Sciloop
Sciloop pays STEM olympiad medalists to write ultra-hard problems for frontier AI labs to test on.
On the record
Checkpoint: Open-source Logics-STEM reasoning dataset release, which could commoditize large-scale STEM problem sets
The call
Sciloop sells expert-written problems into a market where bigger, richer data vendors already operate, with no disclosed customers and a product whose value shrinks as models improve.
Sciloop provides ultra-difficult STEM problems (like olympiad-level math) to companies training advanced AI models. The problem is that well-funded rivals like Mercor and Scale AI can hire the same scarce talent to create similar problems, and AI systems are already starting to solve these hard problems on their own. Meanwhile, free open-source problem collections are appearing online, which means companies may not need to pay for Sciloop's data anymore. The one thing that could change this is if Sciloop's problems stay genuinely harder than what AI can do, but that window is closing as AI improves every few months.
Why
01
AI pressure: Multi-agent AI systems and frontier models (GPT-5.4 Pro, Gemini 3.1 Pro) closing the gap on olympiad-level problems. A multi-agent system already won physics-olympiad gold in Sept 2025, narrowing the exact gap Sciloop's problems are built to expose. Could AI do the core job? Yes, likely.
02
Bigger players: Mercor, Scale AI, Surge AI. Mercor ($10B valuation), Scale AI (~$950M ARR), and Surge AI (~$1.2B revenue) already run expert-data pipelines for the same labs at far larger scale. Could a giant copy it? Yes, easily.
03
What protects it. Founders are IPhO medalists who can personally vet and source problems models still fail, a narrow credibility edge with no lock-in behind it.
What would change this call
A frontier lab reports over 5% model accuracy on Sciloop's advertised hardest problems. Shows Sciloop's difficulty moat has closed and its core pitch is losing relevance.
Sciloop discloses a named lab customer or revenue figure within 9 months of its pivot. Would indicate the new expert-data model is finding paying demand.
Mercor's client retention after its March 2026 LiteLLM breach becomes public. Shows whether labs are shifting spend toward smaller vendors like Sciloop or staying with incumbents.
Sciloop discloses a named lab customer or revenue figure within 9 months of its pivot. Would indicate the new expert-data model is finding paying demand.
Mercor's client retention after its March 2026 LiteLLM breach becomes public. Shows whether labs are shifting spend toward smaller vendors like Sciloop or staying with incumbents.
Key numbers
- YC F25 batch, ~$500,000 seed via SAFE
- Pays problem writers $225+ per problem
- Claims frontier models score 0-5% on its hardest problems
Named risks
- Frontier models closing the performance gap on olympiad-level STEM problems, undercutting core value proposition.
- No disclosed revenue, ARR, or named lab customers as of July 2026.
- Two-person team competing against Mercor, Scale, and Surge for same scarce expert talent pool.
- Business model already pivoted once (Oct 2025 to mid-2026), signaling ongoing product-market-fit search.
- Open-source STEM reasoning corpora (Jan 2026) could reduce lab demand for paid boutique data vendors.
Rivals, as read in July 2026
| Rival | Direction | Evidence |
|---|---|---|
| Mercor | losing to | Mercor hit $10B valuation and ~$500M ARR run-rate by October 2025, recruiting same expert pool |
| Scale AI | losing to | Scale reported ~$950M ARR by August 2024 with ~500 employees, far outscaling Sciloop |
| Surge AI | losing to | Surge reportedly crossed $1.2B revenue in 2024 with 1 million annotators serving major labs |
| Turing | contesting | Turing publishes its own STEM datasets and VLM-Bench, directly overlapping Sciloop's pitch |
| Datacurve | contesting | YC W24 peer raised $15M for frontier coding data, adjacent niche to Sciloop's STEM focus |
Questions people ask
- Is Sciloop profitable in 2026?
- No disclosed revenue, ARR, or customer figures exist for Sciloop as of July 2026. The two-person, YC F25 startup has not published any traction numbers, and its earlier product (an AI research agent) also had no quantified metrics, per its October 2025 YC launch page.
- Could AI replace what Sciloop sells?
- Yes, that is the core risk. Sciloop's pitch rests on models scoring 0-5% on its hardest problems, but a multi-agent system already won physics olympiad gold in a September 2025 arXiv paper, directly closing the gap Sciloop monetizes.
- Who is Sciloop's biggest competitor?
- Mercor, which raised $350M at a $10B valuation in October 2025 with a reported run-rate near $500M ARR, competing for the same scarce pool of IMO and IPhO medalists Sciloop relies on for problem writing.
- What does Sciloop actually do now?
- As of July 2026, Sciloop sells expert-authored, extremely hard STEM problems (math, physics, expanding to chemistry, biology, law) as training and evaluation data to frontier AI labs, paying problem writers $225-plus per problem via a piece-rate model.
- Who owns or funds Sciloop?
- Sciloop is privately held with no parent company, founded in 2025 by MIT-affiliated IPhO 2022 medalists Bilal Asmatullah and Osman Siddique, and backed by Y Combinator's F25 batch, which typically provides about $500,000 in SAFE seed funding.
- Did Sciloop change its business model?
- Yes. Sciloop launched on YC in October 2025 as 'The AI Scientific Researcher,' an autonomous ML-experiment agent, then by mid-2026 repositioned entirely as an expert STEM reasoning data vendor for frontier AI labs, a full pivot within about nine months.
Timeline
- 2026-03Mercor suffers LiteLLM supply-chain breach exposing up to 4TB of data, clients including Meta pause work
- 2026-01Open-source 'Logics-STEM' 10M-scale STEM reasoning dataset published, threatening boutique data vendors
- 2025-10-28Sciloop launches on YC as 'The AI Scientific Researcher' with autonomous ML-agent product
- 2025-10Mercor raises $350M Series C at $10B valuation, ~$500M ARR run-rate reported
- 2025-09-29Multi-agent AI system wins gold-medal-level physics olympiad performance per arXiv paper
- 2025-09OpenAI publishes GDPval framework, shifting labs toward expert-judgment task evaluation
- 2025Sciloop founded, joins Y Combinator Fall 2025 (F25) batch, San Francisco
- 2024-08Scale AI reported ARR reaches approximately $950M
- 2024Surge AI reportedly crosses $1.2B in annual revenue with about 1 million annotators
Revision history
27 July 2026
First assessed. Little protection left. The text above is fixed as of this date and is not edited afterwards.
In the same band
Sources
- Sciloop: Expert STEM reasoning data for frontier AI labs
- wikipedia.org (Y Combinator Programs) (2026-07-27)
- Launch YC: Sciloop: The AI Scientific Researcher (2025-10-28)
- Jobs at Sciloop
- Sciloop, The Hardest Problems, Built by the Best Minds
- Data Annotator Recruiting for AI Labs
- Report: Mercor Business Breakdown & Founding Story
- LinkedIn for data labelers (2025-01-07)
- STEM Datasets and Evaluation
- Datacurve: Frontier coding data for training and evaluating LLMs
- Winning Gold Medals in the Latest Physics Olympiads with a Coevolutionary Multimodal Multi-Agent System (2025-09-29)
- Empowering LLM Reasoning via Failure-Driven Post-Training and Document Knowledge Enhancement (2026-01-19)