Tomorrow Tuesday — September 2026 (Issue #9)
Welcome to another issue of Tomorrow Tuesday, where we cover the revolution running through film, video games, music and the creative industries.
A newsletter by SOUTS, where we explore what happens when artificial intelligence stops being just a tool and starts changing the way we create and consume stories.
Stop the AI, I want to get off
You’ve heard plenty already about whether AI is going to kill us, whether it needs to be regulated, and so on. So let’s sum up everything that happened last week with emojis.
📢 Last week an Anthropic researcher resigned. Their warning: there’s a chance AI goes terribly wrong and nobody is doing anything about it.
😴 The problem is we’ve been hearing versions of this for years. Seven years ago OpenAI decided not to release GPT-2 because it was “too dangerous.” The thing is, Silicon Valley figured out that sounding the alarm brings in more investment money. VCs want to invest in the machine that could destroy the world.
⚠️ But on Saturday Dario Amodei (Anthropic’s CEO) published a blog post, We Must Pace the Frontier. He says AI has started participating more and more in building the next generation of AI, and that we need to slow down the rate at which we improve AI.
That’s called recursive self-improvement.
Amodei’s concern is fairly simple: the speed at which these systems improve is outrunning our speed at understanding and controlling them.
⚠️⚠️ Who agreed is what caught people’s attention. Sam Altman (OpenAI) said he agrees and that OpenAI will allow independent evaluators. Elon Musk agrees. Demis Hassabis (Google DeepMind) agrees. Satya Nadella (Microsoft) agrees.
😡 But the big opponent showed up: Donald Trump. His argument is the great argument against regulation: fear of China. If the United States hits the brakes and China doesn’t, then what?
As usual, Trump posted on Truth glorifying himself and saying he’s the only one who can stop AI.
🇨🇳 But what about China? There’s still been no communication from the country or the party. The understanding is that scientists in China have the same fears, but they don’t want to lose the AI race either.
Everyone is playing chicken.
Dario says in his blog post that the effort has to be international and not just from the US, and that seems to be the hardest part.
❓ What does “slowing down” mean?
The plan proposes three things:
- External auditors inside the labs, with access almost equivalent to that of the internal safety teams.
- Common rules among the companies at the frontier, including limits on how far they can go without passing certain safety tests.
- International coordination, ideally including China, for the truly dangerous capabilities.
This story is changing day by day, so we’ll see where it goes.
They taught a fly to play Doom
I’m pretty worked up about this story. Stay with me, because it says a lot about the big science-communication problem that social media makes worse.
It all started with news from Google Research and HHMI Janelia presenting a map of the brain and central nervous system of a fruit fly: 166,000 neurons, 125 million synaptic connections. That map of connections is called a connectome.
It’s a reconstruction of every connection from microscope images, processed with AI tools and reviewed by specialists.
It’s a tremendous neuroscience experiment.
Second act: along came DOOMFLY, a project connecting a simulation based on that connectome to a Doom environment. Game frames get turned into signals for simulated neurons, and the resulting activity is translated into movement and shooting.
So:
- They reconstructed the brain.
- They turned the brain into a mathematical model. Numbers go in, things happen inside.
- They feed it images, some parts of the brain switch on, others switch off.
- We read the parts that switch on and translate them into actions.
- They put the brain inside Doom and set it to play. When the fly sees demons, it shoots them.
Viral.
From there we started getting the fly playing Beat Saber, the fly inside Minecraft, the fly controlling a real physical robot. All viral.
And this is where I get worked up.
I said: I’m going to train my own fly for the newsletter.
I started opening up the projects going around. Almost none of them had published code, and that’s where I started getting suspicious. I went through the ones that did have code.
There are three categories of project:
- Pure smoke.I won’t name them, but there’s one that promises to connect a fly brain to a physical robot. You open the brain file and every function says “this doesn’t work, for now it’s hand-made for the Twitter video.”
- Real projectsthat, because the fly’s vision fails, decide to fake it by stimulating the neurons by hand. They pair that with another mapping they made two years ago, this time of all the fly’s joints: Flybody.
- Real projects where the results are bad. I tried some of these. I installed one, downloaded the 258 MB connectome and ran it on my machine. You stimulate the threat-detection neurons on the left side and a group of neurons lights up. You stimulate the detection neurons on the right side and a different group activates.
The problem with the third category: the vision part either isn’t understood or doesn’t work, because in the best case, with two buttons, the fly pressed the right one 60% of the time (vs 50% from chance alone). The fly never won a match. In fact, in DOOMFLY the fly killed the same number of demons with vision off as with it on.
The problem here is that you can upload a file saying “we tried to teach a fly to play DOOM and we couldn’t,” or you can upload a faked but pretty video of the fly playing. Guess which one goes viral.
So I decided to do the second one:
I taught a fly to do the 6-7
I took FlyBody, an extremely detailed physical model of a fly developed by DeepMind and Janelia and published in Nature in 2025.
FlyBody doesn’t have the fly’s brain, it has the body.
Legs, wings, joints, adhesion to the floor, physics, gravity.
To walk, you have to control 59 dimensions — meaning you can’t animate it by hand. You have to tell it what to do with each of the 59 muscles, five hundred times a second, without falling flat on its face. Nobody can do that by hand. You have to teach it.
The goal was to do the 6-7 with the front legs without falling over.
So we built a gym. 1,024 flies practicing at the same time, each one trying minimally different variants, and a jury that grades every attempt:
- Did the hands get where they were supposed to get?
- Did it stay upright?
- Did it stay in place without walking?
- Did it fall?
The ones that score best pass what they learned to a shared brain, the brain corrects itself a little, and they start over. Thirty-two million times.
This is the result:
For now we have an extraordinary map, a pile of honest experiments that don’t work yet, and a brain we still understand very little of.
You don’t need to sell it as a digital consciousness for it to be worth looking at.
That was Tomorrow Tuesday, the weekly newsletter. If you want to subscribe, subscribe!
At SOUTSwe try to always stay on top of what’s happening, and every time there’s something relevant we’ll bring it to you in this newsletter.
See you next Tomorrow Tuesday.
Yves
