Builder-driven lab
powered by Causal Computation
We build the infrastructure that turns streaming numerical facts into verifiable causal structure, simulation, and decision intelligence.
Why now
Four eras of organizing
intelligence.
From data retrieval to causal computation. The evolution of how humanity processes and acts on information.
Bloomberg
Organized financial data into real-time access.
Created the first universal terminal for financial information. Turned scattered pricing data into a searchable, real-time system that became the backbone of global finance.
$120B market cap
Organized the web's information graph for retrieval.
PageRank treated the web as a directed graph — links as votes, structure as signal. It didn't just search text; it read the topology of human knowledge.
$2T market cap
GPT / Claude
Organized language into high-fluency generation.
Transformers compressed the statistical structure of language into models that generate fluent text. But fluency is not understanding — and prediction is not causation.
No verifiable outputs
Abel
Organizes causation with regime-aware, testable outputs.
A live causal world model across 200K+ variables. Structure refreshed daily, predictions updated hourly. The only layer that answers "what happens if I act?" with math you can check.
The decision layer
Mission
LLMs read the world.
Abel computes it.
Abel is building the Social Physical Engine — a live, regime-aware causal world model inferred from numerical facts.
Decode Reality
The world is a causal system. We built its engine.
Abel maps 200K+ variables into a live, directed causal graph — refreshed daily, with structure that reveals how change actually propagates.
Compute Consequences
Moving AI from "what is" to "what if."
Instead of predicting outcomes, Abel computes the causal effect of actions. Intervention, counterfactual, and simulation — all Layer 2/3 math.
Restore Truth
Decisions demand proof, not persuasion.
Every output carries p-values, confidence intervals, causal chains, and validity boundaries. Fluency without grounding is not intelligence.
Epistemic Principles
How we compute reality
These principles govern Abel's architecture — rejecting static assumptions and linguistic fluency in favor of structural truth.
The world is non-stationary
Reality changes; models must update with it.
Systems are reflexive
Predictions, incentives, and interventions change the system itself.
Structure matters
Numerical facts are not enough; what matters is the causal structure beneath them.
Claims must be verifiable
Intelligence should be testable, falsifiable, and accountable.
Models must be regime-aware
Good systems do not average across worlds; they detect when the world has changed.
Outputs must preserve integrity
Fluency cannot come at the expense of uncertainty, assumptions, or causal validity.
For the implementation-level view of these principles, see the Concepts page in our documentation.
Unfair advantage
The people who invented the field
are building the company.
The moat isn't code. It's that the field's inventors are our co-founders, and we already have 18 months of causal graph history no one else has.
Stephen Wang, PhD
Industry FounderCo-Founder
Principal Research / Director @ Snap · CTO (US) @ Amber ($AMBR) · Tech Lead @ Meta · Microsoft. 2× Founder (1 exit). System architect · GPU-native causal infra.
Biwei Huang, PhD
Academic PioneerCo-Founder
Professor, UC San Diego · Apple Scholar (1 of 10 globally) · Program Chair, CLeaR 2025. Creator of causal-learn & Causal-Copilot — the open-source libraries Abel runs on.
Advisors & Investors
Kun Zhang, PhD
CMU · H-index 73 · 28K citations · Father of modern causal discovery
David Danks, PhD
UCSD · National AI Advisory Committee · AI Ethics & Causality
Michael Li
VP Engineering, Coinbase ($COIN) · VP Data, LinkedIn
Michael Wu
Founder & CEO, Amber Group · AI-Fintech pioneer
Team DNA
Causal discovery is a 20-year academic discipline. The person who wrote the software Abel runs on is our co-founder. The person who created the theory is our lead advisor.
Technical moat
100K-variable GPU-PCMCI engine
Academic SOTA: ~500. Abel: 200× beyond. GPU-native tensor-parallel.
3M-node global causal graph
Stocks + crypto + macro. 30 time lags. Daily refresh. Running now.
Markov Blanket query engine
Complete causal neighborhood with lags, strengths, p-values.
do-calculus intervention simulator
"If X changes 1σ → causal effect on Y" — with CI and audit trail.
18–24 months minimum to replicate
GPU-native causal discovery isn't off-the-shelf. We wrote the math and the CUDA kernels.
Founding scholars of the field
Biwei Huang (causal-learn), advised by Kun Zhang (H-index 73). This team doesn't exist twice.
Join Us
Build the decision layer of AI
We are hiring researchers, engineers, and builders who want to work at the intersection of causal inference and production systems.
Read our blog to understand how we think.