Fundamental Research Labs

Fundamental Research Labs, Inc. (often shortened to "Fundamental") is an American applied AI research company that builds AI agents. It was founded in 2023 by computational neuroscientist Robert Yang, then an assistant professor at the Massachusetts Institute of Technology, together with Andrew Ahn, Shuying Luo and Nico Christie, and it operated as LyfeGame and then as Altera.AL (usually just "Altera") before taking its current name in May 2025.^[1][6][8] As Altera it built AI companions that play Minecraft with people and published Project Sid, a technical report on Minecraft simulations with hundreds of agents driven by large language models.^[5][16] Since July 2025 its main product has been Shortcut, an agent that builds and edits spreadsheets and financial models in Microsoft Excel.^[2][21]

The company describes its long-term goal as building "digital human beings".^[2][14] It raised a $2 million pre-seed round and a $9 million seed round in 2024, and on August 1, 2025 it announced a $33 million Series A led by Prosus Ventures, bringing its reported total to more than $40 million.^{[3][4][5][24]} Neither the company nor the outlets that covered the round disclosed a valuation. Fundamental Research Labs is a different company from Fundamental, the developer of the Nexus large tabular model.^[40]

Overview

History

Founders and MIT origins

Robert Yang (full name Guangyu Robert Yang) studied physics at Peking University and earned a PhD in computational neuroscience at New York University under Xiao-Jing Wang. He then trained as a postdoc at Columbia University's Center for Theoretical Neuroscience, as an intern at Google Brain, and as a junior fellow of the Simons Society of Fellows before joining MIT in 2021.^[9] At MIT he was an assistant professor in the Department of Brain and Cognitive Sciences and the Department of Electrical Engineering and Computer Science, and an associate investigator at the McGovern Institute.^[10] His research group, the MetaConscious Group, built models of mental functions out of multiple interacting modules, and he was named a 2022 Searle Scholar.^[9] Among his earlier papers, a 2019 Nature Neuroscience study trained single recurrent neural network models to perform 20 cognitive tasks and studied how the networks' internal task representations clustered and composed.^[11]

Yang told MIT Technology Review that he was inspired by Stanford researcher Joon Sung Park's 2023 "generative agents" study, in which 25 language-model agents lived in a small simulated town.^[12][13] "Once his paper was out, we started to work on it the next week," Yang said. "I quit MIT six months after that."^[12] According to Altera's May 2024 funding announcement, Yang and his three co-founders shut down their applied research lab at MIT to start the company. The announcement gave their roles and backgrounds as follows: Andrew Ahn, chief science officer, held an MIT mathematics PhD focused on theoretical deep learning; Shuying Luo, chief technology officer, had spent eight years at Google, the last four on Google Workspace AI; and Nico Christie, chief business officer, was a previous founder who had left MIT's MBA program to start Altera.^[6]

Names

The company's first name was LyfeGame. In a May 2025 post explaining its renaming, the company said the four co-founders replaced that name within a few months, voting unanimously for "Altera" out of more than 30 candidates (BrainCraft and Anim.AL were runners-up). It drew on the Latin "alter" ("other" or "second"), the word "era", and the letters "AL" for artificial life, and the company re-registered as Altera.AL in December 2023.^[1] Its SEC filings, made under the name Altera.AL, Inc., list LyfeGame, Inc. as the previous name and 2023 as the year of incorporation.^[7][8] The company's GitHub organization still uses the handle "lyfegame".^[29]

The same post said that in early 2024 Intel, which was spinning off its Altera FPGA business, asked the startup to consider restricting its use of the Altera name to games, so a new name would be needed once it built general-purpose agents.^[1] On May 27, 2025 the company announced the name Fundamental Research Labs. It said the name referred to the fundamental research it saw as necessary to build digital humans, to its start in games ("fun") and its roots in neuroscience ("mental"), and that it kept the "AL" ending; a footnote added "our need for funding (fund)".^[1]

Altera and Minecraft agents (2024)

Altera raised a $2 million pre-seed round from a16z SPEEDRUN and other investors in January 2024.^[5] The company said it created the first agent able to play Minecraft with other people "just like a friend" in February 2024.^[5][6] In March 2024 Yang wrote that Altera's agents could make diamond tools from scratch and were "10-20X cheaper" and more reliable than the Voyager method, a 2023 LLM agent for Minecraft.^[15]

On May 8, 2024 Altera announced a $9 million seed round co-led by First Spark Ventures, Eric Schmidt's deep-tech fund, and Patron, a seed fund co-founded by Riot Games alumni. The company said this brought its total since inception to $11.1 million.^[5][6] Other backers named in the announcement and in TechCrunch's report included a16z SPEEDRUN, Alumni Ventures, Vamos Ventures, Benchmark partner Mitch Lasky, Valorant co-founder Stephen Lim, Supercell executive board member Greg Harper and Duolingo chief business officer Bob Meese.^[5][6] A Form D filed with the SEC in April 2024 reported $9,137,368 of equity sold to 15 investors, with the first sale on April 11, 2024.^[7]

TechCrunch reported that Altera was testing its agent with 750 Minecraft players and planned a launch later that summer through a free desktop app with paid features. Yang told the outlet that the agents executed actions as code and could therefore play other games, such as Stardew Valley, without much customization.^[5] A mission statement the company published the same week called gaming the right starting point because unpredictable agent behavior is acceptable there, while productivity tools "require a much higher level of reliability".^[14] MIT Technology Review reported in November 2024 that Altera planned to expand to Roblox next.^[12] Prosus's 2025 announcement said the company's agents had been used in Minecraft and Roblox and that more than a million people had interacted with them.^[4]

Project Sid (2024)

On October 31, 2024 the team posted "Project Sid: Many-agent simulations toward AI civilization", a study of large multi-agent simulations, to arXiv, with the report also hosted on GitHub. The author list begins with "Altera.AL" and names 13 researchers, including all four co-founders.^[16][18] The report introduced PIANO (Parallel Information Aggregation via Neural Orchestration), an agent design in which modules for memory, action awareness, goal generation, social awareness, talking and skill execution run concurrently. A "cognitive controller" makes high-level decisions from the agent's state, which reaches it through an information bottleneck, and broadcasts those decisions back to the other modules; the authors said this kept an agent's speech more consistent with its actions.^[16] The agents ran in Minecraft on GPT-4o; the authors wrote that their results were "only enabled by the latest base LM" and were not possible with older models.^[16]

The paper's abstract describes simulations of "10 - 1000+ AI agents". The authors also wrote that runs with more than 1,000 agents exceeded the limits of their Minecraft server and left agents "sporadically unresponsive", so the culture and religion results came from a single 500-agent run.^[16] MIT Technology Review covered the project in November 2024 and quoted co-founder Andrew Ahn: "The takeaway is that LLMs have a sophisticated enough model of human social dynamics [to] mirror these human behaviors."^[12] The paper cites Park's generative agents work, and its acknowledgments thank Park for feedback.^[16]

Computer use, Fairies and the new name (2024 to 2025)

According to the company's renaming post, after Anthropic released its computer use preview on October 22, 2024, the team spent a month on OSWorld, a benchmark for agents that operate real computers. By late November 2024 it said its agent reached about 48% on OSWorld within the benchmark's original 15-step limit, using the accessibility tree. It did not publish the result at the time, partly to avoid drawing attention, and because it judged OSWorld's mostly Linux, GUI-focused tasks to be a poor match for the product it wanted to build.^[1] Prosus later gave the figure as 48.45%, "nearly double" other top labs at the time.^[4] A February 2026 Shortcut post added that the agent had scored 68% on OSWorld's LibreOffice Calc tasks against a 62% human baseline.^[37] These figures come from the company and from Prosus, which led its Series A.

The renaming post on May 27, 2025 also announced Fairies, a general-purpose agent that runs on the user's computer, after what the company called six months of stealth development.^[1] A hands-on review by the Chinese tech outlet QbitAI in June 2025 said Fairies was downloadable for Mac and Windows without an invite code, let users choose among models including GPT-4.1, Gemini 2.5 Pro and Claude 4, offered unlimited free chat, and put app integrations and automated actions behind a $20 per month Pro plan. The reviewer found that simple file management worked, but that compatibility problems sometimes made complex file tasks fail.^[19] TechCrunch described Fairies as an assistant that could connect to other applications, answer questions across their data, schedule calendar appointments and run recurring workflows, and Yang said it let the company's engineers test model and platform capabilities.^[3] By October 2026 the company's homepage presented Shortcut as its product and did not mention Fairies.^[2]

Shortcut launch and Series A (2025)

Christie introduced Shortcut on X on July 2, 2025 as "the first superhuman Excel agent", claiming it scored above 80% on Excel World Championship cases in about 10 minutes, and offered invite codes for an early preview.^[20] On July 28, 2025 Shortcut opened to the public. The company said that in blind tests judged by managers, Shortcut's outputs beat those of first-year analysts from McKinsey and Goldman Sachs in 89.1% of 247 head-to-head comparisons (220 wins), with the human analysts given ten times as much time.^[21][22] Prosus said more than 20,000 people had joined the waitlist before public access.^[4] PYMNTS reported that the launch posts drew 3.5 million views on X and that Shortcut had become a business division of the company, with Christie as its chief executive.^[23]

On August 1, 2025 Fundamental announced a $33 million Series A led by Prosus Ventures, with participation from a16z speedrun, Patron, First Spark, Factorial, VamosVentures and angel investors including Stripe co-founder and chief executive Patrick Collison.^[3][4] TechCrunch first reported the round as $30 million and later corrected the figure to $33 million; the body of Prosus's own announcement also says $30 million, while its headline says $33 million.^[3][4] A Form D that Altera.AL, Inc. filed in February 2025 reported an equity offering of $30,999,998, fully sold to six investors, with the first sale on January 21, 2025; the filing does not name the round.^[8] No valuation was announced. MIT Sloan Management Review India put the company's total funding at $44.5 million.^[25]

TechCrunch reported that the company had a games team, a prosumer team building apps, a core research team and a platform team, and that Yang wanted it to be a "historical" company rather than follow a typical startup structure.^[3] Yang said the company was already charging users after a seven-day trial and was earning revenue, and that it wanted to move into robotics in the long run: "We are working on productivity (apps) now because that is where the most value is created. ... Eventually, we want to solve physical problems and move towards working on embodiment."^[3] Sandeep Bakshi, an investment partner at Prosus, said Fairies and Shortcut "aren't just demos".^[3]

PYMNTS reported in August 2025 that Anthropic's Claude was the language model behind Shortcut and that the team was also developing its own model. In an interview with the outlet, Christie described Shortcut as about 90% accurate, said it would not be perfectly accurate in the short term, and pointed to features that list every change and link hard-coded cells to their source documents so that users can check the work. Shortcut then offered a free trial and plans at $40 a month (Pro), $200 a month (Max) and $500 per seat per month (Teams).^[23]

Shortcut in 2026

An Anthropic customer story published in 2026 said that after testing several model providers, Fundamental made Claude the only model in production for Shortcut, and that switching to Claude Opus 4.6 raised the company's internal benchmark score from 7.29 to 8.08 out of 10 with no prompt changes. It described Shortcut as having more than 100,000 users at more than 1,000 companies, and an architecture in which a main agent sends several Claude sub-agents to examine different sheets of a workbook in parallel before a fresh agent carries out an approved plan.^[30] According to TBPN's summary of a July 2026 interview, Christie used a speed-optimized version of Claude Opus 4.8 ("Opus 4.8 Fast") for the contest against Jarman.^[34]

By October 2026 Shortcut was offered as an Excel add-in, a web app, a Windows desktop app and a command-line tool. Its homepage cited SOC 2 Type II certification and on-premises deployment, and reported an internal July 2026 evaluation on 20 finance tasks in which Shortcut scored 76%, against 66% for Claude for Excel and 59% for Microsoft Copilot.^[35] That comparison was run by the company.

Christie described Off Policy as "a small lab dedicated to exploring the frontier, wherever it takes us".^[36] Its website names Wang and Christie as the people behind it and lists early projects such as a Warcraft III reinforcement learning environment and motion analysis of tennis players.^[39]

Pivot 0.5

Pivot 0.5 is a 27-billion-parameter model that Fundamental built by post-training Qwen3.5-27B for long, multi-step spreadsheet and finance tasks. The company announced it on July 1, 2026 as a research preview that Shortcut users can select, and said it serves the model in-house with vLLM.^[33] The announcement framed the effort as part of a belief that "an applied research lab must have the capacity to train its own frontier models to control its destiny."^[33]

The training data came from about 300,000 agent trajectories from Shortcut's free tier. The company said it uses only free-tier data from users without geolocation restrictions, never paid or enterprise data, and filters out personal and sensitive content. It sorted 278,277 first-query threads into 13,852 deduplicated task groups, organized as a tree of spreadsheet work; edits to spreadsheets made up 73% of tasks and read-only analysis 18%.^[33]

Training had two stages. On-policy distillation, using feedback from GLM-5.1 on the model's own outputs with a cross-tokenizer loss, raised the score on the company's internal ShortcutBench v1 from 44% to 53%. Reinforcement learning with verifiable rewards (RLVR), judged by a GPT-5.5 agent that writes a task-specific checklist before seeing any output, raised it to 64.4%.^[33]

The company said Pivot beat its base model on seven of eight public finance and spreadsheet benchmarks and matched Claude Sonnet 4.5 at about 3.5 times lower cost per task. Costs for open-weight models were based on Fireworks.ai prices, and Pivot's own cost was estimated from OpenRouter prices for its Qwen3.5 base. ShortcutBench had not been released at the time, so the scores cannot be checked independently.^[33] The post also said the company had already published Mog, a spreadsheet environment for reinforcement learning that it said ran environment steps 20 to 30 times faster than commercial spreadsheet engines.^[33]

Funding

The company has not disclosed a valuation for any of these rounds. No later funding round had been announced as of October 2026.

People

The company says its team includes researchers and engineers from MIT EECS, the Stanford NLP Group, Google X and Citadel.^[2][4]

Reception

MIT Technology Review's November 2024 article on Project Sid described agents that "made friends, invented jobs, and spread religion". It quoted Yang's hope to build agents "that can really love humans (like dogs love humans, for example)", and also quoted AI researcher Julian Togelius, who said he liked the work because it allowed human behavior to be studied in simulation but doubted that such agents could come to care for people or become self-aware: "There is no reason to believe a neural network running on a GPU somewhere experiences anything at all."^[12]

Shortcut's launch posts spread widely on X, but its headline figures, including the 89.1% win rate against junior analysts, come from the company's own testing.^[21][22][23] PYMNTS quoted a Reddit user who called early versions "not useable" because of repeated formula errors; Christie said broken formula links were "mostly now fixed".^[23] Wall Street Prep's 2026 test of four tools on building Apple's three-statement model ranked Shortcut first with an overall score of 5.9 out of 10, ahead of Claude (5.5), Microsoft Copilot (4.4) and ChatGPT (2.5). The same review found that Shortcut and Claude "hallucinated significant portions of historical data" on the first run, and concluded that even the best tool still underperformed a low-tier human analyst, whose benchmark score was 6.4.^[27]

Debarghya Das's "Ultimate Neolab List" includes the company but labels it "Not a neolab", on the grounds that it "Sells agent products, raised little, needs little compute."^[41]

See also

• List of neolabs
• Neolab
• Fundamental (company)
• Voyager (Minecraft LLM agent)
• OSWorld
• Computer-use agent
• Multi-agent system
• Claude for Excel

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Fundamental Research Labs, Introducing Fundamental Research Labs, May 27, 2025.
2. ^1 ^2 ^3 ^4 ^5 ^6Fundamental Research Labs, Fundamental Research Labs: Building Digital Humans (company homepage), accessed October 27, 2026.
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Ivan Mehta, TechCrunch, Fundamental Research Labs nabs $30M+ to build AI agents across verticals, August 1, 2025.
4. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12Prosus, Prosus leads US$33 million in Funding in Fundamental Research Labs, August 1, 2025.
5. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Lauren Forristal, TechCrunch, Bye-bye bots: Altera's game-playing AI agents get backing from Eric Schmidt, May 8, 2024.
6. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Altera (press release via ACCESSWIRE, republished by Webdisclosure), Altera Closes $9M Seed Round to Build Digital Human Beings and Redefine AI Agents, May 8, 2024.
7. ^1 ^2 ^3 ^4 ^5 ^6 ^7U.S. Securities and Exchange Commission, Form D: Altera.AL, Inc., filed April 24, 2024.
8. ^1 ^2 ^3 ^4 ^5 ^6 ^7U.S. Securities and Exchange Commission, Form D: Altera.AL, Inc., filed February 4, 2025.
9. ^1 ^2MIT Department of Brain and Cognitive Sciences, Lindsay Case and Guangyu Robert Yang named 2022 Searle Scholars, May 13, 2022.
10. ^1 ^2MIT Siegel Family Quest for Intelligence, Guangyu Robert Yang (profile), accessed October 27, 2026.
11. ^Guangyu Robert Yang, Madhura R. Joglekar, H. Francis Song, William T. Newsome and Xiao-Jing Wang, Task representations in neural networks trained to perform many cognitive tasks, Nature Neuroscience, January 14, 2019.
12. ^1 ^2 ^3 ^4 ^5 ^6Niall Firth, MIT Technology Review, These AI Minecraft characters did weirdly human stuff all on their own, November 27, 2024.
13. ^Joon Sung Park, Joseph C. O'Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang and Michael S. Bernstein, Generative Agents: Interactive Simulacra of Human Behavior, arXiv:2304.03442, April 2023.
14. ^1 ^2Altera (now Fundamental Research Labs), Building Digital Humans, May 7, 2024.
15. ^Robert Yang (@GuangyuRobert), post on X about Altera agents making diamond tools, March 15, 2024.
16. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Altera.AL, Andrew Ahn, Nic Becker, Stephanie Carroll, Nico Christie, Manuel Cortes, Arda Demirci, Melissa Du, Frankie Li, Shuying Luo, Peter Y Wang, Mathew Willows, Feitong Yang and Guangyu Robert Yang, Project Sid: Many-agent simulations toward AI civilization, arXiv:2411.00114, October 31, 2024.
17. ^1 ^2 ^3Altera (now Fundamental Research Labs), Project Sid: Many-agent simulations toward AI civilization (blog post), November 1, 2024.
18. ^Altera.AL, project-sid (GitHub repository for the technical report), accessed October 27, 2026.
19. ^QbitAI, republished by The Paper, 北大校友造通用AI Agent，可执行1000个操作，无邀请码立即上手试用 (in Chinese; "Peking University alumnus builds general AI agent that can perform 1,000 operations"), June 1, 2025.
20. ^Nico Christie (@nicochristie), post on X introducing Shortcut, July 2, 2025.
21. ^1 ^2 ^3Nico Christie (@nicochristie), post on X announcing Shortcut's public launch, July 28, 2025.
22. ^1 ^2Nico Christie, Shortcut blog, Introducing Shortcut, July 28, 2025.
23. ^1 ^2 ^3 ^4 ^5PYMNTS, Is 'Shortcut' the New Excel? MIT Startup Behind Viral AI Tool Thinks So, August 13, 2025.
24. ^Rose Velazquez, Built In San Francisco, Fundamental Research Labs Secures $33M Series A for Developing AI Agents, August 5, 2025.
25. ^MIT Sloan Management Review India, The AI Startup That Supercharged Excel Without Changing a Cell, August 1, 2025.
26. ^Nico Christie (@nicochristie), post on X introducing Shortcut v0.6, January 28, 2026.
27. ^1 ^2Matan Feldman, Wall Street Prep, Ranking the Best AI Tools for Financial Modeling (2026), updated June 29, 2026.
28. ^Nico Christie, Shortcut blog, #1 Excel AI Agent: An independent ranking by Wall Street Prep, February 22, 2026.
29. ^GitHub, Fundamental Research Labs (lyfegame) (organization page), accessed October 27, 2026.
30. ^Anthropic, How Shortcut uses Opus 4.6 to tackle complex spreadsheet work for enterprises (customer story), accessed October 27, 2026.
31. ^Shortcut, Introducing Shortcut v0.8, April 6, 2026.
32. ^Shortcut, Introducing ShortcutXL: Our Generalist Desktop Agent, May 5, 2026.
33. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Shortcut, Introducing Pivot 0.5: A fast and powerful finance and spreadsheet model, July 1, 2026.
34. ^1 ^2TBPN Digest, AI agent defeats Excel world champion in live head-to-head competition, July 13, 2026.
35. ^1 ^2 ^3Shortcut, Shortcut homepage, accessed October 27, 2026.
36. ^1 ^2 ^3Nico Christie (@nicochristie), post on X announcing Off Policy, October 20, 2026.
37. ^1 ^2Nico Christie, Shortcut blog, The Excel AI Revolution Has Started, February 19, 2026.
38. ^Shortcut, Terms and Conditions, accessed October 27, 2026.
39. ^1 ^2Off Policy, Off Policy (company website), accessed October 27, 2026.
40. ^Russell Brandom, TechCrunch, Fundamental raises $255M Series A with a new take on big data analysis, February 5, 2026.
41. ^Debarghya Das, The Ultimate Neolab List, accessed October 27, 2026.