Key Takeaways

  • Abhishek Malani and Sequoia Capital partnered with OpenAI's Noam Brown and Citadel Securities to run "No Limit," an AI tournament that drew 7,000 collegiate applicants and 700 active bot builders.
  • The competition enforced a strict one-millisecond response latency, eliminating the ability to rely on heavy test-time compute.
  • Organizers added novel poker variants, including bidding rounds over a five-card flop, ensuring pre-computed solvers could not dominate the board.
  • The winning team from Carnegie Mellon built an opponent classification system that profiled rival playstyles (loose, tight, or hyper-aggressive) in under 15 hands.
  • The project moved from weekend concept to live collegiate finals in six weeks.

The Method

Standard AI hiring relies on polished take-home assignments or resume pedigree. Malani took a different route. Six weeks before the finals, he was playing poker with friends, watched recent model releases perform, and asked: “Can we build a poker bot?” (play from 1:30:41) That spark turned into "No Limit," a collegiate tournament co-run by Sequoia, Noam Brown from OpenAI, and Citadel Securities.

To separate genuine systems engineering from developers wrapper-building around public solvers, they built the competition around three design choices.

First, they broke public solvers. If you run a standard No-Limit Texas Hold'em bot tournament, someone downloads a canned solver and cleans up without writing novel code. Malani and his team altered the game mechanics by adding bidding mechanisms over a five-card flop. Suddenly, pre-computed game trees failed. Every model had to reason about value under modified auction rules.

Second, they capped latency at one millisecond. “This prevents you doing 30 minutes of test time compute before you make a decision,” (play from 1:34:36) Malani said. When you restrict an agent to one millisecond, you force builders to optimize small architectures, pre-calculate risk distributions, and manage tight inference budgets under hard constraints.

Third, the winning architectures focused on rapid dynamic classification. Carnegie Mellon's winning team realized that game-theory optimal play was too passive against unpredictable humans or erratic bots. Instead, they built an opponent classification engine. Within 15 hands, the bot analyzed betting patterns to classify an opponent as tight, loose, or prone to shoving, then shifted its strategy to exploit that specific weakness.

Where This Breaks Down

Tournaments like this reveal raw algorithmic speed and game-theory chops, but poker remains a closed-information game with clear mathematical boundaries. Even with altered flop bidding, the state space has rules.

Building production enterprise systems requires messy data ingestion, edge-case failure tolerance, and ambiguous requirements that no 15-hand poker match simulates. A developer who builds an aggressive opponent classifier in a 1-millisecond environment might struggle when asked to architect an asynchronous pipeline where API downtime and messy human inputs break formal logic. Fast classification works when the action space is constrained to check, call, bet, and fold.

What to Do With This

Replace your standard technical take-home test with a live, constrained simulation this week. Take an internal toy problem, modify one core business rule to break standard open-source libraries, and enforce a strict compute or time budget. Grade candidates on how quickly their system classifies anomalies and adapts under pressure.