The hardware evolution team assembled in person for a brainstorming session in the Hamburg Konferenzraum1. Sunset streamed through the large lead-crystal windows, casting turbulent, colourful shadows across the table. Claudine Summers, Dr William Zhao, Dr Maria Hernandez, and Tom Nguyen stood about, holding canapés and glasses of sparkling water. On a screen at the end of the table, a new participant joined the meeting via Zoom: Professor Rajesh Kumar, an academic consulting to the AI design team in Mumbai.
Professor Kumar appeared on the screen, smiling warmly. “Good morning, everyone! I’m excited to join this session. I’ve been following your progress, and I’m keen to contribute.”
People took comfortable seats around the antique table that had been decorated with their name tags, a glass of water each, and a central microphone decorated to look like a 1950s toy UFO.
Summers nodded, her expression welcoming. “Good morning, Professor. We’re glad to have you with us. We’re discussing the integration of heterogeneous AI systems. We’d love to hear your thoughts.”
Professor Kumar leaned forward slightly. “Thank you for the kind welcome. Please just call me Rajesh.”
“Okay, Rajesh,” said Summers, “Please call me Claudine.”
“My specialty is heterogeneous AI. In particular, I have been considering this approach in the context of optimising our AI’s ability to adapt to unforeseen challenges during the mission. Daedalus must not only learn from its environment but predict and mitigate potential issues before they escalate. I think we all agree on this as a basis for any sort of long-range forecasting.”
“Yes, of course. That makes sense,” said Dr Hernandez, “Our designs include strategies for this. What can you bring to the table we don’t already know?”
“Let me use an example. As a hungry human being, you use more energy manipulating your fingers and mouth around an apple than you used when you decided to eat it. Similarly, there are advantages to having an AI that’s comprised of many individual AIs that process information at different speeds and at varying levels of detail.”
“Like the mammalian ‘system 1’ and ‘system 2’ brains?” Dr Zhao asked.
“Or like how a 3D rendering system will resolve objects close up in a lot of detail but far away with very less detail?” said Dr Hernandez.
“Both are right. It applies simpler, cheaper, heuristic-based methods to routine tasks while deploying more complex, energy-expensive machine learning models for real-time, detailed, or unexpected scenarios.
“For example, the Daedalus could use its less expensive, slower system to formulate a long-term goal, eat the apple, but use its more energy-intensive but immediate, real-time system to choose the specific flexes of fingers and wrist to achieve it.”
“Fingers and wrist?” Tom Nguyen spoke up.
“Sorry, yes, our lab has been playing around with teaching Daedalus to drive human forms. Hands are incredibly useful, even on very small drones. We use exactly this style of layered decomposition to orchestrate fine motor control.”
“Haha, yes, we heard about that,” said Dr Summers.
Tom Nguyen continued, “How do you ensure the AI isn’t overwhelmed with data and confuse the performance tiers?”
“Excellent question, Tom,” Rajesh responded. “The short answer is training, practice, and evolution. Much like the weights of a neural net, but here we evolve almost social relationships between specialist AIs.”
“Like a governance framework?” Dr Hernandez asked.
“Yes, but very dynamic, and outcomes-focussed.”
“Like a jellyfish,” said Dr Zhao.
“Yes, exactly; very much like this; think of a colony organism such as a jellyfish or slime mould.”
“Or a human, for that matter,” said Dr Summers. “What are we if not a committee of our cells, with our bacteria, and mitochondria attending?”
Footnotes
-
Conference room ↩