Tiling Agents

TagLast edit: 29 Feb 2024 19:02 UTC by steven0461

An agent might have the ability to create similar or slightly better versions of itself. These new agents can in turn create similar /​ better versions of themselves, and so on in a repeating pattern. This is referred to as an agent tiling itself.

This leads to the question: How can the original agent trust that these recursively generated agents maintain goals that are similar to the original’s objective?

In a deterministic logical system, assuming that all agents will share the same axioms, “trust” arises from being able to formally prove that the conclusions reached by any subsequently generated agents will be true. The possibility to be able to have this form of trust is influenced by Löbs theorem. The inability to form this trust is called the Löbian obstacle.

See Also: Löbian obstacle, Löbs theorem, Vingean Agents, Vingean Reflection

References :

Tiling Agents for Self-Mod­ify­ing AI (OPFAI #2)

Eliezer Yudkowsky6 Jun 2013 20:24 UTC
88 points
259 comments3 min readLW link

Walk­through of the Tiling Agents for Self-Mod­ify­ing AI paper

So8res13 Dec 2013 3:23 UTC
29 points
18 comments21 min readLW link

The Löbian Ob­sta­cle, And Why You Should Care

lukemarks7 Sep 2023 23:59 UTC
18 points
6 comments2 min readLW link

Tiling agents with trans­finite para­met­ric polymorphism

Squark9 May 2014 17:32 UTC
6 points
11 comments2 min readLW link

Para­con­sis­tent Tiling Agents (Very Early Draft)

IAFF-User-42 Apr 2015 7:27 UTC
8 points
5 comments1 min readLW link

Vingean Reflec­tion: Reli­able Rea­son­ing for Self-Im­prov­ing Agents

So8res15 Jan 2015 22:47 UTC
37 points
5 comments9 min readLW link

The al­ign­ment sta­bil­ity problem

Seth Herd26 Mar 2023 2:10 UTC
24 points
10 comments4 min readLW link

Vingean Reflec­tion: Open Problems

abramdemski3 Jul 2015 18:44 UTC
3 points
3 comments5 min readLW link

Higher Di­men­sion Carte­sian Ob­jects and Align­ing ‘Tiling Si­mu­la­tors’

lukemarks11 Jun 2023 0:13 UTC
22 points
0 comments5 min readLW link
No comments.