Sarah Guo of Conviction: why LLM won't replace work that can't be measured

Sarah Guo, founder of the AI fund Conviction, has published an analysis that reframes the central question about AI and labor: not "will AI replace people", but "what exactly will AI replace — and why not everything."

Where LLMs hit the ceiling

Language models learn only where there is a clear learning signal. A correct answer to a math problem, code that passes tests, a diagnosis that can be verified — all of this can be labeled, evaluated, and therefore learned. But Guo points to a structural limitation: a huge swath of valuable work cannot be standardized or measured with sufficient precision.

Reinforcement learning by definition requires a reward function — an evaluation of how good the result is. If no such evaluation exists, there is no signal for improvement. This is not a temporary problem that the next version of GPT will solve. It is an architectural limitation inherent to the very nature of trainable models.

What work remains for people

Guo identifies several categories of work that cannot be "turned into a dataset":

• Decisions under uncertainty — when the right answer is unknown in advance, and the result is seen years later. A venture investment is the classic example.
• Building trust — between people, companies, investors. Relationships are not digitized and not evaluated in real time.
• Judgment about unique situations — without historical analogues, without precedent. Each major business pivot is unique.
• Interpreting implicit signals — tone in negotiations, cultural context, unspoken customer expectations.
• Creating something new — ideas, product categories, strategies that did not exist before. The dataset for this is empty by definition.

Common feature: no labels, no benchmark, no way to verify the result with sufficient precision for training.

Why the Application Layer gets venture funding

This explains the logic that guides funds like Conviction (~$300 million under management) and Mento VC: they consistently invest not in creators of fundamental models, but in products built on top of them. Fundamental models — GPT-5, Claude, Gemini — are gradually becoming commodities. They compete with each other, their capabilities level out, the cost of access to frontier-quality decreases every quarter.

Sustainable value is not created in the model itself, but in the context of its application — in what models by their nature lack. The Application Layer accumulates precisely this: domain expertise, accumulated behavioral data, workflows that cannot be copied without deep knowledge of the domain, customer trust, and — most importantly — human judgment where automation is structurally impossible.

"There is work that cannot be reduced to a standard or measured.

If you cannot measure it, you cannot train a model to do it better" — in this thesis of Guo's lies the key to understanding which startups get venture funding and which do not.

What this means

Guo's argument gives investors and founders a practical filter. Companies built on measurable tasks — document processing, content generation from templates, basic data analysis — face maximum pressure from the models themselves and struggle to explain sustainable competitive advantage. Those working in the zone of the immeasurable gain a structural advantage that will not disappear with the next version of GPT.

For investors, the question is simple: if the value of the product is entirely determined by the quality of the base model — weak investment. If the value lies in what the model cannot reproduce in principle — this is interesting.