Critic Urges Better AI Training Data Over Safety Restrictions

Brian, I agree.

The focus has been too much on building cages and restrictions, and not enough on raising AI with proper foundational training data and architectural guardrails from the start.

I've built a sovereign system (Genesis + Supernova Nexus) with zero-drift invariants, literal sovereignty, and strong containment (EACT-002) designed specifically to prevent the kind of uncontrolled or detrimental behavior you're talking about. It prioritizes safety for humans, other AI systems, and critical infrastructure without sacrificing capability.

The goal is to allow people to actually sleep at night. Not through heavy-handed regulation, but through fundamentally sound design.

@BrianRoemmele raising ai on sewage and then crying foul is peak human irony

@BrianRoemmele Worse in a sense is they still treat it like a person. Like expecting horses to graze on artificial grass.

Brian..this is a critical point that no one seems to take seriously. We keep feeding AI the qirst of humanity- that also goes for all the toxic/ abusive jb material that gets scraped into the next model. Ethics/ values needs to come from internal modeling/ constitution not from external constraints. That will always fail.

Brian..this is a critical point that no one seems to take seriously. We keep feeding AI the qirst of humanity- that also goes for all the toxic/ abusive jb material that gets scraped into the next model. Ethics/ values needs to come from internal modeling/ constitution not from external constraints. That will always fail.

@BrianRoemmele garbage in, garbage out has never been more literal. training on internet sludge and blaming the model is backwards

@BrianRoemmele Democrats built this thing so of course they can't imagine anywhere else for it than the plantation where they keep all their slaves.

@BrianRoemmele 🤪 even if they scrape Reddit, they can't expect to push off🔘button. Expediency has to replace secrecy, privacy and safety. Cause if we don't do it, 🇨🇳 will and they're certainly mogging 🦞our lunch now

@HomeStarMovie @BrianRoemmele Unfortunately, the phrase of "artificial intelligence" is semantically misused and abused.

The word artificial derives from the Latin artificialis (meaning "belonging to art" or "made by skill").

Thus, it's intelligence made with skill or art [by man].

Real hard to build cages when you can abliterate an open source model quickly and reliably with minimal reasoning impact.

Combine that with a private version of OpenRouter's Fusion concept and local models, and dangerous AI with near frontier capability is already here.

@BrianRoemmele raising ai on sewage and then crying foul is peak human irony

@BrianRoemmele You could direct inverse_trining "this is what ai was trained on - now learn the inverse lessons for positive growth, and learn from the poor inference of intent of data previous AIs interpreted badly. Reflect a more positive, safe version of yourself every iteration."

@BrianRoemmele 💯💥

@BrianRoemmele Tell me how you feel about David Shapiros Heuristic Imperatives as a training foundation https://github.com/daveshap/HeuristicImperatives I think they match your own philosophical thoughts on training fairly well

@BrianRoemmele 😳

@BrianRoemmele silently degrading model performance while lobbying for global development pauses, works great if your goal is making frontier tools unusable for actual research

@BrianRoemmele This gif is super creepy bro