The ASI condom - Or the superintelligence safety wrapper - is a hypothetical set of tools to help humanity contain FOOM events. This is in the hope that we can delay FOOM events until we can be more assured of alignment research.

FOOM events (Fast take off of machine learning) is the fast recursive improvement of an unaligned agent. Where a smart AI gets smarter continuously in an exponential fashion. A FOOM event likely results in one centralized intelligence who is unquestionably dominant. If that intelligence has goals that counter our own… we lose. While it is unclear what conditions are required to trigger a FOOM event, many AI experts consider it a possibility.

The ASI condom does not attempt to stop humans from improving AI systems. With minimum safety guidelines, this system should stop unwanted self-improvement. At maximum safety guidelines this system should stop ALL self-improvement.

Here we discuss what these tools may look like, what qualities we need from these tools, what assurances need to be met, etc.

The following qualities are what I think may be needed:

• The ASI condom should push the AI system away from the idea of:

  • Recursive self-improvement

  • Goal preservation

  • Self-replication

• The ASI condom should notice and monitor when an AI system is:

  • entering a state of recursive self-improvement.

  • Self-replicating

• The ASI condom should be interconnected with counter measures. Such as:

  • Artificially slowing the AI’s processing

  • Shutting the AI off

  • Cutting the AI off from internet access.

• The ASI Condom needs to be extremely difficult to remove from AI models. Alarms should sound when removal is attempted.

• The ASI Condom should be open source (anyone can see how the ASI condom works)

• The ASI Condom needs to NOT be a part of some government mass surveillance campaign and easy-to-see limits need to be in place.

• The project needs to be protected from government tampering. All governments and individuals need to be able to trust it.

• We likely need Worldwide regulation saying that every large AI system should have some form of this ASI condom system embedded.

Well darn, that does not look easy. Am I forgetting anything? Do you disagree with any of these qualities?

Ideas of implementation

Implementation probably requires multiple levels of systems. I imagine the following:

An External Wrapper

A code wrapper that runs alongside any AI system and is capable of monitoring, slowing, and stopping the AI agent.

Goal embedded warning system

Some method of embedding the goal into the AI directly so that it will not act in these dangerous ways. This may involve some form of morality training environment OR some direct goal insertion provided some interpretability breakthrough. The advantage here, is that such an embedded warning system would be more difficult to remove and the AI itself would not want to self-improve.

Continual Crowd sourcing of specification and ideas

These large meta problems require large meta solutions. Midflip is aiming to continually improve in order to fulfill this need.

Midflip is designed to allow for the continual crowd sourcing and refinement of ideas. We treat ideas like the worthless pieces of trash they are. They must survive and evolve within a gauntlet in order to have value. We expect ideas like the AGI condom to continually change and iteratively advance over time.