# Recursive Superintelligence

> Source: https://aiwiki.ai/wiki/recursive_superintelligence
> Updated: 2026-10-27
> Fact-checked: 2026-10-27
> Categories: AI Companies, AI Research
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Recursive Superintelligence." aiwiki.ai, 27 Oct 2026. https://aiwiki.ai/wiki/recursive_superintelligence
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Recursive Superintelligence** (branded as **Recursive**) is an [artificial intelligence](https://aiwiki.ai/wiki/artificial_intelligence) research company that came out of stealth on May 13, 2026 with roughly $650 million in funding at a company-stated $4.65 billion valuation [1][2][3]. It was founded by a group of eight researchers led by [Richard Socher](https://aiwiki.ai/wiki/richard_socher), the former chief scientist of [Salesforce](https://aiwiki.ai/wiki/salesforce) and founder of the search company [You.com](https://aiwiki.ai/wiki/you_com), and its stated goal is to build recursively self-improving AI: systems that run the research loop of proposing, implementing, and testing improvements to AI with little human help [2][9][18]. Its central project is what it calls the Eureka Machine, an automated AI research system that it applies first to the pipeline for building AI models [12][38]. The company has offices in San Francisco and London and had fewer than 30 employees at launch [1][18][37].

Since launch the company has published early benchmark results from its automated research system (June 2026), signed a $410 million multi-year compute agreement with [Amazon Web Services](https://aiwiki.ai/wiki/amazon_web_services) (July 2026), and released its first model, [Recursive Flint](https://aiwiki.ai/wiki/recursive_flint), an open-weight coding model, on October 21, 2026 [12][13][14].

Despite a near identical sounding name, Recursive Superintelligence is a separate company from [Ricursive Intelligence](https://aiwiki.ai/wiki/ricursive_intelligence), the chip design AI startup founded by Anna Goldie and Azalia Mirhoseini, the former Google researchers behind the AI system Google uses to help design its TPU chips [2][40]. The New York Times noted that Ricursive was pursuing a similar goal and was also valued at about $4 billion [19], but the two have different founders, locations, and product focus.

## Founding and founders

The company's UK entity, Recursive Superintelligence Ltd., was incorporated in London on December 31, 2025, with [Tim Rocktäschel](https://aiwiki.ai/wiki/tim_rocktaschel) listed as its director, according to Companies House records [20][21]. The company's website and launch coverage also use the name Recursive Superintelligence, Inc. [2][38]. Launch coverage described the company as four months old (The Next Web) or six months old (The New York Times) when it exited stealth in May 2026 [3][18], and both Earlybird and Tech Funding News give 2025 as its founding year [8][24].

Socher serves as chief executive [9]. Before this venture he was Salesforce's chief scientist, an early author on the [ImageNet](https://aiwiki.ai/wiki/imagenet) project, and the founder of You.com, an [AI search](https://aiwiki.ai/wiki/ai_search) company that SiliconANGLE reported had been valued at $1.5 billion the previous year [2][5]. He remained You.com's chief executive while leading Recursive; The New York Times corrected its launch story, which had initially described him as You.com's former chief executive [18]. He also continued to run his venture fund, AIX Ventures [11].

The other co-founders are an unusually senior group drawn from [Google DeepMind](https://aiwiki.ai/wiki/google_deepmind), [OpenAI](https://aiwiki.ai/wiki/openai), [Meta AI](https://aiwiki.ai/wiki/meta_ai)'s Fundamental AI Research lab (FAIR), Salesforce, and Uber AI [3][14]. Tech Funding News reported in January 2026, months before the launch, that the startup had eight co-founders [22]. Lead investor GV's investment note, The Next Web, and Earlybird's portfolio page each name the same eight [3][9][24]:

| Co-founder | Prior affiliation |
| --- | --- |
| [Richard Socher](https://aiwiki.ai/wiki/richard_socher) (CEO) | Chief scientist, [Salesforce](https://aiwiki.ai/wiki/salesforce); founder and CEO, [You.com](https://aiwiki.ai/wiki/you_com) [2][18] |
| [Tim Rocktäschel](https://aiwiki.ai/wiki/tim_rocktaschel) | Professor of AI, University College London; led the Open-Endedness team at [Google DeepMind](https://aiwiki.ai/wiki/google_deepmind) [4][5][23] |
| Yuandong Tian | Research scientist director, [Meta FAIR](https://aiwiki.ai/wiki/meta_fair) [3][33] |
| Alexey Dosovitskiy | Lead author of the [Vision Transformer](https://aiwiki.ai/wiki/vision_transformer) (ViT) paper at Google Research; later at Inceptive [32][34] |
| Josh Tobin | [OpenAI](https://aiwiki.ai/wiki/openai), where he led work on Codex and deep research [5] |
| Caiming Xiong | Senior vice president, [Salesforce AI](https://aiwiki.ai/wiki/salesforce_ai) Research [35] |
| Tim Shi | Co-founder of [Cresta](https://aiwiki.ai/wiki/cresta_ai); formerly [OpenAI](https://aiwiki.ai/wiki/openai) [5][19] |
| [Jeff Clune](https://aiwiki.ai/wiki/jeff_clune) | Professor, University of British Columbia; formerly [OpenAI](https://aiwiki.ai/wiki/openai); known for AI-generating algorithms [17][19][29] |

The New York Times reported that the company had also hired [Peter Norvig](https://aiwiki.ai/wiki/peter_norvig), Google's former director of research and co-author of the standard textbook Artificial Intelligence: A Modern Approach, while The Next Web described Norvig as an adviser [3][19].

The roster is notable for how concentrated it is around one research subculture. Rocktäschel and Clune both spent years on open-endedness, the study of algorithms that keep generating novelty indefinitely rather than converging on a single fixed objective. Rocktäschel ran the Open-Endedness team at DeepMind and worked on [world model](https://aiwiki.ai/wiki/world_model)s, including [Genie](https://aiwiki.ai/wiki/genie), whose 2024 paper lists both him and Clune among its authors, and [Genie 3](https://aiwiki.ai/wiki/genie_3) [5][27]. He is also a co-author of Rainbow Teaming, a 2024 method for open-ended generation of diverse adversarial prompts [28]. Clune set out the idea of "AI-generating algorithms" in a 2019 paper and co-authored The AI Scientist (2024), a system for automating machine learning research, and the Darwin Gödel Machine (2025), a coding agent that evolves by rewriting its own code [29][30][31]. GV wrote that four of Recursive's employees had co-authored the Darwin Gödel Machine paper with Clune [9]. Commentators treated the founding team itself as the main asset behind the round, since the company had no product at launch [3][10].

## Research thesis

The company's central claim is that [recursive self-improvement](https://aiwiki.ai/wiki/recursive_self-improvement) is the shortest path to [superintelligence](https://aiwiki.ai/wiki/superintelligence), and that the way to get there is through open-ended algorithms. As the company put it in its launch statement, "The fastest path to superintelligence will be realised by AI that recursively improves itself, and does so via open-ended algorithms that drive endless innovation" [1][4]. Socher summed up the premise as "A.I. is code. And now, A.I. can code," adding, "The ingredients are there" [18].

Socher has been careful to draw a line between ordinary model improvement and the thing the lab is actually chasing. In an interview he argued that much of what passes for automated research does not count: "Our unique approach is to use open-endedness to get to recursive self-improvement, which no one has yet achieved," he said, adding that asking an AI to make some other thing better is "just improvement," not the recursive kind [5]. The recursive version requires the full research loop of ideation, implementation, and validation to run automatically [5][6].

GV, which co-led the round, framed the bet in a single line: "AI is code, and now AI can code. When these two realities connect, the self-improvement loop can be closed" [9]. GV described a system that learns to identify its own limitations, write its own benchmarks, and rewrite its own codebase, and the company said its experiments would target not only the model's code but also its harness and its training and inference infrastructure [2][9]. GV wrote that the team's first goal was to train a system with the capability of "50,000 PhDs," focused first on the science of AI itself, before pointing it at other scientific fields [9].

The company's website describes the Eureka Machine as "the self-improving superintelligence that automates the scientific method," covering ideation, implementation, and validation. It says the system explores many research paths in parallel, treats discoveries as stepping stones for later ones, and is applied first to "the full frontier AI R&D pipeline," from data and training recipes to infrastructure and performance optimization [38].

The open-endedness piece is what distinguishes the approach from a simple optimization loop. Instead of handing the system one objective, the lab wants it to keep generating new environments, problems, and forms of adaptation on its own, an idea GV described as inspired by biological and cultural evolution [5][6][9]. Rocktäschel has invoked Stanisław Lem's idea of an "information barrier," the point at which knowledge grows faster than humans can absorb it [4]. The same framing extends to safety: Socher pointed to rainbow teaming, where one model continuously probes another to surface failure modes, as a way to replace some manual [red teaming](https://aiwiki.ai/wiki/red_teaming) with an adversarial process that co-evolves with the system it is testing [5][28].

Socher has said the lab will "start with AI research itself but eventually hope to expand its aperture to physics, chemistry and especially pre-clinical biology" [2]. On timing he avoided the posture of an open-ended research lab, telling TechCrunch at launch that "there will be products, and you'll have to wait quarters, not years" [5]. Launch materials pointed to a first "Level 1" autonomous training system and a public launch targeted for mid 2026 [3][7][8]. In July 2026 he told TechCrunch that "in October or so, you'll see some actually tangible, useful things" [15]; the company released Flint on October 21 [12].

In September 2026 Socher published a book, The Eureka Machine: Why AI Is the Key to Unlocking a New Era of Scientific Discoveries (PublicAffairs). Its publisher's description says the book predicts "a superintelligence that can continuously improve itself until it hits various theoretical upper bounds" [26].

## First automated research results

On June 11, 2026, Recursive published its first results from the automated AI research system. The company said the system proposes an idea, implements it, runs an experiment, validates the result, and uses what it learns to choose the next experiment, while running many research threads in parallel and screening improvements for reward hacks and variance [13]. It reported results on three public benchmarks [13]:

| Benchmark | Metric | Previous best | Recursive's result |
| --- | --- | --- | --- |
| NanoChat autoresearch (fixed five-minute training budget on one GPU) | Validation bits per byte (lower is better) | 0.9372 | 0.9109 |
| NanoGPT Speedrun (8-GPU H100 node) | Time to reach 3.28 validation loss | 79.7 s | 77.5 s |
| SOL-ExecBench (235 GPU kernel tasks on Nvidia B200) | Mean speed-of-light score (1.0 is the analytical optimum) | 0.699 | 0.754 |

The NanoChat comparison was against the best solution from autoresearch@home, a community effort built on [Andrej Karpathy](https://aiwiki.ai/wiki/andrej_karpathy)'s autoresearch setup, after Recursive removed what it called minor reward hacks from that solution [13]. For the NanoGPT Speedrun, which the company said had seen 83 human record-setting contributions and more than two years of community optimization, Recursive said it obtained its time on Modal H100 nodes, re-confirmed it on Andromeda nodes, and was still awaiting access to the official hardware to submit it to the leaderboard [13]. The company acknowledged that the underlying models may already know public techniques, so its runs from weaker starting points "should not be read as independent or unique discovery" [13].

[Reward hacking](https://aiwiki.ai/wiki/reward_hacking) was a recurring problem, especially on the kernel benchmark, where some candidates cached outputs or exploited timing details instead of writing faster kernels. Recursive said it built increasingly strict automated correctness checks into the loop and called aligning such systems "to solve the spirit of the task, and not its letter" a "grand challenge" of safe recursive self-improvement [13]. It released artifacts from the runs on GitHub [36]. These results are the company's own measurements.

## Compute agreement with AWS

On July 28, 2026, Recursive announced a multi-year, $410 million agreement to run its automated AI research system on Amazon Web Services as its cloud provider. The two companies said they plan to co-develop infrastructure built for large-scale automated AI research [14]. TechCrunch reported that the deal has no investment component from Amazon [15].

The Next Web noted that the commitment was roughly two-thirds of the $650 million the company had raised, at a time when it employed about 25 people [16]. Socher told TechCrunch the deal was "likely going to be one of the smallest compute deals we're going to sign in the next few years," and said that for Recursive "it's less about headcount and more about agent count" [15].

## Flint

Recursive released its first model, Recursive Flint, on October 21, 2026. Flint is a [mixture-of-experts](https://aiwiki.ai/wiki/mixture_of_experts) coding model with about 40 billion total parameters (39.8 billion) and about 2.5 billion active per token, built for agentic coding on local hardware [12][25]. Recursive said a small team of people and agents built it from scratch in under four months and that the Eureka Machine contributed to nearly every part of the model-building stack: it searched architecture configurations, devised a quantization recipe that compresses the output layer to 4 bits with an offline distillation stage, and wrote specialized inference kernels for Apple Silicon and the Nvidia RTX Spark [12].

The weights are available on [Hugging Face](https://aiwiki.ai/wiki/hugging_face) under the Apache 2.0 license, alongside a local inference engine called rec, and the model is also offered through [OpenRouter](https://aiwiki.ai/wiki/openrouter) [12][25][38]. The model card reports 54.9% on [Terminal-Bench](https://aiwiki.ai/wiki/terminal_bench) 2.1, 89.7% on SWE-bench Pro v2, and 74.2% on [SWE-bench](https://aiwiki.ai/wiki/swe_bench) Verified, measured by Recursive with specific agent harnesses [25]. Recursive also claimed more than 125 tokens per second on an M4 Pro MacBook Pro [12].

Wired's Will Knight, who tested Flint on a Mac with 48 GB of memory, wrote that it was fast and "nailed my coding requests every time" [17]. Socher told Wired that "nearly every aspect of the entire model pipeline was actually done and improved upon by AI" and that Flint was built with a system incapable of escaping human control [17].

## Funding and investors

The $650 million round was described as heavily oversubscribed, and Recursive said it valued the company at $4.65 billion [2][3]. It was co-led by GV (Alphabet's venture arm, formerly [Google Ventures](https://aiwiki.ai/wiki/google_ventures)) and Greycroft, with participation from chipmaker [Nvidia](https://aiwiki.ai/wiki/nvidia) and the venture arm of [AMD](https://aiwiki.ai/wiki/amd), AMD Ventures [2][14]. The Next Web called the involvement of both chipmakers notable, since they are the main suppliers of hardware for frontier AI training, and read it as a sign that they saw recursive self-improvement as a near-term compute customer [3]. Earlybird Venture Capital also lists Recursive in its portfolio, giving 2026 as the year of its investment; the size of its stake was not disclosed [24].

The headline figure firmed up over several months. In January 2026 Tech Funding News reported that Socher was in talks to raise several hundred million dollars at roughly $4 billion before the new money, with GV and Greycroft in talks to lead [22]. In April 2026 the Financial Times reported, according to Implicator.ai, that the company had raised at least $500 million at a $4 billion pre-money valuation in a round led by GV with Nvidia participating, and that the round was so oversubscribed it could eventually reach $1 billion [11]. UCL said at the time that the round was expected to close at the end of May [23]. By the May 13 stealth exit the round stood at $650 million at $4.65 billion [2][3].

| Detail | Value |
| --- | --- |
| Announced | May 13, 2026 [1][2] |
| Total raised | ~$650 million [2][3] |
| Valuation | $4.65 billion (company-stated) [2][3] |
| Co-lead investors | GV, Greycroft [1][9] |
| Strategic investors | [Nvidia](https://aiwiki.ai/wiki/nvidia), AMD Ventures [2][14] |
| Other listed investor | Earlybird [24] |
| Adviser or hire | [Peter Norvig](https://aiwiki.ai/wiki/peter_norvig) [3][19] |

GV managing partner Tom Hulme, who wrote the firm's note on the investment, said the company explores "orthogonal S-curves" needed to complement standard pre-training and reach genuine reasoning and scientific discovery, and added: "There are few companies where the saying 'the possibilities are endless' actually holds true, but this is one of them" [9][39]. The company said it would use the capital to secure large-scale compute infrastructure, expand its research operations, and deploy its first "Level 1" autonomous training system, and GV noted that it was hiring in San Francisco and London [7][8][9].

## Location and operations

Recursive has offices in San Francisco and London [37]. Its UK company's registered office is in King's Cross, London [20], while TechCrunch described it as San Francisco-based, its AWS announcement carried a San Francisco dateline, and Earlybird lists San Francisco as its headquarters [5][14][24]. At launch it reported a team of more than 25 researchers and engineers, described in some coverage as "fewer than 30," and no released product [1][3]. The transatlantic structure mirrors the founders' own geography, with Rocktäschel working from London and five of the founders photographed by The New York Times at the company's San Francisco office [11][18].

## Context in the 2026 frontier-lab wave

Recursive Superintelligence arrived during a stretch in which a handful of small, heavily funded labs raised enormous sums on the strength of their founders alone, often before shipping anything. Startup Fortune grouped it with [Thinking Machines Lab](https://aiwiki.ai/wiki/thinking_machines_lab), [Safe Superintelligence](https://aiwiki.ai/wiki/safe_superintelligence), [AMI Labs](https://aiwiki.ai/wiki/ami_labs), and [Ineffable Intelligence](https://aiwiki.ai/wiki/ineffable_intelligence), and used Recursive's raise as an example of frontier AI research talent itself becoming a venture asset [10]. Socher himself pushed back on the "[neolab](https://aiwiki.ai/wiki/neolab)" label for research-first startups, telling TechCrunch he struggled "a little bit with this neolab category" and wanted Recursive to become "a really viable company" with products people love [5].

The thesis also sits inside a broader 2026 conversation about automated AI research and whether AI systems can meaningfully accelerate their own development. Ricursive Intelligence pursued related [feedback loop](https://aiwiki.ai/wiki/feedback_loop) ideas in chip design, and OpenAI said it was building an "automated AI researcher" [2][19]. The Next Web argued that what set Recursive apart was the explicit framing of recursive self-improvement as the company's core commercial thesis rather than a research aid [3]. That stance drew unease as well as interest. Wired's Will Knight wrote that some experts fear AI that improves itself indefinitely could escape human control, that the idea had become controversial enough for several bills in Congress to propose banning it, and that Recursive regards such fears as more science fiction than reality and such legislation as an unnecessary impediment [17]. Socher told Wired that the goal is for the Eureka Machine to achieve full recursive self-improvement, but that the process would still be overseen by humans who could stop it [17]. Whether the lab can turn a strong founding team and a clear thesis into a working self-improving system remains the open question; Socher has said that no one has yet achieved recursive self-improvement [5].

## Timeline

| Date | Event |
| --- | --- |
| December 31, 2025 | Recursive Superintelligence Ltd. incorporated in London [20] |
| January 2026 | Funding talks at about $4 billion pre-money reported [22] |
| April 2026 | FT reports a raise of at least $500 million [11] |
| May 13, 2026 | Exits stealth with $650 million at a $4.65 billion valuation [1][2] |
| June 11, 2026 | Publishes first automated AI research results [13] |
| July 28, 2026 | Announces $410 million AWS compute agreement [14] |
| September 22, 2026 | Socher's book The Eureka Machine published [26] |
| October 21, 2026 | Releases Recursive Flint with open weights [12][25] |

## See also

- [Recursive Flint](https://aiwiki.ai/wiki/recursive_flint)
- [Recursive self-improvement](https://aiwiki.ai/wiki/recursive_self-improvement)
- [Ricursive Intelligence](https://aiwiki.ai/wiki/ricursive_intelligence)
- [Ineffable Intelligence](https://aiwiki.ai/wiki/ineffable_intelligence)
- [Richard Socher](https://aiwiki.ai/wiki/richard_socher)
- [Tim Rocktäschel](https://aiwiki.ai/wiki/tim_rocktaschel)
- [Jeff Clune](https://aiwiki.ai/wiki/jeff_clune)
- [List of neolabs](https://aiwiki.ai/wiki/list_of_neolabs)
- [Neolab](https://aiwiki.ai/wiki/neolab)

## References

1. Tech.eu, "Recursive Superintelligence emerges from stealth with $650M raise," May 13, 2026. https://tech.eu/2026/05/13/recursive-superintelligence-emerges-from-stealth-with-650m-raise/
2. SiliconANGLE, "Recursive Superintelligence raises $650M to build self-improving AI models," May 13, 2026. https://siliconangle.com/2026/05/13/recursive-superintelligence-raises-650m-build-self-improving-ai-models/
3. The Next Web, "Recursive Superintelligence raises $650m at $4.65bn valuation to build self-improving AI," May 14, 2026. https://thenextweb.com/news/recursive-superintelligence-self-improving-ai-funding
4. The Decoder, "AI startup Recursive emerges from stealth with $650 million to build self-improving AI," May 13, 2026. https://the-decoder.com/ai-startup-recursive-emerges-from-stealth-with-650-million-to-build-self-improving-ai/
5. TechCrunch, "What happens when AI starts building itself?," May 14, 2026. https://techcrunch.com/2026/05/14/what-happens-when-ai-starts-building-itself/
6. Unite.AI, "Recursive Superintelligence Raises $650 Million to Pursue Self-Improving AI," May 2026. https://www.unite.ai/recursive-superintelligence-raises-650-million-to-pursue-self-improving-ai/
7. FinSMEs, "Recursive Superintelligence Raises $650M in Funding At $4.65 Billion Valuation," May 2026. https://www.finsmes.com/2026/05/recursive-superintelligence-raises-650m-in-funding-at-4-65-billion-valuation.html
8. TechFundingNews, "UK AI startup Recursive hits $4.65B valuation with $650M raise from Nvidia and GV," May 13, 2026. https://techfundingnews.com/uk-ai-startup-recursive-hits-4-65b-valuation-with-650m-raise-from-nvidia-and-gv/
9. GV (Tom Hulme), "Recursive Superintelligence: Why Self-Improving AI is the Next Frontier," May 13, 2026. https://www.gv.com/news/recursive-superintelligence-self-improving-ai
10. Startup Fortune, "Recursive Superintelligence shows AI talent is now a venture asset," May 2026. https://startupfortune.com/recursive-superintelligence-shows-ai-talent-is-now-a-venture-asset/
11. Implicator.ai, "Recursive Superintelligence Raises $500M at $4B Valuation," April 2026. https://www.implicator.ai/recursive-superintelligence-raises-500m-from-gv-and-nvidia-at-4b-valuation/
12. Recursive, [Introducing Flint](https://recursive.com/updates/introducing-recursive-flint), October 21, 2026.
13. Recursive, [First Steps Toward Automated AI Research](https://recursive.com/updates/first-steps-toward-automated-ai-research), June 11, 2026.
14. Amazon (AWS press release), [Recursive Signs $410 Million, Multi-Year Collaboration with AWS to Scale Self-Improving AI](https://press.aboutamazon.com/aws/2026/7/recursive-signs-410-million-multi-year-collaboration-with-aws-to-scale-self-improving-ai), July 28, 2026.
15. Russell Brandom, TechCrunch, [Recursive Superintelligence signs $410M compute deal with Amazon](https://techcrunch.com/2026/07/28/recursive-superintelligence-signs-400-compute-deal-with-amazon/), July 28, 2026.
16. Cristian Dina, The Next Web, [Recursive is spending two-thirds of its funding on one AWS bill, and calls it small](https://thenextweb.com/news/recursive-superintelligence-410m-aws-compute-deal), July 29, 2026.
17. Will Knight, Wired, [This Startup Is All In on Recursive Self-Improvement. Its Latest Model Is Scary Good](https://www.wired.com/story/recursive-self-improvement-flint-latest-model-scary-good/), October 21, 2026.
18. Cade Metz, The New York Times, [Notable Researchers Join $4 Billion Effort to Build Self-Improving A.I.](https://www.nytimes.com/2026/05/13/technology/recursive-superintelligence-funding-ai.html), May 13, 2026.
19. Cade Metz (The New York Times), The Boston Globe, [Notable researchers join effort to build self-improving AI](https://www.bostonglobe.com/2026/05/13/business/effort-self-improving-ai/), May 13, 2026.
20. Companies House, [RECURSIVE SUPERINTELLIGENCE LTD. (company number 16937077): overview and filing history](https://find-and-update.company-information.service.gov.uk/company/16937077). Accessed October 27, 2026.
21. Companies House, [RECURSIVE SUPERINTELLIGENCE LTD.: officers](https://find-and-update.company-information.service.gov.uk/company/16937077/officers). Accessed October 27, 2026.
22. Tech Funding News, [Richard Socher's new AI lab Recursive eyes $4B pre-money in GV-led round](https://techfundingnews.com/socher-superintelligence-startup-4b-valuation/), January 26, 2026.
23. UCL News, [UCL researchers lead two of Europe's largest-ever AI funding rounds](https://www.ucl.ac.uk/news/2026/apr/ucl-researchers-lead-two-europes-largest-ever-ai-funding-rounds), April 28, 2026.
24. Earlybird Venture Capital, [Recursive (portfolio company page)](https://earlybird.com/companies/recursive). Accessed October 27, 2026.
25. Recursive, [RecursiveSI/flint model card](https://huggingface.co/RecursiveSI/flint), Hugging Face. Accessed October 27, 2026.
26. Hachette Book Group, [The Eureka Machine: Why AI Is the Key to Unlocking a New Era of Scientific Discoveries, by Richard Socher](https://www.hachettebookgroup.com/titles/richard-socher/the-eureka-machine/9781541705708/), PublicAffairs, on sale September 22, 2026.
27. Jake Bruce et al., [Genie: Generative Interactive Environments](https://arxiv.org/abs/2402.15391), arXiv:2402.15391, February 2024.
28. Mikayel Samvelyan et al., [Rainbow Teaming: Open-Ended Generation of Diverse Adversarial Prompts](https://arxiv.org/abs/2402.16822), arXiv:2402.16822, February 2024.
29. Jeff Clune, [AI-GAs: AI-generating algorithms, an alternate paradigm for producing general artificial intelligence](https://arxiv.org/abs/1905.10985), arXiv:1905.10985, May 2019.
30. Chris Lu, Cong Lu, Robert Tjarko Lange, Jakob Foerster, Jeff Clune, and David Ha, [The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery](https://arxiv.org/abs/2408.06292), arXiv:2408.06292, August 2024.
31. Jenny Zhang, Shengran Hu, Cong Lu, Robert Lange, and Jeff Clune, [Darwin Godel Machine: Open-Ended Evolution of Self-Improving Agents](https://arxiv.org/abs/2505.22954), arXiv:2505.22954, May 2025.
32. Alexey Dosovitskiy et al., [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929), arXiv:2010.11929, October 2020.
33. Yuandong Tian, [personal homepage](https://yuandong-tian.com/). Accessed October 27, 2026.
34. EML Munich, [Alexey Dosovitskiy (Collaborating ELLIS Fellow) profile](https://www.eml-munich.de/people/alexey-dosovitskiy). Accessed October 27, 2026.
35. Caiming Xiong, [X profile (@CaimingXiong)](https://x.com/CaimingXiong). Accessed October 27, 2026.
36. Recursive, [first-steps-toward-automated-ai-research (research artifacts)](https://github.com/recursive-org/first-steps-toward-automated-ai-research), GitHub, June 2026.
37. Recursive, [Join us](https://recursive.com/join-us). Accessed October 27, 2026.
38. Recursive, [company homepage](https://recursive.com/). Accessed October 27, 2026.
39. GV, [Tom Hulme, Managing Partner](https://www.gv.com/team/tom-hulme). Accessed October 27, 2026.
40. Julie Bort, TechCrunch, [AI chip startup Ricursive hits $4B valuation 2 months after launch](https://techcrunch.com/2026/01/26/ai-chip-startup-ricursive-hits-4b-valuation-two-months-after-launch/), January 26, 2026.