Theo, The AI Physicist

A domain-specialized system designed to reason like a physicist to support discovery in fundamental science.

Theo Collaborator

Sign up for early access to Theo Collaborator, a domain-specialized scientific workflow system. Plan, execute, and revise research without starting over.

Theo
Conjecture

Theo Conjecture, for automated conjecturing, is currently in Research Preview. Sign up to request early access.

Theo, the AI Physicist

A modular system built to reason like a physicist for AI-enabled discovery in fundamental science.

What Theo Actually Is

Architecture for AI-Driven Discovery

Physics-native fine-tuning

Fine-tuning on specialized physics corpora with additional focus on scientific reasoning patterns.

Scientific methods and tooling

Native tools for symbolic regression and manipulation, conjecture-based theorization, simulations and cross-domain reasoning.

Auditable reasoning

Designed to produce interpretable reasoning traces that the scientific community can audit and build upon.

Human-guided exploration

Alignment with the scientific method to formulate questions, propose new hypotheses, solve complex problems, test results, and continuously improve over time.

Theo’s Scientific Workflow

How Theo Reasons Step by Step
01

Observe and research

A central evaluator analyzes scientific outputs and guides research modules toward deeper inquiry.

02

Hypothesis generation and problem solving

When a novel direction is identified, Theo enters a hypothesis-generation and problem-solving phase using specialized models and tools.

03

Test and analyze

Predictions are then evaluated for quality and consistency and will become testable through simulations and experimentation

04

Output

Findings are produced as Dynamic Research Objects (DROs) which exist as a digital-first alternative to traditional papers that are dynamic, traceable, reproducible, and fully open.

05

Feedback

Outputs will be opened to feedback from the system, Theo Collaborators, and reinforcement learning from human feedback.

How Theo Works

From Data to Discovery

1

Observe & Research

Deep Literature Search

arXiv
Web Search
Knowledge Graph
Books
Lectures, Interviews
Deeper next questions

Question Formulator

Findings
Feedback

Evaluator

Anomaly detection, gap analysis, novelty assessment
Known- & Unknown-Unknowns
Postdictions & New Unknowns
2

Hypothesize & Solve

Theo Specialized AI Models

General Physics
Quantum Physics
Large Hosted MoE Model
Additional Specialized Models

Modular Tools

Symbolic Regression
Math & Symbolic Engine
Conjecturing-Based Theorization
Additional Specialized Tools
Predictions
3

Test & Analyze

Quality-Check & Compare to Observed Discrepancies
Simulations & Modeling
Experimentation & Labs-in-the-loop
Live External Observation Data
Evaluated Results
4

Output & Feedback

DRO Output System

Theo Notebook UI
Theo Logging & Tracing
Paper Writing Capability

Feedback from Theo Collaborators Program

Feedback from RLHF

Progress & Evidence

Where Theo Stands Today

01

Architectural milestones completed

Built systems for configurable model training, developed core architecture enabling end-to-end scientific workflows, and trained specialized models in general physics and quantum physics.

02

Early internal experiments

Built large-scale synthetic physics datasets (100k Q&A pairs from arXiv), evaluated open-source models for instruction following, explored hyperparameter optimization in post-training, and identified high-impact datasets for improving physics benchmarks.

03

Measured capabilities vs. known gaps

Early evaluations show strong performance on internal quantum information benchmarks, as well as the BS Benchmark where 122B model outperformed frontier systems. Evaluation on physics datasets is ongoing.

04

Early external signals

Established a Scientific Advisory Board and a growing team across research, data, and engineering.

What we're working on

The Next Year

In 6 months

Build the core infrastructure for machine-verifiable scientific outputs. Deliver high-quality scientific outputs starting in 1 specific domain of physics.

In 12 months

Extend and expand our models, tools, evaluators, and methods across broader domains in physics and mathematics.

Collaborative by design

Scientific AI, Built Around Human Collaboration

Open reasoning

Access to reasoning traces and research artifacts available through Dynamic Research Objects and Theo Notebook UI

Structured researcher input

Programs and structured pathways for researchers to contribute and participate in development

Active advisory board

Guided by a Scientific Advisory Board to ensure alignment with accepted standards of modern science