Intelligence should be plural.
Thousands of models
No single intelligence covers the whole map.
01
There is no best model.
AI is increasingly measured as if intelligence exists on a single axis.
We don't think it does.
The right intelligence depends on the task, the environment, the objective, and the person using it. A model can be more capable without being more useful. Two models with similar capabilities can communicate differently, behave differently, embody different preferences, and produce dramatically different experiences.
As the number of models grows, understanding those differences becomes as important as measuring raw capability. We're building the tools to measure them.
02
What we study
Evaluation
We build benchmarks for dimensions of intelligence that traditional evaluations miss—from long-horizon performance to the behavioral characteristics that shape how humans experience AI.
Environments
We build open environments where models can act, receive feedback, and be evaluated through experience rather than isolated questions.
Open Infrastructure
We build tools for working across models—routing requests, capturing traces, evaluating behavior, generating training data, and creating environments.
Fieldwork
Some questions shouldn't begin with a benchmark. We experiment with AI in the world to understand what models can do, how they differ, and how humans respond to them.
03
Our thesis
We are betting on a future with thousands of models. Different architectures. Different capabilities. Different personalities. Different values. Different costs. Different ways of communicating and reasoning.
We want the infrastructure around AI to make those differences visible rather than eliminate them.
People should be able to understand the intelligence they are using. People should be able to choose it. And people should be able to shape it.
04
Open by default.
Benchmarks are more useful when others can challenge them. Environments are more useful when others can build on them. Research is more useful when others can reproduce it.
Whenever possible, the work of Last Labs will be published and open sourced.
05
Build intelligence that amplifies people.
AI should expand human agency—not replace it with a single definition of intelligence.
That's the future we're working toward.