Episode summary
Why AI Agents Will Quietly Replace 80% of Investment Teams
How I Invest With David Weisburd
· · 21 min · Hosts: David Weisburd · Guests: Ali Ansari
Summary
Ali Ansari argues that micro1 sits in the “human intelligence” layer of AI: it vets PhDs, professors, and industry experts to generate the new data frontier model labs need for post-training, and it also helps enterprises evaluate agents before production. His core claim is that model improvement is still overwhelmingly driven by humans, from internet-scale pretraining to preference labeling to today’s expert-generated data in fields like medicine, finance, law, and investment banking.
He says the main reason AI agents are not broadly useful in production is not model quality but evaluation. Enterprises need rigorous frameworks that test an agent’s action space like a real workflow, because impressive demos can hide failure rates that would be unacceptable in live systems. He thinks coding is the clearest killer app so far, and expects similar gains to spread into other domains once evaluation becomes standard.
Ansari also describes micro1’s own pivot: the company stopped serving some startup recruiting customers, focused on AI labs and foundational-model work, and says that shift helped it grow more than 30x in 2025. He argues that market selection, ruthless focus, and close customer contact matter as much as product quality for a fast-scaling AI business.
Key takeaways
- Ali Ansari says micro1 helps frontier labs gather expert human data for post-training and helps enterprises evaluate agents before production.
- He argues AI progress is still “almost entirely humans” teaching models, from pretraining to preference labeling to expert data.
- Ansari says production agents need evaluation frameworks for each action, because a demo can work one out of five times and still fail in deployment.
- He says coding is the only AI agent use case that is clearly working well in production today.
- Ansari says micro1’s decision to focus on AI labs and human-data infrastructure helped the company grow more than 30x in 2025.
- He says CEOs should constantly question meetings and stay close to customers and experts to get to ground truth.
Notable quotes
“Micro One is the AI platform for human intelligence”
“It's almost entirely humans that teach the models in some way or another”
“there needs to be an evaluation framework for each of the actions that the probabilistic software needs to do”
“the only use case that is very useful in production now”
How the episode opens
At a high level, how do you explain Micro One?
Micro One is the AI platform for human intelligence. So, what that means is we vet highly skilled people mainly PhDs, professors, uh, and industry experts mainly in medical, finance, and legal, but also many other domains and we help train frontier large language models. So you could think of uh you know the AI labs the the way they're kind of improving their model capabilities is by uh gathering net new human data uh for their post- training pipelines and we help them gather that net new human data.
And who are your customers? customers are the frontier labs um that build foundational models and we also have um uh enterprise customers you know max 7 and and kind of fortune 500 broadly that are building also foundational models but but also they're building uh enterprise agents that we help them evaluate and kind of get ready for production. One of the reasons I wanted to chat today is because upstream of the LM's improving, there's these improvements to the models.
Maybe you could unpack on why are LM models improving and how much of that is this recursive AI improving itself and how much of it is the PhDs and and these other professionals.
It's almost entirely humans that teach the models in some way or another. Um, of course that started with uh the pre-training phase where humans taught models by first creating the internet. Of course that was the you know the largest set of uh uh human data that we that we had initially which the models kind of took a unsupervised route of training and uh you know that was kind of the initial initial state to the foundational models and then afterwards where the models really got useful is when um humans kind of uh started to do a bunch of preference labeling kind of choosing which answer is better and so forth uh based on the model responses and then uh once we pass that phase now we're in this kind expert data training where humans are creating really complex data from scratch whether it's doctors, lawyers, finance experts, and investment banking and and other areas.
From the episode’s captions: roughly the first two minutes, lightly cleaned. Keep watching on YouTube →
Topics
- ai agents
- expert-generated data
- evaluation frameworks
- ai post-training
- market timing in investing
Companies and tickers mentioned
- micro1
- AlphaSense
Watch the episode
Watch “Why AI Agents Will Quietly Replace 80% of Investment Teams” on YouTube
Summary, takeaways and quote selection by Octagon. Quotes are short excerpts; the full conversation is in the episode.
More from How I Invest With David Weisburd
-
Hunter Walk: Why Venture is Getting Harder, Picking Great Founders, and Investing Against Consensus -
True Anomaly’s Even Rogers on Space Warfare, China’s Military Advantage, and the Defense Tech Boom -
Finding the Next Anthropic, Quantifying Venture Capital, and the Science of Picking Winners -
Why 95% of AI Projects Fail, Anthropic still at $2T?, and Why AI Is Still in the First Inning -
Peter Thiel's 27-Year Partner on the AI Bubble, the Venture Capital Crisis, and the PayPal Mafia -
Why Venture Is Eating Public Markets, Picking Great VCs, and the AI Boom w/Brijesh Jeevarathnam
How I Invest With David Weisburd, this episode and its artwork belong to David Weisburd. Octagon is not affiliated with the show; this page is an independent summary.
