Building the economy AI runs on.

Loka Labs trains models of how people behave, and builds the products the AI economy lends, pays and settles with.

MIT Harvard Princeton Columbia UC Berkeley Yale NUS ETH Zurich Stanford

Behavior model · In training

Not an assistant.A person.

A model that bargains, loses patience and misjudges, the way people do.

“We can renew your GPU lease at the current rate. Shall I send the contract?”

Assistants

The same answer, every time.

People

Many answers. Larger ones are more common. Examples are illustrative.

How it is built.One model, three steps.

Start from a base model1 / 3

Training a model to be helpful is what makes it stop acting like a person, so we start before that step, from the full range of how people talk.

Questionsit can answer.

The kinds of questions we are building it for, once training is complete.

  • How do impatient users react when an agent asks for one more step?Where do confused users give up, and what do they say when they do?Which wording makes a delay feel fair rather than evasive?
  • How hard will a tenant push back on a renewal price?Which offer gets a counterparty to sign for longer?When does holding firm end the conversation?
  • Does the same message land differently by age, culture or temperament?Which customers leave after a price change, and which stay?
  • Would a person find this reply pushy, cold or helpful?Which of two agents feels more trustworthy, and why?

How we test it.Results follow training.

01Unseen people

Test personas never appear in training.

02Four levels of fidelity

One person, a population, a whole conversation, and the response to a change.

03No leakage

Hard bargaining stays out of support conversations.

Loka World.A civilization, simulated.

Many people and agents in one world, living, trading and deciding together. Loka World is a simulated civilization for watching how a whole society behaves.

Coming soon

AIUSD

The dollar Loka Capital settles in.

Loans, escrow and repayments need a unit that holds its value and settles on-chain. AIUSD is redeemable 1:1 for USDC and is the dollar lenders use on Loka Capital. It does not pay interest; returns come from lending.

Redeems for
1 USDC
Get it with
USD wire or USDC
Interest
None

Payments

Agents pay for data and work they cannot inspect first. Money should move only when the work is done.

Atara

Payments that release when the agreed condition is met. If it is not, the funds return to the payer.

loka.cash
Atara website
How a payment moves
  1. 01Funds wait in a contract
  2. 02The condition is confirmed
  3. 03Released to the payee, or returned
Built for Apps and agents Conditional settlement through an API

Capital

Compute is built before it earns. Lenders need security they can see, and a claim on the rent.

Loka Capital

Senior secured loans to GPU compute centers. Each loan has its own escrow, and rent from AI companies repays lenders first.

Principal outstandingAs of Oct 16, 2026 · sample
$21.26M
SecurityFirst lien on GPUs
Weighted interest rate 11.4% Fixed, paid monthly · sample
Debt service coverage 1.28× 26 of 26 monthly payments on time · sample

About Loka Labs

Loka Labs designs, builds and operates each product, and every product has its own site, terms and users. A Turing Award laureate and researchers from the world’s leading labs stand behind our research, with published, peer-reviewed work.

Harvard MIT Stanford Princeton UC Berkeley NUS
12+papers at top-tier venues
Formal verification · Systems · Economics · Data
A.M. Turing Award A.M. Turing Award
Joseph Sifakis

Co-invented model checking, the formal-verification method behind mission-critical systems worldwide.

Joseph Sifakis A.M. Turing Award · 2007

In-network consensus and Byzantine fault tolerance: OSDI, SOSP, SIGCOMM.

Jialin Li NUS · Distributed Systems

Oberholzer-Gee · Frey

Harvard · Economics

Motivation crowding, markets, and behaviour.

Yao Lu

AI-native systems · Data & AI Systems

Large-scale analytics and AI-native infrastructure.

Contact us

Tell us what you are working on and the right team will reply by email.

Send opens your email app with this message filled in.