Sabrina Halper Show · 3 min · Oct 15, 2026

Robots need human data pipelines to learn consequences

Founder & CEO, Micro1

Founded micro1, pivoting it from AI recruiting to expert training data, growing it past $1B ARR.

Sabrina HalperHost · Host, Sabrina Halper Show
Has worked as an investor at HOF Capital and founded SHS Media.

About this segment

Micro1 collects egocentric task footage, then annotates and structures it into training data. Ali says robot experience alone cannot provide human-like reward or a sense of consequences.

From “AI Will Always Need New Data to Stay Aligned With Humans | Micro1’s Ali Ansari”, Oct 15, 2026. Watch the full episode

“you have to take that data and you have to turn that into training signal”

Ali Ansariat 41:10 · Sabrina Halper Show41:10

Takeaways

Robotics data starts with real-world, egocentric task footage.
Annotation and structure turn raw footage into training signals.
Robots need external reward signals, not just their own experience.

Watch 2:44 from 40:52

In this episode

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

40:57
Does that mean there's like a camera attached to people maybe across the world just filming everyday tasks, something like that? Yeah. So there's two, for any like environment you build or any data set you build, there's usually seeding data, which is like data you collect from the real world. And digital world is a similar thing. And then you have to take that data and you have to turn that into training signal, which is like you have to structure it in some way you have to annotate it you have to build verifiers etc similar thing is in robotics so the first is seeding data which is the egocentric large scale you know cameras um uh cameras doing a bunch of different tasks um all around the world and then and then more importantly you have to take that and you have to annotate in very different ways
41:41
to to to turn it into actually like training ready data which is the latter is a little bit more of our focus but but we also have um a lot of hours in terms of like data collected in the other world and do you think like once we get robots out there they will be able to create their own data like it'll be an amazing feedback loop or not necessarily i think yes um but but but it doesn't remove this idea of you still need humans in the loop and you still need some level of structuring and annotation and the reason is if you think about what happened in the digital world right there's um like the companies are doing very well are the ones that have a lot of product traffic data which is the equivalent of saying like if the robot goes and acts in the real world yeah that's their product traffic data right yeah the lms like if it's you know some coding tool or
42:28
whatever it is if there are users that are using and they're giving feedback maybe it's thumbs up thumbs down or or like there's just like some innate signal that you can take from that that's very useful but but it's not sufficient um it is just part of what you can use for training pipelines but but but it's not the it's not the entire thing um and and and the reason is like a sort of a a bit of an obvious statement which is like models don't have models don't have a innate sense of consequence right like if if a robot goes and does something you know touches a stove and it's hot and like they're not going to understand that there's there's no like you know i got this emotion of of uh of like my arm hurting and i'm not going to do that again there's no such thing and that's like the reward signal that exists within humans that that models
43:15
just don't have and won't have so so at some point somebody something needs to give it some reward signal yeah which which can partly come from its own like traffic data but but it's not entirely going to come from there and this may be so specific but like our ordinary behavior is worth