What this covers

Patrick talks with Ash, an early-stage investor in machine learning companies, and most of the conversation is a working toolkit for evaluating them. The Five-Factor Dataset Assessment scores data on uniqueness, fungibility, dimensionality, breadth, and perishability. The Four-Factor Model Assessment asks whether accuracy can beat a human baseline, how much data that takes, whether the payoff is convex or concave, and whether the model stays stable as it learns. Ash adds a timing rule: wait until accuracy, data scale, and business value suggest a runaway lead, and weigh the cost of waiting against the risk of going in early. Lilt's translation tool, which recomputes after every user correction, is set against Google's static system. Salesforce illustrates the data-rights problem for incumbents.

The range is wide. Ash lays out a Four-Phase AI Adoption Model, running from consumer recommendations to AI deciding in grids and healthcare, and argues that the cost of error sets the pace. Cambridge Cancer Genomics serves as a medical example. He also covers model bias, citing the Microsoft Twitter bot, and makes the contested claim that simpler, decomposable models let you locate a bias in a way deep networks do not. Other contested forecasts are that regulation will trail the technology and that SaaS companies will be poorly valued in seven to ten years. He also argues that society must let AIs run loose in energy, weather, and biology.

The last stretch turns to venture itself. Ash objects to calling investments "bets" with binary outcomes, since that ignores total loss. He describes early investors as translators to later-stage funds, and calls very large funds chasing $10–20 billion companies unproven. He also doubts that selling a unique dataset in bulk to traders confers an edge.

Sharpest takeaway

Fontana argues that durable competitive advantage in AI-era businesses comes from virtuous data-model feedback loops built into the product, not from static data or models, and that this favors vertically focused, B2B companies that automate decisions in markets large incumbents ignore.

  • Static datasets and single models are perishable, fungible, and easily beaten, so advantage must come from a self-reinforcing loop
  • Competitive advantage is shifting from building software, now commoditized, to automating decisions, which is hard to do
  • Small, unglamorous markets that incumbents ignore give startups room to build defensible positions

The argument · threads1 threads · 80 claims
0.85

Stationarity of the modeled outcome determines whether machine learning works well.

1 pointcentrality 5/5
  • Prices and other outcomes that are non-stationary make machine learning difficult, whereas cash flow growth and sales growth are more stationary and therefore easier to model.

    “if the outcome that you're trying to match or optimize for is non-stationary, things get pretty tricky. That's certainly the case with price”