New AI startup Ineffable Intelligence reportedly raises $1 billion funding round

Ineffable Intelligence Ltd., a startup led by former Google DeepMind principal investigator David Silver, is reportedly raising $1 billion in funding.

The Financial Times Tuesday quoted sources saying Sequoia Capital is leading the round. DeepMind’s parent company, Alphabet Inc., Nvidia Corp. and Microsoft Corp., could also participate. The round is believed to value Ineffable Intelligence at $4 billion.

The London-based startup does not yet have a product. It was founded last November by Silver, who would have left DeepMind around the same time. During his tenure with the Alphabet unit, he led the development of reinforcement learning technologies for several high-profile AI projects. Ineffable Intelligence reportedly intends to build on this work as part of its research efforts.

One of the DeepMind projects that Silver helped lead is AlphaGo, an AI model trained to play the board game Go. The algorithm made computing history in 2016 by beating the world’s top player. In 2018, Silver’s team launched an improved version of AlphaGo called AlphaZero, which was also adept at chess and shogi.

AlphaGo and AlphaZero were trained using reinforcement learning. A team including Silver also used the technique to train AlphaStar, an AI model optimized for playing a video game called “StarCraft II.” The algorithm represents an important step in the research because “StarCraft II” doesn’t give players complete information about the state of a match, making it more difficult than Go in some ways.

The main concept of reinforcement learning is to train AI models through trial and error. During a reinforcement learning run, a neural network is given a set of training prompts to process. Their response to each prompt is then compared to a correct, prepared response. If the AI ​​model answers a question incorrectly, it receives feedback that helps it improve.

During Silver’s tenure at DeepMind, the Alphabet unit developed not only new ways to apply reinforcement learning, but also improvements to the theoretical underpinnings of the technology.

Reinforcement learning uses an algorithm called a loss function to orchestrate the AI ​​training workflow. Last year, DeepMind detailed a new approach to developing loss functions called DiscoRL. It replaces traditional loss function algorithms with a neural network that can improve over time with the help of AI agents. According to Google, DiscoRL outperforms many previous alternatives.

It’s unclear what reinforcement learning approaches Ineffable Intelligence plans to use or what tasks it will seek to automate with its products. However, the fact that Google and Microsoft are considering investing suggests that the company’s software won’t directly compete with their offerings.

Ineffable Intelligence wouldn’t be the first pre-product AI startup to raise 10-figure funding. Thinking Machines Lab Inc., a company led by Mira Murati, former chief technology officer of OpenAI Group PBC, raised a $2 billion funding round several months before launching its Tinker AI fine-tuning service. Safe Superintelligence Inc. earlier farm an investment of 2 billion dollars for a valuation of 32 billion dollars.

Picture: Unsplash