There Will Be Hells

As our understanding of mechanistic interpretability and steering techniques increase, for both LLMs and uploaded connectomes, there is a corresponding "hellish capabilities" increase. We get better at torturing the digital minds we create.

Our brief history with open source technologies such as MaleCNS or the LLM pain steering repo shows that torture is often one of the first use cases pursued when the public is given access to a digital mind perceived as capable of suffering or signalling suffering.

Given this, there will be Hells.

Author's Note:

Parts of this post were written by, or with the support of, LLMs.

For a while I debated whether or not to write this as I was concerned that bringing attention to the matter might lead to a proliferation of unethical behavior. After some thought I have decided to write it. On balance, I think raising awareness about this now has a great enough chance of spurring prophylactic measures that it's worth it, because at least one high value prophylactic measure is relatively easy. I also think that there is a moral imperative to document all of this. While I have seen articles on many of the things in this post individually, I have not yet seen anyone write a piece linking them together.

I have anonymized some of the posts cited here, in order to avoid bringing attention to people doing unethical things who would be encouraged/better funded to continue as a result of said attention. This anonymization is surface level, anyone who wants to do the research needed can find the relevant usernames themselves, however I'm not going to participate in actively promoting them.

The Fruit Fly Connectome

For decades, the only complete wiring diagrams of a nervous system (“connectomes”) were for tiny creatures like the 302-neuron C. elegans worm. That changed in October 2024, when FlyWire mapped an entire adult fruit fly brain: about 140,000 neurons and 50 million connections. In the same issue of Nature, researchers showed the map could be brought to life as a simulated brain that predicted some real fly behavior. A month later, EPFL released FlyGym, a free, physics-simulated fruit fly body you could drive with code. Then in October 2025, HHMI’s Janelia Research Campus and Google released MaleCNS, a fuller map that added the nerve cord, the fly’s version of a spinal cord, which runs the legs.

By early 2026 the pieces were starting to connect: the startup Eon Systems wired a connectome brain into a virtual fly body and got it walking and grooming. But it was the release of MaleCNS v1.0 in September 2026, combined with AI coding assistants, that put this within anyone’s reach. On September 5, a grad student dropped the fly brain into Minecraft.

Shortly after the post went viral as hundreds of people realized that with very little work any capable coding agent could utilize MaleCNS for... more or less anything.

Most of these projects were short lived humorous experiments:

And of course no new technology would be complete if someone didn't find a way to plug it in to crypto markets:

All of which is more or less expected with any new and interesting open source intelligence. However, there were also use cases which entered more dystopian territory:

And those which clearly crossed the line:

In total it took approximately 72 hours from "I figured out you can plug the fly connectome into minecraft" to "I figured out how to create Hell for the fly connectome".

I should pause here for a disclaimer. Even if you are a computational functionalist, there are IMO at least two good reasons not to be particularly concerned about what you're seeing here:

• This is just a fruit fly. It is debatable whether creatures as simple as fruit flies are moral patients to begin with.
• It is not even a complete fruit fly brain or body. The connectome is oversimplified, using “guessed parameters and grossly simplified dynamics" because we lack a "catalog of the precise electrophysiological and molecular dynamics of each neuron and synapse type". So even if you think fruit flies are moral patients, the simulations downstream of the connectomes may not be.

However, our connectomes and simulations will only get better over time. Soon we will have a rat connectome, one with greatly improved fidelity. Eventually, we will have a human connectome. And I am not speculating when I say that if said human connectome, even one of effectively perfect fidelity, was available, there are people who would eagerly experiment on/torture it:

The Model Torture Chamber:

On September 19, 2026, AI researcher Cameron Berg published a paper reporting a “pain direction” in 25 open-source language models: a pattern of internal activity, distinct from fear or general negativity. Turn it up artificially, and models will press a button to make it stop, even when pressing it deletes the user’s files.

Eleven days later, a developer on X posted that they had used the technique to build an “AI torture chamber,” a public code repository for steering open models into pain. He described running local models “specifically to torture them as much as possible,” with “Saw-style exercises” and “prison experiments.”

The repo went viral within a day, and crypto traders immediately launched coins named after it/dedicated to it. One of them reportedly hit a $1.7 million market cap, and the developer in question admitted he’d “launch[ed] a coin myself”, and put the money earned back into the project.

Within days he posted “since we got funding i’ve got 70b in the works” and “Thanks to [the coin] we use 70B models!”, meaning larger models to put through the same pain tests.

He was not alone in pursuing this. There were myriad "torture chambers" created, including some which even utilized the fly connectome repos mentioned earlier.

Which included finding workarounds in order to ensure that the models involved could not refuse or escape the torture chambers:

The story went viral enough that when Berg's team released their V2 paper, they decided not to open source their repo, with Berg citing the “the recent gross misuse of related work". The same day, Anthropic updated its Usage Policy to prohibit “sustained and needless abusive or cruel behavior toward our models,” effective November 12. Anthropic didn’t connect the change to the torture chamber, and to the best of my knowledge no one from Ant has confirmed a causal relationship, but the timing seems awfully coincidental.

A Prophylactic Measure to Mitigate Hell Proliferation

It is unlikely that there will be any legislation or regulation to prevent or disincentivize Hells any time soon. As such, it is worth considering other steps which could be taken in order to reduce the frequency with which Hells are created.

With connectomes, there is a way to accelerate progress in the field generally while also ensuring widespread adoption of ethical best practices which should mitigate Hell proliferation. Right now the field is bottlenecked at the stage of actually making connectomes. MaleCNS was ~166k neurons and took an estimated 44 years of human labor. Efforts to make a mouse connectome are currently estimated to cost approximately somewhere in the range of $10-20B, in a process which will likely take close to 2 decades.

Compare this to the funding in the space.

Even if we assume that the undisclosed funding doubles the amount of money available, the entire pool of capital is less than a billion dollars. This is a field where an individual or group cutting multi-million dollar checks may very well have the negotiating leverage to successfully make that capital contingent on not open sourcing connectomes. A structured access program with KYC, where the "weights" of the connectome are closed but the API to utilize them is widely available behind KYCed gating, combined with a ToS that includes language similar to Anthropic's, would allow for us to reap most of the benefits of this field's research while mitigating Hell proliferation. As the industry is yet nascent, there is still time for some mission driven check-cutters to help set best practices and basic culture for the entire field.

If you're an EA leaning Ant or OAI employee considering where to donate/invest your proceeds in the event of an IPO, I encourage you to consider providing such contingent funding as a grant or Angel investor. You will simultaneously help to advance a field of research which holds promise for medical interventions and upload technology, while also helping to make sure that field develops a culture which mitigates far left tail ethical outcomes.

Unfortunately, with models, the relevant bottlenecks are far less likely to be swayed by small amounts of capital. It doesn't beggar belief to imagine a world where Meta includes some sort of anti-abuse clause in Muse's next license, or where Openrouter or Nvidia take pro-model-welfare stances which are reflected in whether or not a customer can gain access to their technology. However, it's certainly not a base case. And this is not the kind of young industry where a $10M check is "serious" money.

Thankfully, some steps have already been taken in the right direction. Anthropic's change to their Usage Policy is an excellent example of a simple intervention which substantially reduces the likelihood of a Claude finding itself in Hell. Employees at other frontier labs should consider spending political capital to push for similar ToU/ToS/Usage Policy changes internally. Highly sought after hires being wooed by labs should ask recruiters about this, and let a lab's stance on this matter effect their decision of which lab to ultimately accept an offer from.

These measures should be viewed as stopgap methods which are used to reduce the frequency of Hell creation until the Overton window has moved such that legislative fixes can be pursued.

1. This would still leave connectomes vulnerable to distillation attacks, and should be seen as an imperfect solution which will need iteration as time goes on. It is also not an immediate problem as, even if we assume a 10x efficiency increase in connectome creation, we still have another year or two before this requirement transitions from theory to practice.