AI Will Always Need New Data to Stay Aligned With Humans | Micro1’s Ali Ansari

Summary · 4 min read

Ali Ansari is Micro1’s solo founder and CEO. The YouTube description says Micro1 supplies human expertise for AI training and passed $1B in ARR in 2.5 years. Ansari moved from Iran to the U.S. at age nine and started out reselling textbooks before building companies.

Ansari argues that models need continuing human data because knowledge changes and AI creates new work to automate. He says Micro1 shifted from AI recruiting to training data as demand for expert contributors grew, and that models need structured expert judgment beyond internet text. He sees robotics as a larger data challenge than software and says robot-generated feedback will not remove the need for human annotation and reward signals. He also argues that exporting structured training data can help rival countries catch up, while the data-center buildout depends on unlocking many more large-scale AI use cases.

Minute by minute

1. 2:49

   Ansari credits his parents' sacrifice. Ansari says he was born in Tehran and moved to the U.S. at age nine, learning English in about four months. His parents left a settled life in Iran, including his father’s company with 40 to 50 employees, and rebuilt their lives in the U.S. He says that experience later made him feel a duty to care for them.

2. 7:32

   Textbooks became his first business. Ansari says he began selling items on eBay at about 12, starting with his father’s shoes, which brought in $60 to $70. He later bought goods at garage sales and focused on textbooks, then built Cash Books Now to buy books back from college students. He used Amazon data to set prices, paying 30% to 50% of a book’s value.

3. 12:13

   Micro1 pivoted from recruiting to data. Ansari says he built a software agency at Berkeley, then created an AI tool to screen engineers because vetting them consumed his time. That tool became Micro1’s AI-recruiting product. After a data company requested hundreds of engineers within weeks in late 2023 or early 2024, Micro1 shifted fully to expert training data, giving up a recruiting business with millions of dollars in run rate.

4. 19:40

   Internet text is only a baseline. Ansari says internet data includes substantial low-quality material and incorrect answers, making it a baseline rather than an adequate training set. He points to top lawyers creating verifiers for legal memos as an example of expert data the public web has little incentive to produce. He says tens of thousands of people at Micro1 and hundreds of thousands across the field are now incentivized to create training data.

5. 22:36

   Knowledge keeps changing. Ansari argues that training models does not simply capture a fixed body of knowledge: he cites an estimate that 30% to 40% of medical-school knowledge is no longer relevant or needs updating by graduation. He says AI also frees people to take on new work, expanding the set of tasks that models must learn. As an example, he says programmers now spend close to 0% of their time programming, while building more software and enabling more people to participate.

6. 29:26

   Subjective judgment needs expert debate. Ansari describes creativity as a spectrum from objective tasks to highly subjective fields such as art and music. He says expert rubrics, peer reviews and tie-breakers can help models converge on judgments in areas such as law, but debate may produce more disagreement at the highly subjective end. For advice systems, he says the relevant data should come from appropriate experts, such as mental-health professionals, philosophers or coaches.

7. 35:18

   Robotics is a larger data problem. Ansari says robotics is mainly a data problem because the field lacks an internet-scale pretraining dataset. He describes collecting egocentric footage of household and commercial tasks, then structuring it for training. He says the physical world has a much larger action space than the digital world, and contrasts general-purpose robots with self-driving systems, which he describes as a more constrained task supported by long-term data collection.

8. 40:52

   Robotics needs data across contexts. Ansari says Micro1 began building robotics data infrastructure about seven or eight months earlier, collecting real-world footage and annotating it into training-ready data. He says robots’ own usage data can help but cannot replace human feedback, because models lack an innate sense of consequences. Training also needs ordinary tasks as well as specialized expertise, and he argues that excellence depends on the context and task rather than a universal standard.

9. 47:54

   Ansari opposes exporting training data. Ansari says intelligent systems improve through data, compute and algorithms, and argues that data supplies models with much of their intelligence. He says the U.S. should not export structured training environments containing American expertise and company operational data to adversaries, because those datasets could help rival labs catch up. He argues that data from strong institutions and large-scale data operations is especially valuable.

10. 53:47

    Rapid growth strains founder priorities. The YouTube description says Micro1 passed $1B in ARR in 2.5 years. Ansari says he feels both grateful and stressed, and that his role changes every couple of weeks as he decides what to prioritize. He says choosing the wrong focus can waste time, while increasing context switching leaves him frequently unsure what is happening.

11. 56:53

    AI infrastructure needs more use cases. Ansari says data-center investments depend on expectations, a decade out, of much higher inference and model usage. He warns that relatively few large-scale production use cases have emerged so far, and says the buildout could be a disaster if adoption does not expand. He argues that data investment must grow alongside infrastructure because data improves model capabilities.

Published Oct 15, 2026.