Hiverge

Monday was a blast! An evening dedicated to AI for discovery, packed with inspiring talks and insightful discussions. Thank you to our great speakers Tian Xie Konstantinos Meichanetzidis and all the participants. Let’s keep pushing the frontier of science with AI. Stay tuned for upcoming events...

4 days until our AI for Discovery meetup in Cambridge! From discovering new materials to advancing quantum computation, AI is pushing the frontier of science. We are hosting an evening of talks and discussions on AI for scientific discovery featuring: - Konstantinos Meichanetzidis, Head of Scientific Product Development and AI strategy Quantinuum - Tian Xie Project Lead Microsoft Research, AI for Science - Johannes Bausch, Staff Research Scientist Google DeepMind Bradfield Centre, Cambridge | Monday 24th November | 5:30pm - 9pm A few seats remaining - grab your spot here: https://luma.com/lcw9qnaj

We often talk about improving computing efficiency by optimizing hardware — faster GPUs, denser racks. But efficiency isn’t just about hardware. It’s about where and how computation happens, across three layers: 1. The compute layer: the software itself that performs the computation. 2. The orchestration layer: how computation is scheduled and distributed. 3. The physical layer: the hardware and data centers where it runs. For decades, developers could rely on Moore’s Law to make software faster without changing a line of code. That era is over. Between 2010 and 2020, single-threaded CPU performance grew only about 2–3x. Compare that to ~30x from 1990 to 2000. The orchestration layer, in contrast, has seen remarkable progress thanks to large open-source systems like Kubernetes. But the compute layer; i.e., the software itself remains underexplored. As Wirth’s Law famously put it: “Software is getting slower more rapidly than hardware becomes faster.” Performance engineers who optimize code at this level have a difficult job. They need to dive deep, find bottlenecks, and implement optimizations manually. Tools like profilers and compilers help, but they can’t make higher-level algorithmic decisions — they won’t replace a Bubble Sort with a Quick Sort or automatically vectorize a complex loop. Yet this is precisely where the largest performance gains lie. Rewriting or restructuring algorithms can lead to orders of magnitude improvements; sometimes tens of thousands of times faster, as seen in optimized matrix multiplication through techniques like vectorization and parallel divide-and-conquer. And it’s not only about speed. As data centers approach near-perfect Power Usage Effectiveness (PUE ≈ 1), the software layer increasingly dominates total energy consumption. To continue improving efficiency, we must learn to automate software optimization at scale. This — the automation of compute-layer efficiency — may well define the next frontier of computation. It’s not just an enabler of faster systems; it’s a prerequisite for sustainable and intelligent computing in the decades ahead.

From discovering new materials to advancing quantum computation, AI is pushing the frontier of science. Join us for an evening in Cambridge, focused on AI for scientific discovery, featuring speakers from Quantinuum, Microsoft Research and Google DeepMind. Bradfield Centre, Cambridge | Monday 24th November | 5:30pm - 9pm Register fast - space is limited: https://luma.com/lcw9qnaj

DeepSeek and Moonshot's Kimi K2 proved that with smarter algorithms you could outperform the world's biggest AI labs with only a fraction of the investments. Where is the catch? Finding such novel ideas requires world-class AI talent working relentlessly. Still, not everyone can afford that. At Hiverge, we strive to enable anyone to automatically discover breakthrough algorithms. Meet Abdulkarim Alsayed Ramadan, our intern who had zero machine learning experience. And for his internship, we asked him something “simple”: Use the Hive, our algorithmic discovery engine, to break the CIFAR-10 speedrun record, a benchmark for which AI experts have competed for years training computer vision models faster than ever. In just weeks, Abdulkarim broke the world record by 20% breaching the 2-second threshold. Now is the time for algorithmic creativity. And we're on it. Read the full story in our blog: https://lnkd.in/ebss8saH

Our team was present at the Airbus Aircraft Saint-Martin site in Toulouse to receive the 1st prize for the Beluga™ AI scalability challenge, which focused on a real-world planning problem from Airbus operations. When dealing with massive, complex operations such as plane production, planning the most efficient set of actions becomes an exponentially hard problem. Our platform, the Hive, discovered algorithms for planning that outperformed classical methods and other competitors by a large margin while being 1000x faster! Bernardino Romera Paredes, our CTO, had the opportunity to present our results and methods to Airbus decision makers. We are grateful to the Airbus teams for a great welcome. Stay tuned for a detailed blog post on our method soon...

We are #hiring! Hiverge is looking for talented engineers and researchers to join our on-site team in Cambridge. Apply today: https://lnkd.in/dHGXK48W #ArtificialIntelligence #Algorithms #Engineering

Excited to see our journey featured in Business Insider. Thanks Robert Scammell for the great summary of our vision to bring world-class algorithms to every company, democratizing access.