Argentina · MMXXVI

Once you see it, you can't stop thinking about it.

Zahir is an independent AI research lab in Argentina. We build the architectures, verifiers, and world models that come after LLMs.

— after Borges, El Zahir (1949): an object that, once seen, cannot be unseen.
The Thesis

The transformer era is peaking, not ending.

Large language models are sequential and brittle. They hallucinate, they depend on the cloud, and they continue text rather than reason about it. They cannot plan, and they do not ground themselves in causality. This is not a flaw we can patch. It is the shape of the paradigm.

The incumbents, OpenAI and Anthropic and Google, are too committed to the bet they have already placed to go looking for the next one. That is the opening. The 2025–2027 window is when an independent lab can get there first, before the largest players notice they need what comes after.

“We are not racing the frontier labs. We are building what they cannot afford to chase, because they are too committed to the bet they already have.”
Zahir — founding memorandum
The Work

Five bets on what comes after.

i

Hybrid SSM + attention architectures

State-space cores with selective attention, designed to run on-device and cheaply. Not another model that only lives in the cloud.

Architecture
ii

Neurosymbolic verifier infrastructure

A neural model proposes, then a formal verifier (Lean, Z3) checks the result. The output is proven, not just plausible.

Verification
iii

World models for science & industry

Learned simulators for agriculture, mining, and energy, where being grounded in the physical world matters more than writing fluent prose.

Simulation
iv

RL on verifiable rewards

Reinforcement learning where the reward can actually be checked. We are applying it to agriculture, fraud detection, and clinical triage in Latin America.

Reinforcement
v

Sovereign AI appliances

Small, distilled vertical models running on-premise. Argentine data, processed on Argentine soil.

Deployment
“The Library of Babel has every book, and no way to know which one is true. We are building the reader that verifies.”
The Place

Why Argentina.

The talent

UBA, ITBA, and CONICET produce world-class researchers at a fraction of what they cost in Silicon Valley. It is a deep bench of ability that has barely been drawn on.

The market

Latin America is a large, underserved market with real problems in agriculture, fraud, and clinical care that Silicon Valley has never had a reason to solve.

The sovereignty

Argentine data should not have to travel to American or Chinese servers to be processed. Here, on-premise is not a constraint we are working around. It is the entire point.

The moment

The window is narrow, and it is open now. A lab rooted here can move before the incumbents do, and build the sovereign infrastructure the continent is going to need.

Two readers. One Zahir.

Address us as an investor, or as an institution.

For Investors

An asymmetric bet beyond the LLM

A credible team going after a defensible IP position, before the frontier labs understand they need what comes next.

Request the investor memorandum →
For Government & Institutions

Sovereignty, jobs, applied return

Partnerships with INTA, AFIP, and the Ministry of Health. High-skill jobs and sovereign capability, built at home.

Open an institutional conversation →