Safe Superintelligence, an AI startup founded by a team including former OpenAI chief scientist Ilya Sutskever, is seeking funding at a valuation of at least $20 billion.
This reports Reuters based on sources. Safe Superintelligence (SSI) is in talks with both new and existing investors. SSI’s current investors include Sequoia, DST Global, Andreessen Horowitz, SV Angel and NFDG. This consortium led a $1 billion funding round last September, valuing the company at $5 billion. According to sources, SSI is now seeking funding at a valuation at least four times higher, indicating that investors are optimistic about the company’s technology.
Superintelligence
Safe Superintelligence launched last June with the goal of developing AI models that possess superintelligence, with built-in security measures to prevent malicious output. The founding team included Sutskever, AI researcher Daniel Levy, and Daniel Gross. The latter is a former partner at YCombinator. Gross, SSI’s CEO, also previously led AI development efforts at Apple.
Sutskever stated that SSI focuses exclusively on developing AI models with superintelligence. It does not engage in any other activities. This suggests that the company is not generating revenue in the short term. A few details about how SSI plans to build such models are available. According to Reuters, Sutskever has indicated that the company is pursuing a new research direction rather than using existing AI development methods.
The amount of data is too limited
Developers improve the output quality of large language models by increasing the number of parameters. And by using more hardware and training data. During a recent talk at the NeurIPS machine learning conference, Sutskever suggested that this approach reached its limit. He claimed that developers are struggling to find enough high-quality training data to continue improving LLMs’ output quality. He then argued that people have reached the peak of available data on the Internet, and not much more will be added. According to him, that the industry must work with the data it has. After all, there is only one Internet.
In 2012, Sutskever was part of the academic team that developed AlexNet, one of the first modern computer vision models. This algorithm inspired much of the deep learning research that laid the foundation for contemporary large language models. After this project, Sutskever worked at Google for several years before co-founding OpenAI in 2015.
Reasoning models
Sutskever was ChatGPT’s developer chief scientist until last year. Among other things, he worked on OpenAI’s suite of reasoning-optimized language models. Reasoning models were also an essential part of his recent keynote at NeurIPS.
However, a reasoning system becomes more unpredictable the more it reasons, Sutskever told then. He compared it to chess AIs that are more unpredictable to the best human chess players.
SSI is not the only startup taking a new approach to developing AI models. In December, Liquid AI raised $250 million from investors to establish so-called fluid neural networks. The company claims these algorithms can match the output quality of advanced language models with only a fraction of the hardware required.
Artificial neurons
An AI model consists of artificial neurons, simple programs that each perform a small amount of processing to generate a response to a prompt. Each neuron, in turn, contains components called weights and activation functions. Weights determine which parts of the data an AI model includes in its decisions. Activation functions contain the code that the neuron uses to analyze this data.
In standard LLMs, weights and activation functions remain unchanged after training. Fluid neural networks, on the other hand, can modify these components during inference, allowing such algorithms to continuously adapt to the data they process.
SSI is one of many AI companies reportedly seeking new funding. Last month, it was announced that OpenAI is trying to complete a $40 billion funding round at an estimated valuation of $340 billion. Competitor Anthropic is reportedly seeking $2 billion in funding at a $60 billion valuation.
Also read: Ilya Sutskever wants to do what OpenAI can’t: to develop superintelligence securely