The world runs on algorithms. They don't just dictate search results or social media feeds—they are the lifeblood of modern industry. From orchestrating supply chains to training and running AI models, algorithms power the backbone of business today. But current algorithms fall short of today's challenges. Systems are scaling exponentially while computational resources remain constrained, and problems across industries are more diverse than ever. With algorithmic expertise increasingly scarce compared to growing demand, this gap means we're approaching the limits of what's possible with traditional approaches, curtailing growth and putting efficiency gains out of reach. What's needed is a way to discover breakthrough algorithms—highly tailored to each specific challenge—whilst democratising access.
Hiverge is building algorithmic superintelligence: an AI discovery engine that writes breakthrough code and algorithms. Our system continuously optimises code uniquely tailored to each user's specific challenges, based on concrete metrics. From resource allocation to machine learning acceleration, we're building technology that discovers algorithms capable of transforming entire operational paradigms. Unlike traditional software engineering, which focuses on building applications and systems, our technology specialises in algorithmic discovery—finding novel computational approaches to solve complex problems.
Where are we today?
Over the past few months, we have built our core platform for algorithmic discovery, the Hive. Behind the Hive lies state-of-the-art program synthesis technology that combines language models with advanced search algorithms. This platform is rooted in rigorous scientific methodology, building upon our deep expertise in program synthesis research (e.g., AlphaTensor, FunSearch and the early stages of AlphaEvolve). Below, we offer a sneak peek into the first real-world applications of the Hive, with more in-depth posts coming in the next few weeks.
Accelerating AI algorithms
AI models are notoriously resource-intensive, consuming vast amounts of computational power and time for training and inference. Making AI models train and run faster is now of paramount importance—not just for cost efficiency, but also for enabling real-time applications and reducing environmental impact. However, achieving this is highly challenging and demands deep technical expertise.
To test the Hive’s ability to make novel contributions to this area, we evaluated it on the task of accelerating the training of a neural network for the popular CIFAR-10 image dataset. The Hive discovered novel optimisations that reduced training time by more than 20% compared to existing best-known human-engineered training algorithms. Figure 1 summarises the algorithmic ideas implemented by the Hive, and the relative contribution of each in speeding up the training. The Hive implemented a diverse range of algorithmic ideas, many of which were surprising and novel to us; including a smart selective test time augmentation approach, vectorisation of the optimiser and data augmentation procedure, as well as modifications to the model architecture.
Figure 1: CIFAR-10 training speedup obtained by the Hive
We also ran the Hive to optimise the training of a language model — a smaller version of the GPT-2 model. The Hive discovered new ideas to accelerate training, leading to a ~1.25% reduction of training time relative to the previous state-of-the-art. This improvement comes on top of cumulative efforts by top AI engineers to optimise the training code over the past year. This result marks an important milestone, especially in light of prior works highlighting that current LLMs notably struggle in this task.
Beyond these benchmarks, we found that the Hive’s ideas are generic enough to be applicable across multiple settings. In fact, when we analysed the improvements obtained by the Hive, we found that some ideas (e.g., the vectorisation of the optimiser) were common across both CIFAR-10 and GPT-2 training tasks. We are excited about the opportunity of applying these ideas to a wide variety of models.
Algorithms for sequential planning
Another use-case of the Hive that we have been working on is sequential planning. Planning problems are ubiquitous across a wide range of areas, including robotics, aerospace and supply chain management. Long-horizon planning is notoriously difficult with current approaches either based on advanced classical search methods or reinforcement learning. Indeed, classical algorithms only work for small-medium range problems, while reinforcement learning-based models are difficult and computationally intensive to train. At the same time, while LLMs excel at many tasks, they notoriously struggle with complex multi-step planning.
We have tested the capabilities of the Hive on a challenging real-world planning problem from Airbus operations. The problem involves finding the optimal way of moving jigs from Beluga airplanes to production lines in a coordinated approach, while satisfying several complex constraints. This is a difficult problem where traditional state-of-the-art planning algorithms quickly hit scaling limits and LLMs fail to progress. In contrast, the Hive discovered a planning strategy that not only solved almost all problems in our testing problem set, but also delivered solutions up to 1,000 times faster on average than existing approaches—a performance that earned us the first place in the Airbus Beluga scalability challenge. Figure 2 shows the performance of the planning algorithm discovered by the Hive compared with other state-of-the-art approaches.
Figure 2: The Hive for sequential planning
What makes these results truly remarkable is that we didn’t provide the Hive with any strategic hints, pre-built frameworks, or domain expertise. Starting from nothing more than a barebones function header, the Hive independently proposed an intricate solution spanning over 1,000 lines of code. The discovered algorithm incorporates novel forward-thinking strategies that consider the long-term impact of current actions, along with intelligent safeguards to avoid dead-end states—ideas which are all specifically tailored towards the Airbus problem domain.
What’s next?
We’re just getting started in executing our mission of building algorithmic superintelligence. With the Hive, we aim to empower engineers everywhere to tackle the most complex algorithmic challenges. We're giving early access to the Hive to selected customers. If you are interested in trying out the Hive, please fill the for us to understand your use-case, and we will get back to you.
We're building our vision alongside an exceptional team of researchers and engineers, backed by leading investors and industry experts. We are excited to announce that we have raised a $5M Seed led by Flying Fish Ventures, with participation from Ahren Innnovation Capital, 10x Founders, Alpha Intelligence Capital, 1st Kind, StemAI, Vela Partners, Cur8 Capital and Jeff Dean.
We're actively building our team. If you're passionate about algorithmic discovery and want to shape the future of AI, apply here!
