Shardul Bansal

I'm building the arena where AI learns to run a business.

I'm Shardul. With Seth Schilbe, I started ORO. We build high-fidelity simulations of real businesses so AI can learn the tacit knowledge it takes to run one.

Before this, I studied CS at the University of Toronto, worked with my adjunct professor Ala Shaabana, spent a couple of years at AWS, went through Entrepreneur First in London, tried to fix grid congestion, and then fell deep into decentralized ML on Bittensor. This is where I write about what I'm learning.

Now

SMB Bench

Our first simulation hands a model the produce buying desk of a wholesaler for a 90-day quarter, then plants consequences that only show up weeks after the decision. No frontier model reliably beats a fixed buying rule yet. Read the report.

Subnet 15

We host our environments on Bittensor. Miners hillclimb them, which shows us where an environment can be reward-hacked, and their runs give us traces of how frontier models behave.

ORO is backed by my adjunct UofT professor Ala Shaabana, David Lawee, Y Combinator and a number of others.

Core beliefs

  1. Work on the most important problem you can.

    Fazl Barez once asked me the Hamming question: what are the most important problems in your field, and if you're not working on them, why not? I was automating site assessments for HVAC contractors at the time, and I had no answer.

    From Solve the most important problems

  2. Bet on the model getting better.

    Base models keep absorbing whatever harness improvements we build around them. If a model has a capability gap, the longer-term play is to build the task that closes it and sell that task back to the model provider.

    From What we believe at ORO

  3. Let AI into the arena.

    Most of what makes someone good at their job is tacit, and they learn it from what their decisions lead to, often weeks later. AI needs autonomy inside a business to learn the same way. Simulations come first, then real businesses.

    From ORO's next chapter

  4. Get the incentives right first.

    An employee optimizes for their career and a founder for the company. A miner on our subnet optimizes for whatever the validator measures, and finds the cheapest path to it. If the measure is wrong, you pay for the wrong behaviour.

    From Principal agent problem

  5. Explore widely, then exploit relentlessly.

    In 2024 I left AWS, went through EF in London, worked on heat pumps and then grid congestion, and spent a month in Southeast Asia. Exploring that long was worth it, but only because I eventually picked one thing and went all in.

    From What kind of home do you want to create?

  6. Observe without judgement.

    Judgement leads to a reaction, and a reaction pulls you out of the moment. I first noticed this swimming laps as an 11-year-old, and found it again practicing Vipassana.

    From Observation without judgement

Essays

All writing

Thoughts

All thoughts