Real Alpha
Watch mode

Sabrina Halper Show · 4 min · Oct 15, 2026

Useful training data must match the task and its context

Ali Ansari

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

Ali says models must cover the distribution of work, including ordinary tasks, and be trained by people close to each workflow. A senior title alone may not identify the right expert.

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

“everyone agrees that this is the excellent way to do it. In almost every area, there's a contextual excellence that exists.”
Ali Ansariat 46:40 · Sabrina Halper Show46:40

Takeaways

  1. General automation needs ordinary as well as specialized tasks.
  2. Senior leaders may not do the hands-on work being modeled.
  3. Excellence is contextual to the customer and workflow.
Watch 4:26 from 43:27

In this episode

Full episode
  1. 22:36Human expertise must keep updating AI models5:45
  2. 35:18Robotics is held back by a much larger data problem4:40
  3. 49:24Training data is a strategic asset that should not be exported4:41
  4. 16:07Micro1 gave up a growing recruiting business to focus on data3:25
  5. 19:40Expert data fills gaps the internet was never built to cover3:05
  6. 29:26Expert debate can train models on some subjective work3:43
  7. 40:52Robots need human data pipelines to learn consequences2:44

  8. 43:27Useful training data must match the task and its context4:26

Up next · Core MemoryRealbotix prioritizes autonomy and useful tasks over walking6:23

Transcript

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

  1. 43:32

    just as much as edge cases or it's not separated um no it's it's actually um the the other way to think about this is like one of the one of the people one of the common things people ask is like we're we're always it seems like we're scaling on expertise right which is like at some point we're going to get to like only the best surgeons in the world training models right but but that's actually not true but what you said is is um is more true Is if we're trying to build models that can automate the entire economy, quote unquote, which doesn't exist, we spoke of that earlier, you need to get the you need to get the entire distribution of what exists in the economy and the entire distribution is is a lot of ordinary tasks.

  2. 44:19

    In fact, almost all tasks are more ordinary than specialized. So one way to think about this is when we get a model to do a really good job on legal memos, the partner at the law firm or many partners at law firms that are trying to get the quality of the legal memo to be good is an important part. But then it's also some somebody that's good at writing that maybe is not even like legal or maybe the paralegal that's like extremely good at writing.

  3. 44:48

    They're going to focus on the writing aspect and they're going to get the writing part to be good. And they're not necessarily like a partner at a law firm. They're just like a paralegal. So so so you need to match you need to match the distribution of all actions that happen in the world if you want to have an impact on all actions. And that means lots of ordinary actions. Yeah. OK, one one question on this is. There's definitely this concept of like, OK, if all the ordinary stuff is automated, people that are exceptional at something will always be valuable and they'll be the most valuable even to train data, like even in the world.

  4. 45:23

    i know you don't believe in this world where everyone's full-time job is just training ai and like creating data um people who are exceptional are the most uh are worth the most and you're saying that's like kind of actually not true but i guess the question i would have is like if you could get mna like workflow and like decision data from like the ceo of goldman Sachs is that actually worth like is there anyway is that worth so much so the the CEO of Goldman Sachs would would probably not have a clue on how to redline M&A documents right and there's actually this is a good example because if you if you overdo it you actually get people that are not in the weeds of what matters and what you're trying to train. So even if we take a step back on like, let's say it's not the CEO of Goldman Sachs, but it's general counsel at Goldman Sachs.

  5. 46:25

    And we get the general counsel to give us redlining data. That person is probably also not the best redlining individual at the company, right? So you have to like pick the right person and the right data pipeline for the job. There's no notion of this is objectively excellent globally and everyone agrees that this is the excellent way to do it. In almost every area, there's a contextual excellence that exists.

  6. 46:54

    And that's what you've got to get the models and robots and so forth to do a good job with. And the last thing I'll say on this is I'll use an Alex Karp quote. Love it. He says, I think it was something along the lines of, like i don't give a shit about your sweet bench pro which by the way sweet bench is great like you know but but he said i don't give a i don't give a shit about your sweet bench pro i care about what my customers care about uh what what they what what is good intelligence in the models for then is what we're going to build towards which goes again it's the same idea of like if you hill climb an arbitrary benchmark it doesn't mean that it's good for that enterprise it's it's actually probably irrelevant for that enterprise so so that's why there is no um there's no global

  7. 47:38

    kind of uh objective excellence and and it's it's just very context yeah also i guess excellence is perpetually reborn like exactly a lot of the biggest companies today are run by people who