Ineffable Intelligence Ltd., a British artificial intelligence startup founded a few months ago, has raised $1.1 billion in seed funding.

The investment values the company at $5.1 billion. CNBC reported today that Lightspeed Ventures and Sequoia Capital led the deal. They were joined by Nvidia Corp., Google LLC, the U.K’s Sovereign AI Fund, DST Global, Index Ventures and others.

Ineffable Intelligence is led by prominent AI researcher David Silver. He previously spent more than a decade at DeepMind, Alphabet Inc.’s machine learning unit. He was the lead developer of the groundbreaking AlphaGo model that the lab debuted in 2016.

AlphaGo bested the world’s highest-ranked Go player in a series of matches that were watched by more than 200 million people. DeepMind later used the technologies that underpinned the model to build a math-optimized AI called AlphaProof. In 2024, AlphaProof became the first neural network to win a medal at the International Mathematical Olympiad.

AI models answer user questions based on existing knowledge they collect from sources such as the public web. According to Wired, Ineffable Intelligence is seeking to build an AI model that can obtain entirely new knowledge. The startup believes that such an algorithm, which it refers to as a superlearner, could accelerate scientific research and engineering projects. 

Ineffable Intelligence plans to develop the superlearner using reinforcement learning. It’s a common AI training method that lends itself to, among other tasks, building large language models.

The core concept behind reinforcement learning is to provide an AI with sample tasks and observe how well it completes them. That monitoring is usually carried by a second, less advanced neural network called a teacher model. It provides the AI being trained with feedback that it uses to refine its output.

There are many implementations of reinforcement learning. One common approach is to equip an AI model with a database of common situations that it will encounter in production. That approach, which is known as model-based reinforcement learning, speeds up AI training.

Other reinforcement learning techniques don’t use a database and instead prioritize model adaptability. Another method, RLHF, provides the teacher model that guides the AI training process with feedback from humans. The human feedback can significantly boost AI output quality.  

Reinforcement learning is usually applied to models that have already been calibrated through a process known as pre-training. Ineffable Intelligence plans to skip that step. Additionally, it will place its AI models in simulations that will enable them to learn from one another. 

AlphaGo used a similar approach. It played millions of Go matches against itself to develop novel tactics that weren’t available in training datasets.

Ineffable Intelligence’s raise comes two months after AMI Labs Inc., another early-stage AI startup with a prominent founder, secured $1.03 billion in funding. Chief Executive Yann LeCun stated at the time that the company plans to develop world models optimized for tasks such as refining aircraft component designs. 

Image: Unsplash