Safe Superintelligence Inc., a high-profile artificial intelligence startup started by a team that includes former OpenAI Chief Scientist Ilya Sutskever, is reportedly seeking to raise new capital at a valuation of at least $20 billion.

Reuters today cited sources as saying that the company is in talks with both new and existing investors. SSI’s existing investors include Sequoia, DST Global, Andreessen Horowitz, SV Angel and NFDG. The consortium led a $1 billion round for the company last September at a reported valuation of $5 billion.

That SSI is now gearing up to raise capital at a valuation at least four times as high suggests investors are optimistic about its technology. 

The company launched last June to develop AI models that possess “superintelligence” as well as guardrails for blocking harmful output. Its founding team included Sutskever (pictured), AI researcher Daniel Levy and Daniel Gross, a onetime Y Combinator partner. Gross, the company’s chief executive, also led Apple Inc.’s AI development efforts for several years. 

Sutskever has stated that SSI “will not do anything else” besides developing AI models with superintelligence, which suggests it won’t generate revenue in the near term. Few details are available about how SSI plans to go about building such models. According to Reuters, Sutskever has indicated that the company will follow a “new research direction” instead than using existing AI development methods. 

Developers increase large language models’ output quality by boosting their parameter counts, as well as the amount of hardware and training data at their disposal. In a recent talk at the NeurIPS machine learning conference, Sutskever suggested that this approach is nearing its limits. He argued that developers are struggling to source the growing amounts of high-quality training data necessary to keep enhancing LLMs’ output quality. 

“We’ve achieved peak data and there’ll be no more,” Sutskever said. “We have to deal with the data that we have. There’s only one internet.”

In 2012, Sutskever was part of the academic team that created AlexNet, one of the first modern computer vision models. The algorithm inspired a significant amount of deep learning research that helped lay the foundation for large language models. After the project, Sutskever worked at Google LLC for several years before co-founding OpenAI in 2015. 

Sutskever was the ChatGPT developer’s chief scientist until last year. He worked on, among other projects, OpenAI’s series of reasoning-optimized LLMs. Reasoning models were another focus of his recent NeurIPS keynote.

“A system that reasons, the more it reasons, the more unpredictable it becomes,” Sutskever told attendees. “And one reason to see that is because the chess AIs, the really good ones, are unpredictable to the best human chess players.”

SSI is not the only startup taking a new approach to building AI models.

In December, Liquid AI Inc. raised $250 million from investors to develop so-called liquid neural networks. The startup says that such algorithms can match the output quality of cutting-edge LLMs using a fraction of the hardware.

An AI model is made of artificial neurons, simple programs that each perform a small portion of the processing involved in generating a prompt response. Each neuron, in turn, includes components called weights and activation functions. Weights determine which pieces of data an AI model takes into account when making decisions. Activation functions contain the code that the neuron uses to analyze this data. 

Standard LLMs’ weights and activation functions don’t change after training. Liquid neural networks, in contrast, can reconfigure those components during inference. That allows such algorithms to continuously adapt the way they perform processing based on the data they ingest. 

SSI is one of several AI providers reportedly seeking to raise new funding. Last month, word emerged that OpenAI hopes to close a $40 billion round at a $340 million valuation. Rival Anthropic PBC is reportedly seeking to raise $2 billion at a $60 billion valuation.

Photo: OpenAI