AI RL Environments, Evals, and Infrastructure Landscape

0.0(0)
Studied by 0 people
call kaiCall Kai
learnLearn
examPractice Test
spaced repetitionSpaced Repetition
heart puzzleMatch
flashcardsFlashcards
GameKnowt Play
Card Sorting

1/53

flashcard set

Earn XP

Description and Tags

Vocabulary flashcards detailing company profiles, benchmarks, platforms, and execution infrastructure across Tier 1 and Tier 2 of the reinforcement learning and AI evaluation landscape.

54 Terms

1
New cards

Halluminate

A Tier 2 Knowledge Work company that builds training environments for finance and consulting work, such as Finance Diligence Bench. It raised a 30M30M Series A led by Oak HC/FT and counts 4 of the 5 leading US AI labs as customers.

2
New cards

AfterQuery

A company selling expert-made training data, scoring guides, agent environments, and real-task tests sourced from a network of approximately 100,000100{,}000 verified professionals. It raised 30.5M30.5M and lists NVIDIA as a named customer.

3
New cards

Fleet

A company backed by Sequoia, Menlo, and SV Angel that sits on the OpenEnv committee and builds simulated copies of business software (CRMs, spreadsheets, browser, and desktop) for training office agents.

4
New cards

Collinear

A platform that runs a Simulation Lab of environments featuring simulated users and graders, designed for training across coding, computer use, and enterprise workflows.

5
New cards

Chakra Labs

A company that has raised approximately 10M10M to build exact software clones, task sets, and recorded expert examples for agents learning to operate computers.

6
New cards

Veris AI

A startup that raised an 8.5M8.5M seed round to build sandboxes mimicking real enterprise systems and users, targeting office workflows in telecom, fintech, and HR.

7
New cards

BenchFlow

An open-source environment lab that publishes the SkillsBench benchmark, env0 simulated workspaces, the PostTrain Arena, and an Apache-2.0 open-source runtime for RL environments.

8
New cards

General Reasoning

A London-based lab that has raised approximately 10.9M10.9M and operates OpenReward (an open resource for serving RL environments) and KellyBench.

9
New cards

Aptura

A team composed of alumni from Lazard and McKinsey that co-authored SpreadsheetBench 2, a benchmark evaluating full business spreadsheet workflows created from real financial reports.

10
New cards

Andon Labs

A developer of long-horizon benchmarks (Vending-Bench, Blueprint-Bench, Drone-Bench, and Butter-Bench) that runs real AI-operated businesses and maintains verified ties to Anthropic and xAI.

11
New cards

Good Start Labs

A gaming-focused company that raised 3.6M3.6M from General Catalyst and Inovia to sell game-based training environments, gameplay data, and benchmarks to AI labs in partnership with game publishers.

12
New cards

KellyBench

A benchmark managed by General Reasoning that tests betting-style decision-making where an AI agent must manage resources across many consecutive rounds.

13
New cards

Vmax

An auto-generation player backed by Race Capital and South Park Commons that converts a customer's proprietary data and tests into training environments and researches automated task generation.

14
New cards

Huzzle Labs

A company that raised 6M6M to build coding, computer-use, and enterprise environments, selling long-horizon RL environments and expert-recorded task examples directly to frontier labs.

15
New cards

Idler

A YC S25 company that raised a 9M9M seed led by Paradigm to build tests and training environments from real production work, including the ShelfLife and CorpLaw benchmarks.

16
New cards

Bespoke Labs

A company backed by 40M40M from Wing VC and 8VC that builds training environments and conducts custom post-training on simulated enterprise tasks, counting Intuit/Credit Karma as a customer.

17
New cards

Andromede

A Lausanne-based company in private beta backed by Unusual Ventures that automatically generates environments, tasks, and graders with checkable answers from real-world data.

18
New cards

Trajectory

A continual-learning platform co-founded by alumni from Google DeepMind and Windsurf focused on helping models improve over extended periods of real use (distinct from Trajectory Labs).

19
New cards

Polymath

A YC W26 company building world-generation models that generate simulated worlds for AI agents to learn independent work over long horizons.

20
New cards

Patronus AI

An auto-generation company that raised a 50M50M Series B at an approximately 406M406M valuation and grew revenue 15x in a year; it creates Generative Simulators that continually invent tasks and rules.

21
New cards

Quesma

A Warsaw-based company that creates independent tests for DevOps, security, and database tasks, including the Harbor-based BinaryAudit test.

22
New cards

BinaryAudit

A security benchmark built on Harbor by Quesma that tests AI models on finding hidden backdoors within compiled software, on which the top models achieve scores around 49%49\%.

23
New cards

ARIMLABS

A Warsaw security lab founded in 2025 providing red-teaming and training environments to prevent AI agents from being manipulated into abusing their system access.

24
New cards

Tacit Labs

An AI company that co-authored the LifeSciBench benchmark alongside OpenAI to evaluate models on complex biology research tasks.

25
New cards

LifeSciBench

A life-sciences benchmark co-authored by Tacit Labs and OpenAI containing 750 expert-written tasks across seven workflows evaluated against 19,020 expert criteria, where the best model scores 36.1%36.1\%.

26
New cards

Preference Model

A company founded by members of Anthropic's data team that builds environments for AI research and ML engineering; received investment from a16z in October 2026.

27
New cards

ULAM

An Oxford team based in Warsaw and London that releases the Math RL Tasks dataset, consisting of pure mathematical object construction tasks verified automatically.

28
New cards

Aviro

A YC Spring 2025 startup building environments for long, multi-step tool use covering ML research, live web tasks, and office operations.

29
New cards

HUD

A YC W25 authoring and management platform that raised 16M16M; it enables over 50 businesses to author RL environments and evals, with 2,500+ environments created.

30
New cards

NVIDIA NeMo Gym

An open-source toolkit and library by NVIDIA used for building RL environments and running practice attempts at scale; integrates with Harbor, OpenEnv, and various training frameworks.

31
New cards

Pareto

A human data platform that pivoted from a remote-work bootcamp in 2023 to translate expert judgment into data, evals, MCP tools, and RL environments for labs like Anthropic and Google DeepMind.

32
New cards

OpenReward

An open infrastructure platform hosted by London lab General Reasoning designed for serving, managing, and hosting reinforcement learning environments.

33
New cards

Prime Intellect

Known as the Hugging Face for RL environments; raised a 130M130M Series A at approximately a 1B1B valuation, offering an Environments Hub with 2,500+ community environments alongside compute sandboxes.

34
New cards

DataVendor

HUD's invite-only marketplace enabling creators to sell tasksets, environments, and agents to AI labs while retaining an 80%80\% revenue share.

35
New cards

OpenEnv

An open standard and hub for RL environments created by Meta (PyTorch) and Hugging Face, governed by a committee including Nvidia, Mercor, Fleet, Modal, Prime Intellect, and Microsoft.

36
New cards

Harbor

An open-source framework by the creators of Terminal-Bench used for hosting agent evals and building environments; serves as the official harness for Terminal-Bench 2.0.

37
New cards

Handshake

A college career platform whose AI division reached approximately 1B1B in annualized gross revenue and a 3.5B3.5B valuation; acquired Cleanlab to supply student/expert talent for evals and environments.

38
New cards

Uber AI Solutions

Uber's AI data division that leverages a workforce of 10M+ contributors for data labeling, collection, RLHF, and custom RL gym construction.

39
New cards

Contra

A commission-free freelance creative network of approximately 1.5M members that raised 50M+50M+; its Contra Labs division develops creative RLHF and the Human Creativity Benchmark.

40
New cards

Revelo

A Latin American network composed of 400k+ vetted software engineers operating as a code-focused data lab licensing expert-built coding environments.

41
New cards

Scale AI

A data-labeling incumbent in which Meta acquired a 49%49\% stake for 14.3B14.3B; provides RL Environments featuring simulated web applications, desktops, and tools such as Slack and HubSpot.

42
New cards

Surge AI

A bootstrapped human-data incumbent with 1.2B1.2B in 2024 revenue serving Google, Anthropic, and OpenAI; transformed RL environments into its primary revenue-driving product by August 2026.

43
New cards

Mercor

An expert network reaching 2B+2B+ in annualized revenue that acquired Sepal AI and Deeptune (software replicas of Excel, Salesforce, and Slack) and operates the APEX benchmarks.

44
New cards

micro1

An expert data provider exceeding a 1B1B run rate at a 4B4B valuation that committed 1B1B to license enterprise company data and sells environments under its Realm brand.

45
New cards

Snorkel AI

A Stanford spinout valued at 3.5B3.5B with a 375M375M run rate that provides finished datasets and expert-designed enterprise environments, such as insurance underwriting simulations.

46
New cards

Invisible Technologies

An AI training partner serving clients like Microsoft, Cohere, and AWS by building training environments and expert rubrics based on real accounting, banking, and legal workflows.

47
New cards

Toloka

A Nebius subsidiary serving Anthropic, Amazon, and Microsoft that produces RL gyms simulating CRM, ERP, and ticketing workflows across 11 distinct industries.

48
New cards

Turing

A remote-engineer talent marketplace and research partner valued at 2.2B2.2B with approximately 300M300M ARR that constructs coding and STEM RL gyms for major frontier AI labs.

49
New cards

Daytona

An execution infrastructure provider that raised a 60M60M Series B at a 1B1B valuation; provides cloud sandboxes capable of pausing, snapshotting, and self-copying for running agent RL.

50
New cards

Modal

A serverless AI cloud provider valued at 4.65B4.65B after raising a 355M355M Series C; its Sandboxes product reached an approximately 300M300M run rate executing untrusted AI code.

51
New cards

Runloop

A developer of Devboxes (secure, specialized cloud sandboxes raised via a 7M7M seed) designed for building, testing, and grading software engineering agents.

52
New cards

Browserbase

A cloud platform providing code-controlled headless browsers for AI agents; raised a 40M40M Series B and supports over 1,000 customers.

53
New cards

Cua

An open-source virtualization platform providing secure cloud desktops (supporting macOS, Windows, Linux, and Android) where agents interact via clicking and typing.

54
New cards

Morph

An execution infrastructure provider utilizing Infinibranch technology to snapshot and duplicate complete virtual machines in under 250 ms250\,\text{ms}, lowering the cost of parallel RL rollouts.