François Chollet's New Startup Ndea
An AGI focused AI lab taking some novel approaches. A good or not a good idea? Ndea.
Yet another Google researcher keeps talking about AGI and now has his own startup. Keras deep learning framework, has recently launched a new startup named Ndea. This venture is dedicated to advancing the field of Artificial General Intelligence (AGI).
Are we to expect a new onslaught of AGI and ASI startups in 2025? It appears so.
The Ndea Lab will focus on merging program synthesis with deep learning to create more adaptable AI systems.
The company aims to accelerate scientific progress through AI that can innovate independently
The startup is building a distributed team of program synthesis researchers
François Chollet is starting to become a controversial figure in AGI debates, ever since OpenAI’s o1 model claimed victory on a test he devised.
AGI, ARC and Francois Chollet’s Interview with MLST
FC is also behind the nonprofit, the ARC Prize Foundation, will be led by Greg Kamradt, an ex-Salesforce engineering director and founder of the AI product studio Leverage. Kamradt will serve as president and a member of the board.
OpenAI however had access to the data the entire time. That is, OpenAI's new model, o3, has been reported to have access to the ARC (Abstract Reasoning Challenge) data, which has raised concerns about the integrity of its performance evaluations. The model scored a high accuracy of 87.5% on the ARC-AGI test, which is a significant achievement.
Let me be frank, OpenAI and The ARC AGI Prize appear to have colluded in extremely shady practices here as well as Epoch AI’s Maths benchmark here. OpenAI’s PR ethics knows no bounds.
Which makes a lot of this AGI and ASI debate, more like Tech Optimism propaganda. The is problematic because Ndea is trying to “operationalize AGI”
Ndea
The startup, Ndea, will consist of an AI research and science lab. It’s looking to “develop and operationalize” AGI. But can we even take it seriously? It’s not like it has bad ideas.
Chollet's track record speaks volumes. As the architect behind Keras, a Python deep learning library powering AI development at YouTube, Waymo, and Spotify, he has shaped how developers approach machine learning. But AGI? It’s a mouthful.
While Epoch AI has all the tell-tale signs of OpenAI’s PR manipulation, FC is supposed to be a scientist.
Ndea is building frontier AI systems that blend intuitive pattern recognition and formal reasoning into a unified architecture.
View the Mission Statement on the Website.
Ndea's founding stems from a critical view of current AI limitations. While today's language models excel at pattern matching and specific tasks, they struggle with open-ended problems and efficient learning. It’s not clear if OpenAI’s most expensive model o3 is any closer to that goal in reality. The week before Ndea was announced, Chollet and Knoop transformed their $1 million ARC Prize competition into a formal nonprofit organization. So that they are building a for-profit startup around AGI is sort of lame.
There are going to be a lot of AGI profiteers. A lot of folk have left OpenAI, and will leave Google and Anthropic.
Ndea an AGI Lab for Science?
AI for Scientific Advancement
Unlike all life before us, humanity's ascent is a story of ingenuity, not just biological evolution. Our progress has been driven by the curiosity to acquire knowledge, the ability to pass it on, and an intrinsic drive to innovate. We build technology which gives us leverage beyond our biology.
We stand at the top of a knowledge and technology colossus that we collectively created over the past ten thousand generations. Scientific progress has helped us overcome burdens that long defined human life — famines, plagues, and widespread illiteracy. Science will continue to redefine the boundaries of the human condition.
Today, the acceleration of scientific progress hinges on one factor: AI capable of independent invention and discovery. This capacity is the gateway to advancements beyond our wildest imagination.
Program Synthesis
Ndea plans to use a technique called program synthesis or “guided program synthesis”, in tandem with other technical approaches, to unlock AGI. Chollet thinks that program synthesis, which allows AI to generalize problems it hasn’t seen before from only a few examples, can help to overcome the most intractable problems in AI research.
Ndea plans to build what they call a "factory for rapid scientific advancement," focusing on both known frontiers like drug discovery and unexplored territories.
However drug discovery is not at all an unexplored domain with AI. I’ve watched many big budget labs fail going after that during the last decade or longer.
Chollet's vision for Ndea includes building AI that can continuously enhance its capabilities without encountering bottlenecks, a significant challenge in current AI development.
FC has created a false narrative around how close we are to AGI now in 2025. Good for his startup, not so good for society.
With Ndea, Ilya Sutskever’s SSI, and many other famous researchers going after and talking about AGI it’s clear this is more about profit than about real capabilities. With people like Sam Altman having so much power, we are likely to see more ponzi schemes and false narratives around this technology.
While existing deep learning systems are impressive, Chollet and Knoop argue that they are fundamentally constrained by their reliance on large datasets and their inability to adapt efficiently to new tasks. That’s not unlike what another Google spin-off is saying over in Japan, a startup Sakana AI. But will Ndea, SSI and Sakana Labs come to anything? That’s harder to pin down.
The AGI Startup Labs
Sakana AI
SSI
Ndea
Probably Mira Murati’s new startup with be of this variety (whatever it will be called).
A Path to Formal Reasoning in Machine Intelligence?
The company's technical strategy centers on program synthesis, a method that searches for discrete programs to explain observed data, potentially requiring fewer examples than traditional deep learning approaches. By combining this with deep learning techniques, Ndea hopes to create systems that can both recognize patterns and engage in formal reasoning.
Deep Learning
Current deep learning-based AI, while impressive and economically valuable, is ultimately constrained by its inability to efficiently learn and adapt. It excels at known tasks but crumbles when faced with open-ended problems. It reflects back to us only the knowledge, programs, and abstractions found in its training data. Like the man behind the curtain in The Wizard of Oz, it hides the fact that human general intelligence created the training data.
If AI cannot efficiently adapt, it is forever constrained by what humans teach it. To compress the next 100 years of scientific progress into 50, or perhaps 10, we need general intelligence, not task-specific skill. We need computers that can pose problems and explore new territory, not just apply known solutions. We need computers that can innovate.
The path to AGI is not through incremental improvements to existing methods. The problems with deep learning are fundamental and cannot be addressed superficially. It's time for a new paradigm. The good news is that we know what it is: program synthesis.
I had never heard of ‘guided program synthesis’ before the startup Ndea was announced.
Program synthesis is traditionally computing-intensive. But Chollet thinks this limitation can be overcome — and that overcoming it will help accelerate scientific progress.
According to TechCrunch, Ndea, which Chollet is co-founding with Zapier co-founder and head of AI Mike Knoop, hasn’t disclosed whether it has raised any capital from outside investors.
Ndea’s long-term vision goes beyond the creation of AGI. The lab aims to act as a “factory for rapid scientific advancement,” capable of solving both known and unknown challenges.
A new research focus for AGI
Program synthesis, the cornerstone of Ndea’s research, is still a relatively young field. Chollet likened its current state to where deep learning was in 2012, according to Venturebeat.
Chollet joins others like former OpenAI chief scientist Ilya Sutskever in pursuing alternative approaches to AI development and we can expect more to form in 2025 in a sort of AGI prophecy to coincide with datacenter expansionism and AI glorification we have seen with the rise of OpenAI and Anthropic in a U.S. push for the lead. It’s certainly U.S. venture capital advantages weaponized, relative to the rest of the world. It’s also a U.S. system of democracy and capitalism more skewed to the Billionaires than the common people.
We can’t take OpenAI’s o3 achievements with the ARC challenge seriously if they had access to the data.
The controversy puts Ndea into a potentially shady position. But in Generative AI there is no such thing as failure really, only raising more funds and even being consumed (illegal acquisitions called acquires btw) by BigTech is a kind of “exit” after all. If you are famous enough as a ML researcher (Google calls them “scientists”), there’s virtually no risk. FC is certainly famous enough, lke SSI and IS.
The AGI and ASI programming by Venture Capital is obvious in terms of ideological modification of the public - so Ndea's strategy emphasizes the use of program synthesis, which allows AI to generalize from limited examples, thus overcoming the data limitations that traditional deep learning approaches face. This method is seen as crucial for achieving AGI, as it enables the development of AI systems that can learn and adapt more efficiently than current models.
In classic Google Deepmind style PR: FC said:
“We are at the crest of a pivotal moment in scientific history and the world deserves every possible direct, unique attempt to build AGI.”
But the company is currently hiring for remote research positions, suggesting that there’s at least some financial backing.
I’m not personally one to take people like Gary Marcus or Sam Altman seriously, nor do I think the AGI narrative is factual. I understand how X and LinkedIn are owned by Tycoons and a financial elite class and even on Substack there’s a lot of fairly clickbait GPT-5 speculations. It’s difficult to know the truth in such an atmosphere where the U.S. media can no longer be relied upon.
However, François Chollet and Mike Knoop do at least bring extensive experience to Ndea. Of course it will take more than LLMs to get to AGI and we are nowhere near yet. OpenAI claims that they have AGI internally aren’t factual and their lack of transparency around benchmarks are clear evidence of tampering around this. The problem is this misinformation now is pushed with newer startups. This anthropomorphization of AI breeds a deeply dystopian culture. But is it good motivation for students in the U.S. or Asia to cheat on their homework with ChatGPT? It certainly boosts OpenAI’s revenue in that light.
Looking beyond AI development, Ndea envisions itself as a catalyst for scientific advancement. At a time when there are serious Science based LLM startups raising money, it’s the flashy AGI ones that get a lot of the attention.
A lot of the Google spinoffs have failed in Generative AI or have had trouble getting going. Too much centralization of power from OpenAI and Google spin-offs in AI funding clearly isn’t great for innovation or meritocracy in the U.S. Especially when Anthorpic is mostly owned by Amazon and OpenAI is mostly owned by Microsoft now. The Cloud is eating AI. These research labs thus have a predictable path to monopolization.
FC is well respected in the AI domain, but going after an AGI startup is somewhat deceptive, especially after the hype the ARC prize got. Can I trust that
or Ilya Sutskever have good motives? I cannot.
Nor can we take the Mission Statement of Ndea at face value. OpenAI has created a world of gimmick startups. Will they be more agile and bring the world something new? Or just eat up funds make their founders rich? People have been talking about artificial general intelligence for decades and it seems what it means and the goalposts move a lot with each decade and the corporate atmosphere of the times.
Knoop co-founded Zapier, the world’s largest AI automation company, where he led engineering and product development as well as the company’s early adoption of AI technologies. OpenAI’s o3 did not make legit progress on ARC like tasks, they cheated. And the world isn’t better for it. LLMs are not the final architecture nor is what we call compute today the final mode of machine learning thinking power.
FC is the creator of Keras. Keras went from a small library used by a few thousand enthusiasts to a state-of-the-art framework used by 2 million developers. Today, Keras powers ML solutions in countless companies and touches the lives of people everywhere.
The AGI startups think they are on to something in 2025. It won’t take long to see if they are serious.