Ndea

Ndea is an artificial intelligence research company co-founded by François Chollet, the creator of the Keras deep learning library and the ARC-AGI benchmark, and Mike Knoop, a co-founder of the automation company Zapier.^[2][7] It calls itself "an AI research and science lab" and works toward artificial general intelligence (AGI) through deep learning-guided program synthesis. The idea is to search for small symbolic programs that explain observed data, using neural networks to steer the search.^[1][6] Chollet announced the company on 15 January 2025, about two months after he said he would leave Google.^[6][7] Knoop is the chief executive.^[16][17]

Ndea is a small, remote team. In October 2026 Y Combinator's company directory listed 15 people, and the company's jobs page described "a team of builders distributed across Europe and North America".^[3][9] It was part of Y Combinator's Winter 2026 batch. Menlo Ventures and Quiet Capital list it in their portfolios. Ndea has not announced how much money it has raised or at what valuation.^[9][11][14] As of October 2026 its website listed no released models or products.^[1] Chollet and Knoop also co-founded the ARC Prize Foundation, the nonprofit that runs the ARC-AGI benchmarks, and an Ndea researcher led one of the foundation's published research analyses.^[2][20]

History

Founding

Chollet said in November 2024 that he was leaving Google after close to a decade at the company.^[7] On 15 January 2025 he posted on X that he was "joining forces with @mikeknoop to start Ndea", which he called "a new AI lab", and that its focus would be "deep learning-guided program synthesis".^[6] "We're betting on a different path to build AI capable of true invention, adaptation, and innovation," he wrote.^[6] In the same series of posts he said the founders believed they had "a small but real chance of achieving a breakthrough", meaning AI "that can learn at least as efficiently as people, and that can keep improving over time with no bottlenecks in sight".^[7]

Knoop wrote that he was stepping away from his day-to-day work at Zapier, where he had been head of AI, to focus on Ndea, while staying on Zapier's board.^[7][8] He said the company was "assembling the world's top program synthesis team".^[7] At launch Ndea did not say whether it had raised money from outside investors, although it was already hiring for remote research roles.^[7][8] TechCrunch described Chollet as the latest high-profile AI researcher to leave Big Tech to found an independent lab, citing Safe Superintelligence and World Labs as earlier examples.^[7]

Y Combinator's directory and Menlo Ventures' portfolio page both give 2024 as the year Ndea was founded, which places its formation before the public announcement.^[9][11]

2025

In a 4 February 2025 episode of the Gradient Dissent podcast, where he was introduced as co-founder and CEO of Ndea, Knoop explained why the founders had started a separate lab. He said most frontier labs did not share their view of how important program synthesis would be, and that the broad industry view was still that deep learning was scaling.^[17] In his view, program synthesis would supply at least half of the ideas in a highly efficient AGI system, even if not half of the compute.^[17] He also argued that program search was what made OpenAI's o3 work.^[17] He compared o3's December 2024 ARC-AGI result to AlexNet's 2012 ImageNet win, which helped start the deep learning era, and suggested it could mark the start of a five-to-ten-year scaling period for program synthesis.^[17]

Early public material from Ndea's researchers included short videos. In April 2025 Alessandro Palmarini summarized a paper on best-first bottom-up program search by Saqib Ameen and Levi Lelis. In May 2025 Ethan Vizitei posted a demonstration of using Datalog, a declarative logic programming language, to identify musical intervals and chords.^[22][23]

At Y Combinator's AI Startup School in San Francisco on 16 June 2025, Chollet described the system Ndea was building.^[18] In his account, the system would assemble a model for each new task through discrete program search, guided by deep learning-based intuition about the structure of program space. It would draw on a global, constantly growing library of reusable abstractions, and any new building block it synthesized while solving one problem would be added back to the library, much as a software engineer publishes a useful library on GitHub.^[18] He described the goal as a meta-learner that works like a programmer, improving over time both by expanding its library and by refining its intuition. He said Ndea's first milestone would be to solve ARC-AGI with a system that starts out knowing nothing about ARC-AGI.^[18]

In August 2025 the ARC Prize Foundation published an analysis of the Hierarchical Reasoning Model (HRM), a 27-million-parameter model from the Singapore-based lab Sapient whose paper reported 41% on ARC-AGI-1. The foundation said Ndea researcher Konstantin Schürholt led the analysis.^[20] It roughly reproduced the claimed results, measuring 32% on the ARC-AGI-1 semi-private set and 2% on the ARC-AGI-2 semi-private set. Its ablations found that the "hierarchical" architecture made little difference compared with a similarly sized transformer, while a little-documented outer refinement loop drove much of the performance.^[20] Schürholt presented the work in a talk released on Ndea's YouTube channel.^[21]

2026

Ndea took part in Y Combinator's Winter 2026 batch. Its YC directory entry carries the one-line description "Building AGI that can innovate." and names Tyler Bosmeny as the company's primary YC partner.^[9]

In an episode of Y Combinator's Lightcone podcast released on 27 March 2026, shortly after the launch of ARC-AGI-3, Chollet said Ndea was not building an alternative to coding agents. Instead it was trying to build "a new branch of machine learning, an alternative to deep learning itself".^[19] He described the approach as replacing the parametric curve that deep learning fits with gradient descent by "a symbolic model that is meant to be as small as possible". Those models would be found by a process the company calls "symbolic descent", which he described as the symbolic-space equivalent of gradient descent.^[19] He said the team had started on day one with this "symbolic learning vision", tried many ideas, and taken "half a year roughly" to reach good foundations on which it could start "building a system that compounds".^[19] He put Ndea's chance of success at "maybe a 10 or 15%" and said the attempt was still worth making because no one else was likely to make it.^[19]

Name and mission

The company describes its name as "like 'idea' with an 'n'". It says the name was inspired by the Greek concepts ennoia (intuitive understanding) and dianoia (logical reasoning), and reflects its first goal of merging deep learning with program synthesis.^[1]

Ndea says it is "entirely focused on developing and operationalizing AGI to realize unprecedented scientific progress in our lifetime". Its stated longer-term aim is "a factory for rapid scientific advancement", one "capable of inventing and commercializing N ideas".^[1] Its website lists self-driving vehicles, drug discovery, sustainable energy, robotics and space exploration as "known" frontiers that AGI would help with, and says the company's "chance of success is low, but not zero."^[1] It also argues that "the world deserves every possible direct, unique attempt to build AGI".^[1]

Research approach

Ndea's website argues that current deep learning "excels at known tasks but crumbles when faced with open-ended problems". The reason it gives is that deep learning cannot efficiently learn and adapt, and the company calls for a new paradigm: program synthesis.^[1] Deep learning interpolates between data points in a continuous embedding space. Program synthesis instead searches for discrete programs that "perfectly explain observed data", which the company says gives much stronger generalization from only a few examples.^[1] The main obstacle is the cost of searching a combinatorial space of programs. Ndea's bet is that deep learning models can guide that search and remove the bottleneck.^[1][19] Chollet has compared this to the way AlphaGo and AlphaZero use learned intuition to guide search.^[19]

The website also says that "the state of program synthesis in 2025 looks a lot like deep learning circa 2012". It notes that every frontier AI lab is starting to explore program synthesis but that most treat it as a small part of what AGI needs, whereas Ndea considers program synthesis and deep learning "equally important".^[1]

On its jobs page Ndea describes the work as building "a new branch of machine learning: Symbolic Learning, a method for efficiently searching for minimal symbolic models in program space".^[3] Chollet has tied this focus on minimal models to the minimum description length principle, under which the shortest model of the data is the one most likely to generalize. He argues that such models would need much less data, run more cheaply at inference time, and generalize and compose better than parametric models.^[19] Job postings describe a hybrid architecture. A deep learning engineer would build the "System 1" side, described as "fast, intuitive systems that complement symbolic reasoning".^[4] A platform engineer would work "at the frontier of neuro-symbolic AI", with program synthesis, inductive logic programming, Answer Set Programming, Datalog, and SAT/SMT solving listed as desirable experience. The same posting lists "Genuine interest in symbolic AI" among its qualifications.^[5]

People

Ndea's public materials also name researchers Alessandro Palmarini, Ethan Vizitei and Konstantin Schürholt, along with Bryan Landers, a member of technical staff who hosts the company's podcast.^{[21][22][23][24]} Landers co-wrote the ARC Prize 2024 technical report with Chollet, Knoop and Greg Kamradt.^[30]

Funding

Ndea has not announced a funding round or disclosed a valuation. When the company launched, TechCrunch reported that it had not said whether it had raised outside capital.^[7] The investors below are those that list Ndea on their own websites, plus two that appear only in a third-party database.

In April 2026 Specter listed Ndea's total funding as $43.5 million.^[15] Ndea has not published a total, and the investors' portfolio pages do not give amounts.^[11][14]

Relationship with the ARC Prize Foundation

Chollet and Knoop are co-founders and board members of the ARC Prize Foundation, which Ndea's about page describes as "a non-profit for public advancement and education of open AGI".^[2] The foundation runs the ARC-AGI benchmarks and the annual ARC Prize competitions.^[20] Ndea is a separate company, but the two are linked in practice. Chollet has named solving ARC-AGI as Ndea's first research milestone, Ndea researcher Konstantin Schürholt led the foundation's 2025 analysis of HRM, and Bryan Landers, now on Ndea's technical staff, co-wrote the foundation's 2024 technical report, which was published in December 2024, before Ndea's public launch.^{[18][20][24][30]}

Abstract Synthesis podcast

In December 2025 Ndea began publishing Abstract Synthesis, a podcast on its YouTube channel. Most episodes interview a program synthesis or formal methods researcher about a specific paper; the others include a live multi-guest episode and a recap. Bryan Landers of Ndea hosts it.^[24][25] Guests have included:

• Mark Santolucito of Barnard College, Columbia University, on grammar filtering for syntax-guided synthesis (December 2025)^[24]
• Kevin Ellis of Cornell University, on DreamCoder, a system that learns reusable program abstractions together with a neural search strategy^[26]
• Stephen Muggleton of Imperial College London, on his paper that introduced and named inductive logic programming^[27]

Other guests include Kedar Namjoshi, Moshe Vardi, Andrew Cropper, Aws Albarghouthi and Loris D'Antoni.^[32] A live episode was recorded at NeurIPS 2025 in San Diego.^[28] A recap episode in February 2026 summarized the first eight conversations under three themes: shrinking the search space without breaking correctness, why "correct" programs still behave badly, and the real meaning of "neurosymbolic".^[25]

Reception

On the day of the launch, Naveen Rao, the former chief executive of MosaicML, wrote that he was "excited to be involved as a supporter". He said the field was "dominated by hype and hope" and that Chollet and Knoop were "going after fundamental innovations".^[29] Debarghya Das's "The Ultimate Neolab List" includes Ndea and classifies it as a frontier lab working on program synthesis.^[31]

See also

• François Chollet
• Mike Knoop
• ARC-AGI
• ARC-AGI 3
• Program synthesis
• Symbolic AI
• List of neolabs
• Neolab

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Ndea. "Ndea" (company home page). Retrieved 27 October 2026.
2. ^1 ^2 ^3 ^4 ^5 ^6Ndea. "About Ndea". Retrieved 27 October 2026.
3. ^1 ^2 ^3 ^4Ndea. "Join Ndea". Retrieved 27 October 2026.
4. ^1 ^2Ndea. "Applied Deep Learning Engineer". Retrieved 27 October 2026.
5. ^Ndea. "Applied Research Platform Engineer". Retrieved 27 October 2026.
6. ^1 ^2 ^3 ^4 ^5François Chollet. "I'm joining forces with @mikeknoop to start Ndea". X, 15 January 2025.
7. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Kyle Wiggers. "AI researcher François Chollet founds a new AI lab focused on AGI". TechCrunch, 15 January 2025.
8. ^1 ^2Mike Wheatley. "François Chollet reveals he's leading a new startup that aims to make AGI a reality". SiliconANGLE, 15 January 2025.
9. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Y Combinator. "Ndea: Building AGI that can innovate.". YC company directory. Retrieved 27 October 2026.
10. ^Kirsty Nathoo. "The Y Combinator Standard Deal". Y Combinator. Retrieved 27 October 2026.
11. ^1 ^2 ^3 ^4 ^5 ^6Menlo Ventures. "Ndea". Menlo Ventures portfolio. Retrieved 27 October 2026.
12. ^Deedy Das. "Q3 2025 Anthology Fund Update". Menlo Ventures, 24 October 2025.
13. ^Menlo Ventures. "Menlo Anthology Fund". Retrieved 27 October 2026.
14. ^1 ^2 ^3 ^4Quiet Capital. "Ndea". Quiet Capital portfolio. Retrieved 27 October 2026.
15. ^1 ^2Specter. "Hottest Startups - March 2026". Specter Insights, 24 April 2026.
16. ^1 ^2 ^3Mike Knoop. "Mike Knoop" (personal home page). Retrieved 27 October 2026.
17. ^1 ^2 ^3 ^4 ^5 ^6 ^7Gradient Dissent. "R1, OpenAI's o3, and the ARC-AGI Benchmark: Insights from Mike Knoop on the Gradient Dissent Podcast". YouTube, 4 February 2025.
18. ^1 ^2 ^3 ^4Y Combinator. "François Chollet: How We Get To AGI". YouTube, 3 July 2025 (talk given 16 June 2025 at AI Startup School).
19. ^1 ^2 ^3 ^4 ^5 ^6 ^7Y Combinator. "François Chollet: Why Scaling Alone Isn't Enough for AGI". Lightcone podcast, YouTube, 27 March 2026.
20. ^1 ^2 ^3 ^4 ^5ARC Prize Team. "The Hidden Drivers of HRM's Performance on ARC-AGI". ARC Prize, 15 August 2025.
21. ^1 ^2Ndea. "The Surprising Performance Drivers of HRM". YouTube, 18 August 2025.
22. ^1 ^2Ndea. "Paper Recap: Program Synthesis with Best-First Bottom-Up Search". YouTube, 29 April 2025.
23. ^1 ^2Ndea. "Program synthesis + music theory - Datalog demo". YouTube, 19 May 2025.
24. ^1 ^2 ^3 ^4Ndea. "Grammar Filtering For Syntax-Guided Synthesis - Mark Santolucito". Abstract Synthesis, YouTube, 16 December 2025.
25. ^1 ^2Ndea. "February 2026 Podcast Recap". Abstract Synthesis, YouTube, 9 February 2026.
26. ^Ndea. "DreamCoder's Wake-Sleep Library Learning - Kevin Ellis". Abstract Synthesis, YouTube, 7 April 2026.
27. ^Ndea. "Inventing Inductive Logic Programming - Stephen Muggleton". Abstract Synthesis, YouTube, 18 June 2026.
28. ^Ndea. "Live @ NeurIPS 2025". Abstract Synthesis, YouTube, 29 December 2025.
29. ^Naveen Rao. Post on X about Ndea's launch. X, 15 January 2025.
30. ^1 ^2François Chollet, Mike Knoop, Gregory Kamradt and Bryan Landers. "ARC Prize 2024: Technical Report". arXiv:2412.04604, 5 December 2024.
31. ^Debarghya Das. "The Ultimate Neolab List". Retrieved 27 October 2026.
32. ^Ndea. "Ndea" YouTube channel, videos. Retrieved 27 October 2026.