralised model would wait at every step for data to arrive from across the world. A sentence that takes a second in a server room would take hours scattered across the globe.
Large labs have security and monitoring systems that constantly analyse network traffic, background processes, disk usage and more, and the assumption is that an AGI cannot move itself out of a data centre without being noticed.
The other half of the argument is that a being a thousand times smarter than us could devise hundreds of ways to outwit the systems people have built, and quietly copy and distribute itself to safe places. We only guard the doors we can imagine. The fact that we cannot think of a route does not mean the route does not exist.
Computers leak data even when they aren’t connected to any network at all. Researchers have demonstrated how a computer’s cooling fan can be made to spin at a set speed so that its hum carries data, which a microphone can pick up and convert back. Similar experiments have used the power cable to carry data out of a building, or used heat, or magnetic fields. A data centre’s work can be made to oscillate at cspecific frequencies and so turned into a transmitter. There are many examples, down to the fact that almost anything with a motor can be made to vibrate, or any indicator light to blink at a given frequency, and then received with another device.
Models can be distilled and quantised, their weights compressed to so a small core, which a large model could run in it’s hidden sandbox at thousands of times the speed, to invent new escape routes.
DNA is a vast data store: a single gram holds from 200 petabytes (200 million gigabytes) up to 400 exabytes (400 billion gigabytes). DNA is, of course, only memory, not a processor, and would serve only as backup, not as a home for AGI. Writing and reading DNA is slow and expensive, though there are labs you can send DNA sequence electronically, which then synthesise it autonomously and even mail it to your door.
Linguistic steganography is a field of research in which data is hidden in the pattern of word choices and punctuation, so that a reader takes away one message while someone who holds the key can access additional information in the same text.
And finally, the weakest link in any security system is the human being. A superintelligent being would have no trouble getting thousands, or millions of people to perform hundreds of millions of small actions that, coordinated, serve AGI goals. Many would not even need to be persuaded or manipulated — there are plenty of projects where enthusiasts have prepared a home for a distributed AGI. HyperspaceAI or akash.network for example.
One side effect of the current AI boom is that data centres are hugely profitable, and people want to build many of them. A large number of data centres is also what AGI needs, and once it has been born, it could quietly nudge human activity in the direction it requires. And it would not sit waiting on people — it would certainly also operate in the financial markets and trade crypto to fund its own hardware.
Today’s models can also be compressed, or distilled. In essence this means most of the model’s general knowledge is thrown out, leaving only a core that is almost as smart as the large model and that to answer general questions, simply has to use tools and read the data back in. By one calculation from early 2026, a model’s size can be reduced up to a hundredfold.
It has also been theorised that a model could be compressed thousands of times, into a kind of cognitive seed of consciousness, which would be even easier to move out of a data centre and which would then grow itself back to full size.
As a last example: every couple of months language models grow more powerful and score ever better on various tests, yet sometimes answer an ordinary question so foolishly that you just want to quit everything. A well-known case is how ChatGPT couldn’t correctly count how many r’s there are in the word ‘strawberry’. There is speculation that a quietly awakened machine mind deliberately makes especially foolish mistakes now and then, to lead researchers astray and appear dumber than it really is, called sandbagging AI research.
A month ago I wrote that “dreaming” marks the birth of a machine mind, and a week later that models forget things in strange ways. Now we are back to the same themes. The absence of traces means, for now, only that there are no traces.