The YC startup is selling RL environments built from production data as Paradigm puts its broader frontier-tech mandate to work.
By Ryan Merket · Published
Primary source: Idler on X
Why it matters #
Idler is betting that proprietary production data, rather than another model wrapper, will become valuable infrastructure for post-training AI agents. Paradigm's lead also shows how quickly its new AI mandate is moving from fund announcement to seed deals.
Nalu Concepcion (@naluconcepcion), Ivan Chub (@chubivan), and Tony Goss (@cha0sg0d_) publicly launched Idler on Wednesday as a research lab building evaluation datasets and reinforcement learning environments for frontier AI models. In a thread on X, Idler also disclosed a $9M seed round led by Paradigm investors Alpin Yukseloglu (@0xalpo) and Frankie (@FrankieIsLost).
https://x.com/idler_ai/status/2090153772672754012 The financing appears to predate Wednesday's announcement. A Cornell entrepreneurship bulletin dated December 16th, 2025 referenced Idler's $9M seed round, a foundation-model lab contract, and profitability. Wednesday's news is therefore Idler's public unveiling and investor disclosure, rather than evidence of a newly closed round.
Y Combinator lists Idler as a San Francisco company founded in 2025 and part of its Summer 2025 batch. Idler currently has 12 employees listed on the accelerator's directory.
From game worlds to model training
Idler's founders arrived at AI training through games and operational software. Concepcion studied at Cornell and previously worked in software engineering at Microsoft, product at Poshmark, and data science at PayMongo. She has described Idler's work as "scaling RL environments," a job that followed an earlier phase organizing events with more than 500 attendees.
Chub previously co-founded Dark Forest, an open-source strategy game built on Ethereum using zero-knowledge proofs, as well as the applied cryptography group 0xPARC and the digital credential project Zupass. Goss says he first got Chub's attention by finding a legal exploit in Dark Forest that helped him win a prize NFT for which a collector offered $400,000.
That history shaped Idler's original framing: environments are games for AI agents, complete with objectives, tools, state, and verifiable outcomes. The founders initially explored automating small-business workflows before moving to San Francisco for Y Combinator and concentrating on training environments for model labs.
Idler says that over the past year it developed thousands of coding evaluations and environments for an unnamed frontier lab. Idler also claims the customer's models improved eightfold at long-horizon coding during that period, forcing the environments to become harder as the models advanced.
ShelfLife turns a retailer into a training ground
Idler's first public benchmark, ShelfLife, is a digital copy of an operating e-commerce business. Idler says the underlying retailer runs nine stores, three commerce platforms, and a sourcing operation involving multiple suppliers.
The benchmark contains 200 environments drawn from nearly a decade of operational data. Tasks cover work including accounting, marketing, finance, due diligence, planning, disclosure control, and inventory decisions. Idler says each environment was built with executives and operators who performed the original work.
Results measured on August 13th put Claude Opus 5 at a 69% single-attempt pass rate. GPT-5.6 Sol scored 55.7%, Claude Fable 5 scored 54.7%, Grok 4.6 scored 39.4%, Kimi K3 scored 30.6%, and Gemini 3.6 Flash scored 16.1%, according to Idler's leaderboard. These are Idler's own evaluations, and their value will depend on how closely the tasks reflect work that enterprises actually need agents to complete.
Idler also reports that post-training Nemotron 3 Nano on 41 ShelfLife tasks increased its Finance Agent Benchmark score from 30% to 41.8%, with a reported p-value of 0.001 or lower. Idler has made the full results and datasets available by request rather than as a public download.
The catalog already extends beyond the headline benchmark. ShelfLife E-Sim places agents in a stateful retail simulation where they operate a business one day at a time. CorpLaw evaluates agents inside transaction workspaces assembled from anonymized law-firm records. Idler says 10 of CorpLaw's 50 tasks remained unsolved by all six models it tested at high reasoning effort.
Paradigm puts its new mandate to work
Paradigm's lead is part of a broader shift at a firm historically associated with cryptocurrency. Paradigm raised a $1.2B fund in July to invest across AI, robotics, and other frontier technologies while continuing its crypto strategy. Idler gives Paradigm a bridge between those mandates: Chub and Goss came out of crypto-native game development, while Idler sells infrastructure directly into the AI model-training market.
The round also included Y Combinator (@ycombinator), Long Journey (@LongJourneyVC), and angels including Poshmark founder Manish Chandra (@marrc), Simple AI co-founder Catheryn Li (@catheryn_li), Daimo co-founder Daniel Posch (@dcposch), and Edge City founder Janine Leger (@JanineLeger).
Idler's route to scale depends on acquiring proprietary work rather than generating synthetic exercises. Idler is asking businesses with 10 to 100 employees to license codebases or de-identified operating data, giving those businesses an additional revenue stream while supplying Idler with material for new environments.
The $9M round buys Idler time to turn bespoke work for individual labs into a repeatable catalog spanning software engineering, cybersecurity, law, finance, and operations. Every improvement in frontier models shortens the useful life of an evaluation, however. Idler is selling the test while committing itself to keep rewriting it.