I have been spending an unreasonable amount of time asking what happens when you give an insect brain something biology never gave it:
persistent external memory.
Then I came across a paper from 2009.
And for approximately five minutes my reaction was:
well, screw everything I'm doing, DARPA already turned insects into computers.
That isn't actually what happened.
But what they did do is weird enough that the exaggerated version is almost unnecessary.
In 2009, researchers published a paper called "Insect–Machine Interface Based Neurocybernetics."
It was funded by DARPA's Hybrid Insect Micro-Electro-Mechanical Systems program, or HI-MEMS.
Their experimental animal was Manduca sexta, the tobacco hawkmoth.
And their basic idea was beautifully unhinged:
don't attach electronics to an adult insect. Implant the electronics while the insect is still becoming an adult, and let metamorphosis build the interface for you.
That worked.
Really, really well.
TL;DR: the science
The researchers inserted flexible microfabricated electrodes into Manduca sexta during the pupal stage.
Then they waited.
As the moth continued developing, its tissues formed around the implanted probes.
The researchers even fabricated 200-micrometer holes in the probe tips so developing muscle could grow through the electronics, creating a biological anchor.
The insect's cuticle healed around the insertion points without glue.
At the optimal implantation time, approximately seven days before emergence, 90% of the moths successfully emerged as adults with fully inflated wings.
Across more than 100 implantations, the researchers reported 98% success placing the probes into their intended flight muscles.
They could use those implanted electrodes to record muscle activity.
They could also stimulate the muscles.
Stimulating one flight muscle pulled a wing downward.
Stimulating another pulled it upward.
Stimulating both at sufficiently high frequency could stop wing flapping.
And stimulating one side of a naturally flying moth produced a yaw toward the stimulated side.
In other words:
they implanted electronics during metamorphosis, the animal developed around them, and afterward those electronics could interact directly with the machinery responsible for flight.
That is not a cyberpunk novel.
That is IEEE Transactions on Biomedical Engineering, June 2009.
The really clever part isn't the remote control
The remote-control angle is obviously the thing that gets people's attention.
"Cyborg moth."
"DARPA insect drone."
"Military remotely controls animal."
Sure.
But from an engineering perspective, I don't think that's actually the most interesting part.
The brilliant idea is:
Use development as part of the manufacturing process.
Normally, implanting something into a fully developed animal means fighting an already completed structure.
The tissue is already where it is going to be.
The exoskeleton is already hard.
You're cutting into something that was not designed to accommodate your device.
So instead, the researchers went earlier.
During metamorphosis, an insect is already performing an absurd biological reconstruction project.
Structures are degenerating.
Others are forming.
Muscles are developing.
The adult cuticle is being produced.
So the researchers inserted their interface while that process was still underway.
Instead of forcing the completed organism to accept the electronics, they allowed the organism to develop around the electronics.
They called the method Early Metamorphosis Insertion Technology: EMIT.
That distinction is important because I've seen posts claim that the researchers called the phenomenon "Metamorphic Growth."
They didn't.
Metamorphic growth is the trick.
EMIT is the technique.
They literally put holes in the electronics so the moth could grow through them
This is my favorite detail in the entire paper.
The flexible probes were made from polyimide with conductive traces and gold-coated electrode sites.
At the ends of the probes, the researchers added 200 μm openings.
Why?
So flight muscle could grow through them.
Not around them.
Through them.
The biological tissue becomes part of the mechanical mounting system.
The paper reports that muscle adhesion and growth through those holes substantially improved anchoring.
When researchers later removed probes that had been implanted during the pupal stage, significant tissue came out attached to them.
Probes inserted into adults showed much weaker integration.
That's wild because the electronics aren't merely surviving biology.
Biology is completing the assembly.
No glue required
They also tried something that sounds completely reasonable:
seal the surgical opening with adhesive.
That actually made things worse.
The adhesive could bind tissue layers that needed to separate normally during emergence.
So they stopped using it.
Instead, after inserting the probes, they simply allowed the insect's own cuticle to heal around them.
The animal produced the seal.
Again:
development is doing part of the engineering.
Okay, but was DARPA actually controlling moths?
Yes, with an important asterisk.
This paper does demonstrate externally triggered flight-muscle control.
The researchers stimulated the dorsal longitudinal and dorsoventral muscles responsible for wing movement.
They demonstrated repeatable unilateral and bilateral wing actuation.
During natural flight, stimulation on one side produced a turning response toward that side.
The authors describe these experiments as a proof of concept for insect steering.
There was also a separate 2009 paper from the same research lineage titled "Radio Control of Insects for Biobotic Domestication," reporting radio-controlled neuromuscular stimulation of Manduca sexta during flight.
So "radio-controlled moth" is not something somebody invented for Twitter.
But there is a difference between:
I can electrically induce motor responses that alter the animal's trajectory
and:
I have built a perfectly controllable biological quadcopter.
Those are not the same achievement.
The early work demonstrated controllable outputs and directional influence, not complete deterministic command over every aspect of insect behavior.
The animal is still an animal.
The viral version also oversells the interface a little
One description I saw called this a "high-bandwidth insect-machine interface."
I'd be careful with that wording.
The researchers absolutely created a mechanically stable and electrically functional interface.
They could stimulate selected muscle groups.
They could record electrophysiological activity.
And they explicitly discussed expanding this approach toward neural, sensory and behavioral interfaces.
But this wasn't somebody plugging Ethernet into a moth's brain.
The demonstrated interface was primarily with specific parts of the neuromuscular system.
That's already insane.
It doesn't need embellishment.
Then I got to the end of the paper
And this is where the whole thing suddenly collided with what I've been working on.
The authors start talking about what comes next.
They suggest that future interfaces could stimulate not only muscles but potentially the brain and thoracic ganglia.
They discuss feeding artificial inputs into chemical, mechanical and visual sensory systems.
And then they suggest something especially interesting:
using those interfaces to remotely train individual insects through conditioning.
That's where I stopped reading this as "fun historical cyborg paper."
Because I've accidentally spent the last several months attacking almost the complementary problem.
The thing I'm building isn't a cyborg insect
If you haven't followed the previous experiments, the short version is that I've been playing with a reconstructed insect neural architecture and an external memory system.
The project eventually became Life of Briain.
The important part isn't that I'm simulating a fly.
The interesting question is what happens when something approximately insect-like gets access to a capability evolution normally doesn't provide:
memory that can survive the individual.
Normal learning looks roughly like this:
experience → nervous system changes → animal behaves differently → animal dies
And the learned state largely dies with it.
Evolution preserves information on another timescale, but an individual's learned memories aren't normally copied into its descendants.
So I introduced another path:
experience → nervous system → external memory → death → new individual → inherited information
And that very quickly produces a much stranger question than "can inherited memory help?"
Of course useful information can help.
The interesting question is:
What happens when yesterday's useful information stops being true?
An inherited memory can become an anchor.
An experienced descendant can potentially perform worse than an ignorant one because it enters the world carrying confident evidence about a world that no longer exists.
That is the direction the experiment has taken.
Not merely persistent memory.
Transfer learning across lives.
DARPA attacked one side of the boundary
The 2009 work asks something like:
Can electronics become physically integrated with an insect deeply enough to read from and write to its biological control machinery?
And the answer was:
apparently yes.
My experiments are asking something closer to:
What happens if information acquired by one embodied learner persists outside that learner and becomes available to another?
Those are very different questions.
But they are uncomfortably complementary.
DARPA's system gives electronics access to biology.
Mine gives biology-inspired cognition access to persistent information outside itself.
One externalizes control.
The other externalizes memory.
And once you see those as separate components, a really weird architecture appears.
The cursed architecture
This is speculative.
I am not attaching Mycelium to a moth tomorrow.
Please don't call DARPA.
But conceptually, imagine this:
A living insect has its ordinary sensory systems.
It has its ordinary brain.
It has its ordinary muscles.
An insect-machine interface provides limited channels for reading biological activity and supplying artificial sensory or motor signals.
Outside the organism exists a persistent computational system capable of storing information learned during that insect's life.
Eventually that animal dies.
The memory doesn't.
A new animal receives information derived from previous animals through the machine interface.
Now you don't merely have a remotely controlled insect.
You have something much stranger:
a lineage whose learned information can persist beyond the biological individuals that generated it.
The insects remain disposable bodies.
The memory becomes the persistent organism.
And suddenly the question isn't:
Can we steer a moth?
It's:
What exactly constitutes the continuing individual when bodies are replaceable but acquired information persists?
Which is, unfortunately, exactly the kind of question that caused this entire project to get out of hand in the first place.
Biology keeps solving the hardware problem
There's another reason this paper hit me so hard.
I've been building progressively richer environments for Briain.
Locomotion.
Sensation.
Energy.
Risk.
Reproduction.
Death.
Ecology.
Every one of those becomes an engineering problem when you simulate it.
But an insect already comes with all of them.
Biology gives you:
eyes.
chemical sensors.
mechanoreceptors.
flight control.
power management.
self-repair.
navigation.
actuators.
an embodied nervous system.
and an autonomous organism approximately the size of the computer system you're trying to build.
The HI-MEMS approach basically says:
stop rebuilding all of that. Interface with it.
There is something beautifully backwards about that compared with AI.
In AI we usually recreate more and more biology in software.
Artificial neurons.
Artificial perception.
Artificial agents.
Artificial environments.
Artificial memory.
This research goes the opposite direction.
Keep the biological machine.
Only add the missing computational capability.
And that's why this doesn't make Briain obsolete
My first reaction really was:
"Jesus Christ, they were doing insect-machine interfaces sixteen years before I started playing with fly brains."
But they weren't asking my question.
And I'm not asking theirs.
If anything, discovering HI-MEMS makes the conceptual space around Briain larger.
Because there are now two independently demonstrated ideas sitting on opposite sides of the same boundary.
One says:
electronics can become integrated with a developing insect deeply enough to interact with its neuromuscular machinery.
The other thing I'm exploring is:
persistent computational memory can exist outside an individual nervous system and alter how future learners encounter the world.
One is physical.
One is informational.
The insane version is where they meet.
The sentence I can't get out of my head
My current working principle for Briain is:
Useful memory should lower the cost of learning what is still true without raising the cost of discovering what is no longer true.
I was thinking about that entirely as a computational problem.
Then I found a paper where DARPA-funded researchers were letting moths literally grow around machine interfaces so electronics could participate in their sensorimotor systems.
So apparently the universe would like me to keep going.
Great.
This is fine.
Sources
Alper Bozkurt, Robert F. Gilmour Jr., Ayesa Sinha, David Stern, Amit Lal.
Insect–Machine Interface Based Neurocybernetics.
IEEE Transactions on Biomedical Engineering, Vol. 56, No. 6, June 2009.
DOI: 10.1109/TBME.2009.2015460
Alper Bozkurt et al.
Radio Control of Insects for Biobotic Domestication.
2009.
The first paper explicitly states that the work was supported by DARPA's Hybrid Insect Micro-Electro-Mechanical Systems program.
This is part of my ongoing Life of Briain / Fly and the Graph experiments into insect neural systems, external memory, learning and what happens when information gets to outlive the organism that learned it.
