Is the Recursion Point about to arrive?

The recursion point in AI research refers to the stage where AI systems become capable of autonomously improving themselves, leading to an accelerating cycle of recursive self-improvement. A system reaches the recursion point when it can not only perform tasks effectively but also improve its own architecture, algorithms, and reasoning processes.This could involve refining its neural networks, optimizing its code, or even designing entirely new AI models. Once an AI can improve itself, each improvement could make it better at self-improvement, potentially leading to an exponential rate of advancement. This scenario is sometimes referred to as an "intelligence explosion" (coined by I.J. Good in 1965)"

To illustrate a point towards recursion, all the explanation above was written by Chat GPT. Progress towards the recursion point where AI builds AI leading to an intelligence explosion is almost certainly coming soon. But the exciting news is not only, getting AI to build AI that becomes truly recursive, (one of main focuses of all the major AI research labs from Google Deep Mind to Meta Fair, Deepseek to Alibaba, SSI to Open AI) but that a UK startup established by two PhD Graduates from UCL is enabling them to get there.

The startup named WECO established in London less than 2 years ago built and open sourced in early 2024 an AI powered Agent for machine learning engineering called AIDE. It got some notice amongst data science experts but not much fanfare until Open AI published in the summer of 2024 that their own Machine Learning Engineering benchmark evaluations. This showed AIDE when working on top of GPT 4 delivers up to 4× more medals than any other agents (including the next best from Stanford MLAB) across 75 Kaggle Competitions. Then In November 2024 METR ( Model Evaluation & Threat Research) the well respected nonprofit organization set up by researchers from Deep Mind and Open AI ( in order to be dedicated to assessing the capabilities and potential risks of advanced AI systems) found the AIDE agent outperformed data scientists from the top AI labs for up to a 6 hour time frame.

AIDE heaviest users are now researchers in those top AI labs and so this week the founders are in Silicon Valley to kick off their £10m seed round and unlike most European startups they are nor raising on the back of their commercial traction or the strength of their waitlist (which is substantial) but on the strong prospect that with additional resources, more top class researchers to supplement their top notch research team and more compute power they can reach the recursion point - build AI that can build SOTA AI- in the next 12 months.

This they will prove by first getting AIDE to improve AIDE and then by using the new AIDE to win a high profile Kaggle competitions and then get AIDE to AUTO build a SOTA AI system that can beat the best data scientists on the standard benchmarks across all time frames. Weco then plan to follow the well trodden path of most of the top AI research Labs and release the AIDE system first to the community and then to enterprises so we can all see what AI built by AI can build, what it can generate and the novelties it can discover.

The Recursion point is coming. We at Twin Path Ventures, alongside some of the world best Dara Scientist researchers backed their pre-seed round, and now VC's, hopefully not just in the USA, can get an opportunity to back this outstanding team to be the first to the recursion point.

Check them out at https://www.weco.ai