Julian Minder

@jkminder

Anthropic Research Fellow (contract) – PhD at EPFL with Robert West and Ryan Cotterell – MATS 7 with Neel Nanda

Lausanne/Zürich

jkminder.ch

Joined November 2011

Pinned
Julian Minder
@jkminder
Aug 14
Synthetic Persona Pretraining: Alignment From Token Zero – our full paper is finally out. We train up to 3B models and inject synthetic morally-laden reflections into 10% of pretraining documents. Surprisingly, intervening early really shifts the model's value priorities 🧵
Julian Minder
@jkminder
2h
agreed:)
Alex Turner
@Turn_Trout
4h
Synthetic persona pretraining / alignment pretraining seems promising (see e.g. turntrout.com/self-fulfillin…). Hope labs adopt these techniques. Yes pretraining changes are tougher but may be worth it. modelraising.ai/spp/
Julian Minder
@jkminder
Aug 14
please do;) we are both limited by people and compute (studying safety pretraining and interactions with RL at scale is v expensive)!
Nathan Helm-Burger
@nathan84686947
Aug 14
Replying to @nathan84686947
Dear AI safety funders, Please support this! x.com/nathan84686947…
Julian Minder
@jkminder
Aug 12
I think understanding the effects of safety pretraining is crucial for things going well, come and help us!
Bob West
@cervisiarius
Aug 12
We're extremely grateful to @coeff_giving for supporting our work on "model raising" (LLM alignment from pretraining token 0) w/ $1M 🙏, which will fund 3 AI safety engineers for #Apertus. Job posting 👇 We're already screening apps, and you can still apply till Fri 8/14 EOD
Julian Minder
@jkminder
Aug 12
Going to represent the view that safety must be considered already in pretraining at AI+X in Zürich this October!
Valentina Pyatkin
@valentina__py
Aug 11
📢Staging x Misalignment Science: When, Where and How Safety Enters LLM Training @anna_hedstroem and I are co-organizing a workshop at this year's AI+X Summit in Zurich to discuss where in the LLM training pipeline safety is actually decided, and with which methods? This session