Algorithm discovery
3 labs · 6 models and products Theme: Companies · Generated from this page's data. Click a name to open its panel on the Labs page. Records resting on one source are tagged "one source".An algorithm is a step-by-step method a computer follows, for example to sort data or multiply tables of numbers. Algorithm discovery uses AI to search for new or faster methods and checks automatically that they work. Google DeepMind's systems on this tile include AlphaTensor and AlphaDev, which found faster routines, FunSearch, which found new maths results, and AlphaEvolve, a Gemini-powered coding agent (per the site). Hiverge, based in Cambridge, runs many AI research agents on its platform, the Hive, to improve customers' algorithms; it reports work on an Airbus planning problem.
Context
- AlphaDev, from DeepMind, found sorting routines that were added to a standard library used by C++ programmers. 2023 (Nature vol. 618) nature.com/articles/s41586-023-06004-9
These algorithms have been integrated into the LLVM standard C++ sort library
- FunSearch, from Google DeepMind, used a language model plus an automatic checker to find new results on the "cap set" maths problem. 14 December 2023 (online) nature.com/articles/s41586-023-06924-6.pdf
we discover new constructions of large cap sets going beyond the best-known ones
- AlphaEvolve found a way to multiply two 4×4 complex-valued matrices with 48 scalar multiplications, its authors report. 16 June 2025 arxiv.org/abs/2506.13131
AlphaEvolve developed a search algorithm that found a procedure to multiply two 4×4 complex-valued matrices using 48...
- Hiverge, based in Cambridge, reports its Hive platform delivered solutions up to 1,000 times faster on average for an Airbus planning problem. 17 September 2025 hiverge.ai/blog/introducing-hiverge-our-mission-
delivered solutions up to 1,000 times faster on average than existing approaches.
- Hiverge describes itself as building an AI discovery engine that writes code and algorithms. 17 September 2025 hiverge.ai/blog/introducing-hiverge-our-mission-
Hiverge is building algorithmic superintelligence: an AI discovery engine that writes breakthrough code and algorithms.
Key papers
- Faster sorting algorithms discovered using deep reinforcement learning (AlphaDev) Nature · 2023
- Mathematical discoveries from program search with large language models (FunSearch) Nature · 14 December 2023 (online)
- AlphaEvolve: A coding agent for scientific and algorithmic discovery arXiv preprint · 16 June 2025
Labs Round 8 (31 Oct 2026): introduction and facts as the researchers wrote them; "(per the site)" marks a sentence resting on this site's own description of that lab.
| Record | What it is | Place | Links | Confidence |
|---|---|---|---|---|
| AlphaChipChips, models and products | A reinforcement-learning method that lays out computer chips, used for Google's TPU designs. released (open checkpoint and weights)Key facts Theme: Algorithm discovery. Method first in a 2020 preprint and a 2021 Nature paper; named AlphaChip with a Nature addendum and released pre-trained checkpoint on 26 Sep 2024. Figures (26 Sep 2024): used in the last three generations of Google's TPU; layouts "in hours, rather than taking weeks or months"; MediaTek extended it for its chips.Made by Google DeepMind · 2024-09-26 | London (EC2M 4AG) (via Google DeepMind) | deepmind.google/blog/how-alphachip-transformed-c · deepmind.google/blog/page/8/ | one source |
| AlphaDevChips, models and products | An AI system that found faster sorting and hashing routines by building programs one low-level instruction at a time. released (paper; algorithms open-sourced into LLVM libc++)Key facts Theme: Algorithm discovery. Nature paper 7 Jun 2023; new sorting algorithms added to the LLVM libc++ standard library. Figures (7 Jun 2023): sorting up to 70% faster for short sequences, about 1.7% faster above 250,000 elements; hashing 30% faster in the 9–16 byte range; estimated use "trillions of times a day".Made by Google DeepMind · 2023-06-07Read the paper | London (EC2M 4AG) (via Google DeepMind) | deepmind.google/blog/alphadev-discovers-faster-s · deepmind.google/blog/page/10/ | one source |
| AlphaEvolveChips, models and products | Google DeepMind's coding agent that uses Gemini models plus automatic checkers to discover and improve algorithms. released (early access for academics)Key facts Recovers about 0.7% of Google's worldwide compute via a Borg scheduling heuristic; 4x4 complex matrix multiplication in 48 scalar multiplications; 23% faster Gemini kernel, 1% less Gemini training time; tried on 50+ open maths problems, matched best known in ~75%, improved in ~20% (all 14 May 2025). UK scientists offered priority access under the Dec 2025 UK government partnershipMade by Google DeepMind · 2025-05-14Read the paper | London | deepmind.google/discover/blog/alphaevolve-a-gemi · deepmind.google/blog/strengthening-our-partnersh | lab's own pages |
| AlphaTensorChips, models and products | An AI system that searches for faster ways to do matrix multiplication and checks they are correct. released (paper and code)Key facts Theme: Algorithm discovery. Nature paper 5 Oct 2022; GitHub repository linked. Figures (5 Oct 2022): 4x5 by 5x5 matrices in 76 multiplications (human best 80); found algorithms 10–20% faster on Nvidia V100 and TPU v2. Built on AlphaZero.Made by Google DeepMind · 2022-10-05Read the paper | London (EC2M 4AG) (via Google DeepMind) | deepmind.google/blog/discovering-novel-algorithm · deepmind.google/blog/page/10/ | one source |
| FunSearchChips, models and products | A system in which a language model writes programs and an automatic checker scores them, used to find new maths results. released (paper)Key facts Theme: Algorithm discovery. Nature paper 14 Dec 2023 (doi 10.1038/s41586-023-06924-6). Results as published: cap set problem — "the largest increase in the size of cap sets in the past 20 years"; bin packing with fewer bins. Later work in the same line: AlphaEvolve (May 2025).Made by Google DeepMind · 2023-12-14Read the paper | London (EC2M 4AG) (via Google DeepMind) | deepmind.google/blog/funsearch-making-new-discov · deepmind.google/blog/page/9/ | one source |
| HivergeCompanies | Builds a platform that runs thousands of AI research agents to discover and improve computer algorithms.Profile
Labs Round 7 deep profile (30 Oct 2026), as the researchers wrote it; confidence: verified. Sources Models and products (1)
| 20 Station Road, Cambridge (CB1 2JD) | Companies House · hiverge.ai/ · blogs.nvidia.com/blog/uk-ecosystem-deployment/ · tech.eu/2026/10/15/nvidia-bets-on-five-uk-startu · find-and-update.company-information.service.gov. | verified |
| InherentCompanies | London research lab building Faraday, an AI system meant to let scientists and self-improving AI work together on research.Profile
Labs Round 7 deep profile (30 Oct 2026), as the researchers wrote it; confidence: verified (founders, roles, round: investor page plus trade); Faraday model details single_source. Sources Models and products (1)
Key facts and news (2)
Labs Round 8 (31 Oct 2026), from the researchers' rows; one-source items are marked. | London | Companies House · tech.eu/2026/05/29/london-based-ai-lab-inherent- · uktech.news/ai/london-based-ai-lab-inherent-emer · inherentlabs.ai/ · find-and-update.company-information.service.gov. | verified |
| Recursive SuperintelligenceCompanies | Research company building AI systems that improve themselves, starting by automating AI research and the scientific method.Profile
Labs Round 7 deep profile (30 Oct 2026), as the researchers wrote it; confidence: single_source (two trade sources, no primary); sources are listed in the record panel. Models and products (1)
Key facts and news (5)
Labs Round 8 (31 Oct 2026), from the researchers' rows; one-source items are marked. | King's Cross, London (N1 9FN) | Companies House · tech.eu/2026/05/13/recursive-superintelligence-e · find-and-update.company-information.service.gov. · find-and-update.company-information.service.gov. · recursive.com/ | one source |
| The Hive (Hiverge algorithm-discovery platform)Chips, models and products | Hiverge's platform that runs many AI research agents to write, test and improve algorithms against a customer's own metrics. early accessKey facts Early results (17 Sep 2025): over 20% cut in CIFAR-10 training time; about 1.25% gain on small GPT-2 training; Airbus Beluga planning problems solved up to 1,000 times faster on average. CIFAR-10 training record 1.99 s vs previous 2.59 s at 94% accuracy on one NVIDIA A100 (16 Oct 2025). Logos shown: Airbus, AWS, NVIDIA, QuantinuumMade by Hiverge · 2025-09-17 | Cambridge | hiverge.ai/blog/introducing-hiverge-our-mission- · hiverge.ai/blog/accelerating-ai-training-made-si | one source |