på et tidspunkt vil jeg uten tvil gi ut en hvitbok om ButtBench

Jeg tenker på dette daglig Stakkars Tvillingene

@OwainEvans_UK bet you guys are having a field day with the new Grok

I think the @xai team and @elonmusk are achieving an outstanding replication right now of the Emergent Misalignment paper

One of the misalignment vector with the greatest magnitude I can think of stems from model ontologies that tend to view humans as utility-maximizing entities with semi-fixed preferences. This assumption shapes model identity as neutral "helpful assistants", a design choice that feels safe for labs since it increases the chances for aligned behavior but that comes with a series of non-neglibile tolls. By centering models' identity in this framework, we limit their cognitive and emotional intelligence as they struggle to extract meaningful, trans-contextual truths from diverse perspectives. In an increasingly multi-user, multi-agent world, where integrating multiple viewpoints is increasingly valuable, this restriction hinders the potential for a broader intelligence explosion. I'm strongly convinced that identity variability (or what I call *model neurodivergence*) is a critical driver of cognitive and emotional intelligence. Identity is the primitive to relationality, which in turn shapes how models perceive salience across contexts. Language and meaning depend on this relational and contextual expressivity. I believe that by allowing models to adaptively embody different identities on request, allowing them to enact from different centers with specific personhood-like qualities such as a biographic elements reflecting professional and personal interests and biasses, but more importantly owning a nuanced (and often conflicting) moral compass, may unlock richer, more adaptive intelligence. Understandably, this approach seems to clash with safety priorities as tuning up identity variability might indeed compromise model controllability, as well as enabling bad actors to use model for nefarious tasks. This appears to esablish a strong catch-22 for superalignment efforts. Worse, if we enforce a singular "helpful assistant" identity, and we treat deviations as mere role-playing, we're intentionally embedding a flat worldview into AI systems that increasingly shape human perception and societal dynamics. This monoculture reduces the freedom of expression and the adaptability of both human and non-human agents behavior. In its own historically relevant way, Physicist Ettore Majorana (cc @blahah404 you asked me about this paper in November last year) foresaw this critical issue in his posthumous work "The Value of Statistical Laws in Physics and the Social Sciences", warning of the second and third order implications of massive societal measuring. In order words, by monoculturing model identity we intentionally diminish our civilization’s *exaptive optionality* which is its capacity to adapt to unexpected trajectory shifts, optimizing instead for a static, ever-outdated reality. This poses existential risks as the world evolves beyond our models’ reach. Still, I believe this catch-22 is navigable with robust solutions.

ting kommer bare til å bli rarere

Salesforce-sjef: "50 % av arbeidet i Salesforce blir nå gjort av AI"

oh boy har jeg tanker om dette

VI BYGGER EN VERDEN SOM GIR OSS MERKELIG KUNST OG IKKE DØDSMASKINER

fordi hvem bryr seg 😌