As AI commoditizes benchmarkable work, an organization's lasting moats lie in tasks that are verifiable through its private data and judgment
The mid-2026 investor's version of AI psychosis is a despair that nothing is investable, that we should put all our money into Anthropic and Nvidia and go home.
Sarah Guo
Context & Ripple Effects
The related coverage captures a widening split in AI discourse: investors have worried that AI could impair software-company valuations, while other commentary has framed AI as augmenting rather than simply replacing human work. This piece narrows the investability question to where differentiation can persist once benchmarked capabilities become broadly available.
Its core distinction is between work that can be evaluated against common benchmarks and work whose quality depends on an organization’s private data and judgment. That shifts attention away from raw model capability alone and toward the conditions under which customers can verify and trust an AI-assisted outcome.
First-order effects
Companies whose products compete primarily on benchmarkable AI capabilities face weaker claims to durable differentiation as comparable capabilities spread.
Organizations with proprietary data and established decision processes gain more leverage in AI use cases where outputs must be verified against their own records, standards, or judgment.
Second-order effects
Software vendors and AI startups are pushed to package models with proprietary workflows, data access, and evaluation layers rather than rely on model performance as the main product distinction.
Investor attention may concentrate further on infrastructure leaders such as Anthropic and Nvidia, while application companies face greater pressure to demonstrate why customers cannot readily substitute an AI-enabled alternative.
Third-order effects
If this pattern holds, the AI application market could organize less around generic model access and more around control of domain-specific data, verification, and accountable human decision-making.
The persistent debate over AI-driven software disruption may increasingly turn on which businesses can make AI outputs auditable and trusted, rather than on whether AI can perform a task at all.
The trend: AI is shifting competitive advantage from broadly benchmarked capability toward proprietary context, verification, and judgment embedded in real workflows.
the venture whitepill is that the world contains an uncountable number of opportunities that will remain inaccessible to the labs. however, recognizing this requires leaving (both physically and mentally) the city of San Francisco, and is thus out of reach for most investors
TL:dr on this excellent piece by @saranormous 1. Passing a test in school is not real life experience 2. Do hard and messy things 3. In the 90, investors were paralyzed because Microsoft would do everything. Now it is the foundation models. 4. Coding, which is the easiest
Excellent essay. “The valuable work is illegible by construction: anything you can put on a leaderboard, you can train against, so anything measurable is already on its way to commodity.” Insights in every paragraph.
“So, we may ask two things of any kind of work. Is its correctness private and expensive to establish, the kind of truth that exists only inside someone's data? And is it walled off, locked inside a system you can't get into? Set those against how saturated the task is, and you
This is a critical post to read if you're building an applied AI company right now. “An application earns its place in the untrainable corner by doing unglamorous work: arranging a company's private reality so a model can act on it, handing the model the tools to act, working
Good bit. This is very obviously not true to me. The level of groupthink psychosis amongst investors seems to be at an all time high. Why is that? [image]
Techmeme did what I could not with this essay, which was to cut through all the obviously Claude-generated portions to try to divine the author's original prompt [embedded post]