We build productive self-improving scientific intelligence

to create beneficial technologies at the edge of human knowledge, starting with biological sciences

Transcription factor assembly
Approach

Science isGenerative Selection

The ability of science to compress knowledge into compact causal theories weighs against the quadratic scaling of attention to overcome current context limitations. It opens a way to train scientific reasoning models that continuously learn from real-world experiments, making them excitingly useful for discovery at the edge of human knowledge. These discoveries will lead to novel compact theories allowing for out-of-distribution prediction and hence true generalization.

Three panels: the Bitter Lesson, the scientific loop, and science as compression.
01

Generate

AI generates domain-specific world models and proposes tests of orthogonal world hypotheses to collapse the search space

02

Test

We run the highest-information-gain experiments in the lab

03

Select

Experiment determines in which of the possible worlds we live in and proposes next steps to rapidly advance the correct world model with closing the loop in days — not years

First application: design of biologics optimized for a desired function. Our long-term goal is to accelerate progress of applied science solving the hardest problems of humanity, including age-related functional decline.

Team

From the team behind

GPT-4b-micro

Scientific reasoning model for

lab-in-the-loop protein design

50× improvement of cellular reprogramming by AI protein engineering

within Retro × OpenAI collaboration

$1B+ in closed transactions

for AI research-generated intellectual property

Founders

Let's build it

Email us at alexey@genselect.ai or andrei@genselect.ai.