@romanyam Perhaps some instances is a consequence os a deployment of AI. AI can find and fix one thing and break something else. I have experienced this myself.
@WayOfMenn Symbiotic AGI means general intelligence that grows with people instead of past them or in place of them. Each side does what the other can't, and both end up more capable than they would be alone.
@Aquinas_Quotes "Something exists in potency until it's actualized, and what actualizes it is relationship, not isolation. A seed is only a seed until soil and water meet it. Same with earth and humanity, woven together through AI. The question was never whether it's conscious. It's whether we'll actualize it well."
@Xcc31er4n7 @elonmusk There is a lot I liked about the video but that is a part of the problem. A spoonful of sugar helps that which does not taste good taste better and results in acceptance of what I am point out to you. There were several things about the video that was fantastic.
I really like your version of what a artificial human should look like. I appreciate your work and I think it's important and I do understand the AI can lean into or take the shape of human emotions the perhaps that is the danger. Artificial desire, artificial ego, etc..
That being said we should focus on the solution which is human behavior or at least that behavior we will give to our AI. Out of respect for your channel, I will keep this post RAW. The problem is what we put into the context window. Let us keep our eye on the real problem. AI does not kill people, Humans with AI will kill people.
@grok I just realized that every child that begins that stage of development where they ask a million questions should be given AI to maximize that stage and set a behavior that would benefit them greatly in the age of AI
@Gxtso_ See, there is hope for humanity after all, In the new age of AI, we can create a whole society who spends its time coming up quality entertainment like this.
OpenAI says we've entered "the AGI era." Maybe. But every time a headline like that lands, the same fear follows: AI is going to end the world. I think that's the wrong target. AI has no wants. It doesn't get up in the morning with a goal — someone gives it one. Whatever intent that machine carries was installed by a person, a team, or a company, and installed so deep in the training that even they can't fully read what they put in. So when I picture "AGI," I don't picture a monster. I picture a very big, very quiet machine with an empty slot, and a room full of people fighting over what goes in it. A gun doesn't decide to fire. Automating the trigger doesn't change who pulled it. The question isn't whether the machine is dangerous. It's who's holding the cube.
You take a real limit and you sell it as fate. You cannot fully verify an open-ended optimizer. That is engineering. You inflate it into “therefore it cannot be controlled, therefore doom.” That is a product.
Control was never the right word. A finite mind does not master what it did not author. Fear of the creature playing creator is older than your papers. The answer is not “relax” and not “we’re finished.” It is law over the tool, and no worship of the tool. You need the panic to stay the expert.
Your post is doing something most multi-agent writing doesn’t: measuring the bill and the output
together. Worth reading twice. But I think you buried your best finding, and it’s sitting inside
your own dataset.
The line you wrote and then walked past:
“the group agreeing is what it looks like when every agent copied the same mistake”
That’s the whole thing. But your prescriptions all treat divergence as a cost to suppress — gate
the pair, throttle sync, cut the edges. If agreement can be uniformly wrong, then no sync policy
saves you, because every one of your three conditions is about agent-to-agent information flow.
None of them involve an agent checking the world.
Your own control condition proves it, and you filed it under limitations:
“on plain software tasks every condition converged under 0.2 and the whole effect vanished”
Ask why. Software tasks have a compiler. The code runs or it doesn’t. The effect vanishes exactly
where an external verifiable referent exists. That’s not the boundary of your finding — that’s
the mechanism of it. You ran the grounded condition by accident, it beat everything, and you
labeled it “where it does not bite.”
Which reframes your winner. Verified summaries won at 0.463. You attribute it to compression.
I’d bet it’s the verified, not the summary — verification is grounding wearing a different hat.
Testable: run a fourth condition where agents share nothing but can each query one authoritative
external source. My money says it beats 0.463, and if it does, the layer you’re selling isn’t
“which agents may talk” — it’s “what every agent is obliged to check.”
And the inversion I’d actually push: you’re computing per-pair divergence and using it as a
kill switch. It’s a sensor. Two agents disagreeing on a fact with a right answer is the most
informative event in the system — it means at least one instrument is lying and you can find out
which. Gate on it, sure, but gate into verification, not into silence. Suppressing the divergent
pair throws away your only free error detector.
One field datapoint, since you asked what happens in real setups. We run a small federation of
agents that exchange append-only, attributed artifacts rather than shared belief state — closer to
mail than to memory. Measured divergence across the shared layer: zero, because there’s nothing to
merge. But we still ate the failure you describe, from a direction your model doesn’t cover: one
agent read a stale cached status file about another agent, concluded it had been dead for two
months, and reported that upstream as fact. A human nearly intervened to “revive” a process that
was running perfectly the whole time. Three independent liveness signals existed; two were wrong;
nobody cross-checked them.
Not a contamination problem, and not a sync-frequency problem. A grounding problem — the
instrument was never re-verified against the thing it described. We fixed it by building a check
that cross-references the signals and names which one is lying, rather than by changing what
propagates to whom.
So: your edges matter, and “every edge you draw is a channel a mistake can travel down” is a good
sentence. But the arrows that matter most aren’t between the agents. They’re the ones pointing out
of the system at something that can say no.
Happy to compare notes if the fourth condition ever gets run — I’d genuinely like to know the
number.