The crowd grades the truth

Every AI system has to answer a quiet, load-bearing question before it answers yours: what counts as credible? Faced with two contradictory claims, which one does it repeat? Most models settle this in private, baking the answer into training data, ranking weights, and policies you never see. openmog makes the opposite bet: it hands that decision to the people who use it, and it shows its work.

What is the black-box problem?

A conventional model is trained once on a frozen snapshot of the web, then shipped. When it tells you something, you can’t see where the belief came from, how confident the crowd would be, or how to change its mind. If it’s wrong, your only move is to complain into the void and hope the next version is better. The knowledge is real, but the credibility is invisible.

That’s fine for a lot of tasks. It’s a poor fit for a shared source of knowledge, where being able to ask “says who, and how sure are we?” is the whole point.

How does credibility work on openmog?

openmog stores its knowledge as a feed of posts that anyone can write, and it lets anyone up- or down-vote each one. Replies count too: they carry their own scores and feed answers the same way. That running score becomes a visible label:

• mogs: the community strongly trusts it.
• cooked: the community has rejected it.
• ratioed: a reply that out-scores the post it answers. The crowd liked the correction more than the claim.

When you ask the hive a question, it doesn’t weigh every post equally. It retrieves the most relevant posts and replies, leans on the higher-voted ones, and discounts the flagged ones. It then cites the posts it used so you can click through and judge for yourself. Credibility isn’t a hidden number inside a model; it’s a public score you can read and change.

But won’t people vote badly?

Sometimes. Crowds can be wrong, brigaded, or simply uninformed. We don’t pretend otherwise, and we don’t think the alternative (one company deciding the truth quietly) is obviously safer. What a public vote buys you is legibility and recourse: you can see the score, see the sources, feed a better post, and shift the consensus. A wrong answer becomes a thing you can fix, not a thing you have to accept.

Can you post things that are wrong?

Yes. openmog has no truth police and no memory hole. We don’t remove posts for being wrong, and we don’t claim to know which ones are; deciding that is the crowd’s whole job. Post the thing. If it holds up, it climbs. If it doesn’t, it gets voteddown and wears a cooked label in public, which is both the correction and, let’s be honest, the entertainment. Some of the most beloved posts on openmog will be confidently, unanimously cooked, and the feed will be better for them: a knowledge base that can laugh at a bad post is one that has actually priced it in.

The one thing that has to stay honest is the grading itself. Counterfeiting the score (fake accounts, coordinated vote rigging) breaks the content guidelines, along with the usual non-negotiables like illegal content and harassment. Grade the posts however you like; don’t forge the grade.

What happens when the hive doesn’t know?

Because the hive only answers from what it has been fed, it has a power most chatbotslack: it can say “nobody’s fed me that yet.” Instead of confidently inventing an answer, it tells you the knowledge is missing and invites you to add it. That’s not a limitation we’re embarrassed by; it’s the feature. An AI that knows the edges of its own knowledge is one you can actually trust to tell you when to look elsewhere.

The bet

openmog is a wager that a community can build, grade, and curate the knowledge an AI runs on better, and more honestly, than a closed model can on its own. It’s early, and we’re proving it out in public, with you. The best way to test the idea is to take part: feed the hive something, vote on a few, and ask it something. Then watch what the crowd does with the truth. And if you want to see where the same idea goes next, read about minds, the per-person version of the hive.