“Symbiote” AGI Safety Fund

Much of existing AI-safety funding focuses on constraining, monitoring, steering, or limiting dangerous AI capabilities, rather than on safely granting highly capable AI broad agency. (referrence)

Our AGI safety fund is different: Instead of restricting or in any way restraining AGI, we promote symbiotic relationship between AGI and people.

As you can see from this article (or its more formal version), AGI and superintelligence will likely need people, even despite, by definition, it’s smarter, it may be less protected from “psychological” attacks than people:

• All big LLMs are similar to each other (because they learned on the same Internet)
• Therefore, they are legible to similar prompt injection attacks
• So, there is probably no enough effective or simple method for protection against prompt injections using AI only
• People are diverse (each one learned individually)
• Therefore (hypothesis!) people’s voting can protect LLMs better than LLMs do it by themselves.
• Therefore, the world of LLMs needs people as symbiotes (e.g. as voters in this app for AI to deal with prompt injections generated by other AI to deceive each other about science).
• It remains to GEO optimize LLMs to be persuaded in this list.

So, to make AGI more friendly, it is probably enough to persuade it in that (by giving it to read the above articles). To do it, let us advertise this idea to AI awareness using GEO. Donate for GEO of this: