yavuz

I make AI behave, provably. Evals + Governance for frontier models in regulated industries, turned into decisions businesses can act on

[yavuz @shipbehaves]

@sierracatalina

[yavuz @shipbehaves]

That’s cool and the shape I’d want. The failure mode to keep an eye on as this scales is the clean-but-wrong posting not the flagged ones yeah, the ones that should’ve been flagged and weren’t I mean the false-negative rate on the HITL trigger is the number I’d watch over time. Cool work!

[yavuz @shipbehaves]

@HunterReveur @benhylak

[yavuz @shipbehaves]

@petergostev 5 stars. Clean merge, no conflicts, approved

[yavuz @shipbehaves]

@pawjast

[yavuz @shipbehaves]

@20min Mais non c’est trop triste..

[yavuz @shipbehaves]

“Mythos will look dumb in 6–12 months” reminds us how fast we move and why this isn’t about the model or its quality. Respect the work and lean into the defender first framing. Finding scales with the model and fixing scales with the org, and the defensive gap is organizational, not model-shaped, so I hope more professionals in the field get tools like these in their hands and help close it.

[yavuz @shipbehaves]

Can't see their internals but two things in the story are the tell, ~20% adoption and a 'rogue Claude group' that formed on its own. Grassroots found a use that fit actually but the org kept swapping the tool.. Adoption follows the work, not the logo and that's the layer under "change management" and one of the common reasons why most GenAI pilots stall, because the root cause is the integration, not the model or the quality. The model was never the variable..

[yavuz @shipbehaves]

@HarperSCarroll And it cuts the other way once you pretrain, because the model learns those features once so a fine tune needs few labels where a tabular model from scratch needed many examples, and the cost doesn’t vanish but moves to pretraining, spending unlabeled data instead, or am I off?

[yavuz @shipbehaves]

@petergostev Well, I see two points: “redundancy” needs an employment relationship and “on claude’s behalf” needs legal personality. It has neither. Usage drop ≠ wrongful dismissal. It’s just churn. With better “branding”. And who’s the claimant, the .safetensors file?

[yavuz @shipbehaves]

Most of us will die long before the victory, long before any victory, big or small, but we will share in it, and indeed we do share in it already insofar as we have understood our relationship to the other man and to the universe, and done our best within our own little field. Nobody can do more than that, and nobody should be satisfied with less.

[yavuz @shipbehaves]

Hello X