Essential AI Labs Inc., a startup led by two co-inventors of the foundational Transformer neural network architecture, today announced that it has raised $56.5 million from a group of prominent backers.

March Capital led the Series A investment with contributions from Nvidia Corp., Advanced Micro Devices Inc. and Google LLC. The round also included the participation of venture capital firms Franklin Venture Partners, KB Investment and Thrive Capital, which led an earlier $8.3 million seed raise for Essential AI.

San Francisco-based Essential AI was founded earlier this year by machine learning researchers Ashish Vaswani and Niki Parmar. They previously worked at Google, where they were part of the eight-person team that introduced the Transformer neural network architecture in June 2017. The architecture underpins OpenAI’s GPT-4, Google’s Gemini and most of the other advanced large language models on the market.

The primary innovation in Transformer models is that they include a component known as an attention mechanism. When a Transformer model seeks to understand the meaning of a given word in a sentence, it analyzes all the other words in that sentence. It then identifies the text snippets that most directly influence the meaning of the word being analyzed and uses them to make a decision. This prioritization, which is facilitated by Transformer models’ attention mechanism, allows them to understand prose significantly more accurately than earlier neural networks.

Prior to the introduction of Transformers, there were already AI architectures with the ability to prioritize what text they take into account when determining the meaning of a word. However, those architectures process a large amount of unnecessary information before making a decision. That unnecessary information decreases their accuracy.

The original Transformer model that Vaswani, Parmar and the architecture’s six other co-inventors introduced in 2017 was designed for translation tasks. It comprised two components known as the encoder and decoder, respectively. The encoder processes input data, such as a document uploaded by the user, while the decoder generates output, in this case a version of the document translated into a different language.

There are some overlaps between the two components. A Transformer model’s decoder can carry out many of the tasks an encoder is usually be responsible for performing and vice versa, with the difference that the former component generally considers less data when making decisions. Many of the most advanced language models on the market today, including Google’s Gemini series, use a decoder-only design that doesn’t include an encoder.

Bloomberg reported today that Essential AI will use the proceeds from its new Series A round for “corporate functions” such as data analytics. The company didn’t disclose the valuation at which it raised the capital. 

On its one-page website, Essential AI states that one target use case of the language models it’s building will be making data scientists more productive. Additionally, it’s working to help business users “become independent data driven decision makers themselves.” This suggests the company’s models will enable users to carry out analytics tasks that typically require technical know-how using natural language commands, a feature that OpenAI’s GPT-4 offers as well.

Essential AI says another focus of its AI development efforts is easing the work of financial analytics. According to its website, the goal is to increase the number of companies that analysts can cover and improve the quality of their assessments.

Photo: Essential AI