A Neo Lab betting on "data efficiency" has surfaced with Flapping Planets, which has just raised $1 billion. The company is not only remaking new algorithms, but also restructuring the underlying infrastructure. They believe that the limitations at the bottom will also limit their algorithms at the top, much like adopting a first principles approach to take a fundamentally new path. What is the core doing: Their goal is to significantly improve the efficiency of AI (especially data efficiency and computing efficiency), as they believe that the demand for AI in the future will far exceed the supply of data and computing power. Data efficiency: Adult humans can far exceed the current model by only viewing about a few billion tokens, and they estimate that humans are 100000 to 1 million times more efficient than existing models. The goal is to make a model that can achieve high intelligence without "swallowing half the Internet". System/hardware collaboration: Breaking the traditional software hardware boundary through the co evolution of models and systems. They claim to have developed the 'most complete megakernel framework in the world', which can be completed in a single GPU kernel: complete training, inference services, playing video games, and even all three can be performed simultaneously, overlapping memory bound inference and compute bound training, making the GPU less idle; Support for 'strange architectures and optimizers'. They are also rewriting the underlying Nvidia GPU software to extract more chip performance. ⁠ The content of the CEO's interview today: Company name source: Flapping Airplanes "emphasize their desire to make a" radially different "attempt compared to traditional fixed wing aircraft, rather than optimizing along existing routes. Why bet on data efficiency: The brain can learn with very little data, and the current big model needs to "swallow half the Internet". They believe that data is the real bottleneck, and many fields (such as robotics, science, trading, DNA, etc.) are lacking in data. The Inspiration of Neuralink Experience for AI: The action potential of neurons takes about 1 millisecond, during which the computer can perform massive calculations. The brain is a 'proof of existence', indicating that there are completely different algorithmic paths, but the hardware constraints vary greatly, and the entire brain should not be directly imitated. Why rebuild GPU stack from scratch: (Their runtime framework was just released today) They didn't use Torch, NCCL, or CUDA regular runtime, but instead developed their own megakernel framework (internally called Kittens VM, etc.). A single kernel on B200 can simultaneously perform model training, inference services, and even run Doom games. This can significantly improve GPU utilization and support more "strange" architectures, optimizers, and objective functions, opening up space for new algorithms. Can small 'neo labs' win when big companies achieve RSI first If large companies first implement RSI (Recursive Self Improvement, AI using AI to accelerate their own research and development), they may accelerate on the existing Transformer/data scaling route. But Flapping Planets believes they are working on a "completely different set of problems" (especially data efficiency+new system primitives). If the paths are sufficiently different, small teams still have a chance, as the RSI of large companies accelerates their existing paradigms rather than exploring new algorithmic spaces. They also emphasized that they do not focus on "catching up with the current capabilities of big companies" in their competition, but rather bet on long-term trade-offs. What does ML research look like when no one writes code by hand anymore When the model is able to write most of the code, run experiments, and optimize on its own, the role of human researchers shifts from "writing implementations" to higher-level direction setting, designing new primitives, and determining which strange algorithms are worth exploring. Their own megakernel/Kittens framework is designed to enable "even stranger algorithms" to run - if standard PyTorch struggles to express a certain dynamic/fine-grained workload, AI itself will find it difficult to invent it. So they believe that system level innovation will become a prerequisite for automation research to truly break through. DNA、 Why are robots, trading, and other data scarce fields The current big model is very strong in high data density tasks (coding, search, web content), because the Internet has already provided these data almost free of charge. But many important areas are extremely lacking in data: robot: Physical interaction data must be truly generated (remote operation, real robot trajectories), with extremely high collection costs and publicly available datasets several orders of magnitude smaller than LLM pre training data. DNA/Scientific Discoveries: Experimental data is expensive, slow, and noisy. Trading: Financial data has costs, noise, and adversarial nature. They believe that if data efficiency can be improved by several orders of magnitude, these fields will go from "almost impossible to scale" to viable markets.