A new AI coding challenge just published its first results — and they aren’t pretty | TechCrunch

A new AI coding challenge has revealed its first winner — and set a new bar for AI-powered software engineers.

On Wednesday at 5 p.m. PT, the nonprofit Laude Institute announced the first winner of the K Prize, a multi-round AI coding challenge launched by Databricks and Perplexity co-founder Andy Konwinski. The winner was a Brazilian prompt engineer named Eduardo Rocha de Andrade, who will receive $50,000 for the prize. But more surprising than the win was his final score: He won with correct answers to just 7.5% of the questions on the test.

“We’re glad we built a benchmark that is actually hard,” said Konwinski. “Benchmarks should be hard if they’re going to matter,” he continued, adding: “Scores would be different if the big labs had entered with their biggest models. But that’s kind of the point. K Prize runs offline with limited compute, so it favors smaller and open models. I love that. It levels the playing field.”

Konwinski has pledged $1 million to the first open source model that can score higher than 90% on the test.

Similar to the well-known SWE-Bench system, the K Prize tests models against flagged issues from GitHub as a test of how well models can deal with real-world programming problems. But while SWE-Bench is based on a fixed set of problems that models can train against, the K Prize is designed as a “contamination-free version of SWE-Bench,” using a timed entry system to guard against any benchmark-specific training. For round one, models were due by March 12. The K Prize organizers then built the test using only GitHub issues flagged after that date.

The 7.5% top score stands in marked contrast to SWE-Bench itself, which currently shows a 75% top score on its easier “Verified” test and 34% on its harder “Full” test. Konwinski still isn’t sure whether the disparity is due to contamination on SWE-Bench or just the challenge of collecting new issues from GitHub, but he expects the K Prize project to answer the question soon.

“As we get more runs of the thing, we’ll have a better sense,” he told TechCrunch, “because we expect people to adapt to the dynamics of competing on this every few months.”

It might seem like an odd place to fall short, given the wide range of AI coding tools already publicly available — but with benchmarks becoming too easy, many critics see projects like the K Prize as a necessary step toward solving AI’s growing evaluation problem.

“I’m quite bullish about building new tests for existing benchmarks,” says Princeton researcher Sayash Kapoor, who put forward a similar idea in a recent paper. “Without such experiments, we can’t actually tell if the issue is contamination, or even just targeting the SWE-Bench leaderboard with a human in the loop.”

For Konwinski, it’s not just a better benchmark, but an open challenge to the rest of the industry. “If you listen to the hype, it’s like we should be seeing AI doctors and AI lawyers and AI software engineers, and that’s just not true,” he says. “If we can’t even get more than 10% on a contamination-free SWE-Bench, that’s the reality check for me.”