The Intelligence Age
Article

Better than debate

published on · updated on

The previous article in the sequence is Downsides of debate.

In an earlier article, I pose questions about various interaction patterns for truth-seeking other than debate as described in AI safety via debate1.

The truth-seeking processes I recommend need not be timid; I envision a dynamic where agents press each other. They aren’t trying to beat the other; that would be too low a bar. Instead, they are aiming for correctness. With the appropriate ground rules2, I expect agents can use the other party as a resource for learning rather than an opponent to beat. This is a huge difference from most debate frameworks.

One way to assess an incentive structure is to ask: given the rules of the interaction pattern, do we see a productive progression. Consider the following quote which suggests what beneficial behavior looks like:

Comments should get more thoughtful and substantive, not less, as a topic gets more divisive. 3

What does an “increasingly thoughtful and substantive” process look like? I won’t attempt to pin this down here and now, but for now I will say this: truth-seeking participants often do well by starting by finding points of agreement and disagreement. This is a rational strategy because it allows the agents to narrow down a potentially large discussion tree early on. As the discussion ensues, particular sub-claims are evaluated and refined. Over the course of an interaction, some claims can be tightened up as probabilistic estimates with uncertainty.

Endnotes

1

G. Irving, P. Christiano, and D. Amodei, “AI safety via debate.” arXiv, Oct. 22, 2018. doi: 10.48550/arXiv.1805.00899.

2

I have not yet written down such ground rules.

3

Quoted from the Hacker News (HN) Guidelines https://news.ycombinator.com/newsguidelines.html. This reference does not constitute a endorsement of the HN guidelines generally. In previous writing (not on this site), I’ve written more on the topic of the goals of collaboration and how seeking different objectives is best done with tailored group norms. I hope to revisit and revise that writing and repost it here.

4

To contact me, here is an easy puzzle to solve. My provider is an obvious choice that starts with a “g”; my prefix fits the pattern “firstmlast”. My middle initial is same as the first letter of the last name of the second listed author, above.