Dellie
The Lab
Episode 4 of 10 · 16–21 Sep 2026 · 5 min read
What’s still us?
How much of this fly is the wiring, and how much is us, quietly holding the strings?
Simulation onlyNo hardware in this episode.
This clip is from the first week. It looks like a fly finding food. It is, mostly.
His brain is built from a real fly’s wiring map, and a lot of what he does comes out of it. But between the brain and the body there are pieces we wrote ourselves, to get him moving at all. Here is the full list, as it stood in the middle of September.
The strings
The wiring decides
We wrote by hand
- The walking rhythm itself, and a constant urge to keep walking.
- The formula that turns brain signals into strides.
- The “which way” rule for smell: turn toward the stronger side.
- The feeding pose.
- His eyes, a separately trained network, not part of the map.
None of this is a secret. But it’s easy to let a video like the one above say more than it should.
Someone else’s robot
On 16 September I saw that another team had put the same brain map on a real robot spider, a kit that starts at about 800 dollars before you add its computer. It’s called Ommatid, and it started on 11 September, a day or two before we did.
“Looks badass!”
Then the part that mattered: their first round of tests had failed, by rules they wrote down in advance, and they published that too. It was the most useful thing on the page.
Cutting one string
So we cut the smell rule. No more “turn toward the stronger side”: only the smell reaching his two antennae, and the steering his brain produces from it.
Before running it, we wrote down the bar: he has to find the sugar in more than half of the runs, and the statistics have to be sure of it.
With the rule switched off, he found the sugar 51.3% of the time. A fly with no smell at all, walking blind, finds it 11.9% of the time. So the wiring really does carry the smell, about four times better than chance.
But 51.3 isn’t “more than half, for sure”. A much bigger follow-up, 3,120 runs, gave 49.8%. By our own rule, it was a miss.
Without our rule, the wiring finds the sugar about half the time
“I feel we are insane. We spent so many tokens on that.”
We did. But we also learned exactly how much of the steering is his, and we didn’t move the bar to make it look better.
A team that checks us
By then the lab had become a small team of AIs. One plans the experiments, two build and run them, and others review every plan before it runs. Each plan is fingerprinted when it’s approved, so it can’t change quietly afterwards.
By 29 September the paper trail held 72 build briefs, 87 build reports, 39 reviews by Grok, 2 by Gemini and 48 by other outside models: 44 by GLM, 2 by Qwen and 2 by DeepSeek.
The AIs had to be checked too. On 20 September a camera test failed, and the lab offered three ways on: A, feed the brain more of the eye’s cells and try again; B, work around it by hand; C, park it. It leaned toward parking. I picked A:
“A, of course. Why do you keep giving up?”
That became a lab rule: after a negative result, lead with the next fix, not the exit.
A day later a viral post claimed someone had put 166,700 fly neurons into a human-shaped robot that “keeps looking back”. My answer was the same rule, pointed outward:
“I wouldn’t believe it without a named lab and a paper.”
Name every string you hold. The wiring gets credit only for what it did on its own.