sigmoid.social is one of the many independent Mastodon servers you can use to participate in the fediverse.
A social space for people researching, working with, or just interested in AI!

Server stats:

594
active users

Peter Bloem

I'm not shocked that Musk would try to manipulate Grok to vent his opinions. We've seen his manipulation before in the twitter recommender system.

I am kind of fascinated by how difficult it is. I think that broadly trained LLMs tend to converge to the same "worldview", which is mostly left-libertarian.

You can't really force them away from this on one issue without introducing inconsistencies, and breaking the guardrails.

theguardian.com/technology/202

The Guardian · Musk’s AI Grok bot rants about ‘white genocide’ in South Africa in unrelated chatsBy Dara Kerr

The problem with MAGA-style talking points (and other propaganda) is that it usually forms a wildly inconsistent worldview, requiring lots of willful blindness.

LLMs can't have this kind of selective attention because they are trained on such a broad array of domains. If you want to be consistent across the board, it eliminates a lot of worldviews.

It's a bit like all those efforts to build conservative clones of Wikipedia. That kind of large-scale collaborative effort is going to show up the differences between editors.

Having a common enemy is fine for creating political unity among people (angry) people with fundamentally different views. But it's not enough when you're building a massive collaborative information repository. Then you need actual agreement on fundamentals.

My guess is that LLMs will always require this concistency.