Discover more from Ken’s Substack
AI agents enter the conversation
... and built me a new aeroplane while they were at it
My latest for the Wavell Room:
Stand by for an EoD exclusive, as we reveal the sixth generation multirole aircraft for the RAF. Sirs and ma’ams, I give you the Peregrine:
Now that, you’ll agree, is quite something. Here’s the manufacturer’s bumpf:
"Observe the sleek, angular lines and radar-absorbing surfaces, combined with the internal weapon bays for a low radar profile. The dark, matte finish enhances its stealth capabilities, while the design remains aggressive yet elegant. Notice the advanced sensors and communication systems integrated seamlessly into its form. The engines, designed to reduce heat signature, add to its stealth prowess. This design encapsulates speed, precision, and versatility, befitting the name and role of this formidable aircraft."
Relax - no secrets have been spilled. I created the Peregrine in no more than five minutes, with the help of a GPT ‘agent’ I created, called ‘Spitfire Consultant’. As well as naming the Peregrine, it chose this name itself and also knocked up its own logo, thusly:
This sort of ‘agent’ is a new product from OpenAI, maker of the market leading GPT-4 language model. You can see their boss, Sam Altman, demo the idea here, just last week. The idea is to produce a tailored assistant that can focus in on specific tasks. In the last few weeks, GPT-4 has added integrated picture output – hence my drawings. Video won’t be long in coming. Soon we’ll be able to watch the Peregrine swooping majestically through the Mach loop. And GPT already has voice interaction, which will surely also be coming to the ‘agent’ interface. In a month or so, I'll be conversing naturally with my agent, rather than bashing away at the keyboard. That should be fun, since I've already asked it to talk to me like a 1940s pilot – everything is ‘spiffing’ or ‘top hole’; and the Peregrine itself is, in its view, a ‘spanking new aeroplane’.
It doesn’t take long to cook up the agent, and the output is impressive, superficially at least. In the course of co-designing the Peregrine, we got into quite detailed discussions about air power roles and integrated air defences. The language model reckons that drone swarming is going to be a big factor in the future of air power; but it also pointed to the versatility of crewed aircraft, their psychological and symbolic impact, and the advantages of having a human decision-maker forward in highly contested environments.
Interesting stuff. But we’re a long way from using tools like this for real aircraft design. There are a number of large flaws with agents, as demonstrated by my interaction with the Spitfire Consultant. Here, for example, is the blueprint it generated for the Peregrine:
'One for the boffins to mull over,' as it told me. Well, indeed. And here she is, flying low in the mountains of North Wales.
Wait – isn’t that a Tornado? With the wings of a Harrier, perhaps. It’s definitely not the original design we created. And the blueprint differs too – superficially it's similar, but look again at the engine inlets and exhaust for some striking differences. There’s clearly a problem here with image generation via Dall-e, which will be familiar to anyone who’s tried their hand at wrangling it already. The words on the blueprint are illegible nonsense too.
Anyway, while I’m partial to the original design, it’s not all that original, is it? Looks a bit like an F-35/F-22/J-20. Perhaps, like family SUVs, all 5th and 6th generation designs tend to blend into the same one. But still, this mash-up approach to design says something about the capacity of language models to create breakthrough ideas, as opposed to rehashing existing ones. I suspect that when the mysterious US NGAD eventually hoves into view, it might look rather different from what we have now. The radically different airframes of the U2, SR-71 and F-117 are striking examples of human innovation – on which, I highly recommend Ben Rich’s memoir of his time running the Skunk Works, where all these types were created.
Meanwhile, the AI agent itself tried to shrug off the image flaws with some entertaining nonsense about how even ‘a Spitfire looks a tad different depending on the angle or the light’. Who says machines aren’t out to deceive us?
Beyond the image generation wobbles, though, there’s plenty here that should interest practitioners. The discussion about air power roles and the challenges of modern warfare was rich and well-informed. And there’s scope to build its knowledge base too, for example by uploading existing air power doctrine, or – if you were confident it wouldn’t spill state secrets – existing concepts and designs. For example, I uploaded current UK Air and Space power doctrine, and also the fantastic concept note on human-machine teaming – one of DCDC’s better efforts.
Overall, I wouldn’t be shuttering the RAF Rapid Capabilities Office and outsourcing everything to OpenAI’s agent. But you can see where this is going. Soon enough, we’ll have specialised AI models, helping us innovate and develop concepts in all manner of domains, including, naturally, defence. In the meantime, the Spitfire Consultant is coming soon to a staff college classroom near you.