alth0u, general partner at a16z, disagrees that AI models are earning greater trust for practical coding tasks and says more review of chain-of-thought outputs may be needed
Joanne Jang replied agreeing on the need for greater human agency and model transparency.
hmm i'm not sure where we disagree?
i agree the computer should be invisible too -- i love physical buttons and also hate the tesla screen thing (esp touch screen control for a GLOVEBOX!!)
when i say the model should be invisible in the car case, i mean there should be enough affordances for the rider to have peace of mind _but_ there's no reason to intentionally remind the user that the model is doing the driving
alth0u🧶@alth0umm thats precisely where i disagree: - the models are still and maybe even getting worse on (or just being used in more complicated situations) s.t. actually useful things even in coding that I actually have to read more CoTs - this is the critical issue with the tesla that "everything is computer" - this comment however is about the non-driving parts of the car UX feeling bad (why do i press the screen representation of a door to unlock the door, why not just unlock the door), but I agree that cars in general should melt re: driving ONLY but FULLY but driving is very "low-dimensional" in the sense that other tools/intelligences are used way more for "off-roading" by the average user than cars are
@alth0u super aligned on maximizing human agency and trust -- probs as a function of transparency, reliability, actual alignment
alth0u🧶@alth0uyea i think we are on the same page for software in general but i do think the user and the model have to stand side by side as mutual intelligences and the model has to be very apparent personalization to me is merely calibration of how to escalate takeover decisions to me the model and user are a dyad that tradeoff who is driving
#1114alth0u🧶@ALTH0Uyea i think we are on the same page for software in general
but i do think the user and the model have to stand side by side as mutual intelligences and the model has to be very apparent
personalization to me is merely calibration of how to escalate takeover decisions
to me the model and user are a dyad that tradeoff who is driving
Joanne Jang@joannejanghmm i'm not sure where we disagree? i agree the computer should be invisible too -- i love physical buttons and also hate the tesla screen thing (esp touch screen control for a GLOVEBOX!!) when i say the model should be invisible in the car case, i mean there should be enough affordances for the rider to have peace of mind _but_ there's no reason to intentionally remind the user that the model is doing the driving
Our broader goal with Safety Nudges is restoring user agency. Making hidden interaction patterns visible.
Helping users recognize when a system may be: • sounding more certain than it should • encouraging emotional dependence • presenting probabilistic outputs as truth
Chhavi Yadav@chhaviyadav_As AI becomes more integrated into daily life, safety can’t just live in model training pipelines. It also has to exist at the UX layer. Sometimes the most important intervention is simply helping users pause, reflect, and think critically before accepting an answer.
As AI becomes more integrated into daily life, safety can’t just live in model training pipelines.
It also has to exist at the UX layer.
Sometimes the most important intervention is simply helping users pause, reflect, and think critically before accepting an answer.
Chhavi Yadav@chhaviyadav_One thing we realized while building this: AI-Risk literacy can’t just be taught through courses or policy papers. It has to happen at the moment of interaction. When users are emotionally engaged. When trust is forming. When decisions are actually being made.