Singapore startup Sapient joins global enterprise AI race

wisdomSingapore's first basic model artificial intelligence startup announced the successful completion of a seed round of financing, raising US$22 million and a valuation of US$200 million.

Backed by high-profile investors such as Vertex Ventures, Sumitomo Group and JAFCO, the company hopes to carve a unique path in artificial intelligence development and address what it sees as fundamental flaws in the GPT-style model.

“The goal of this startup is actually to create a new generation of basic model architecture to solve very complex and long-term inference tasks that are really challenging for large language models (LLM), especially the GPT architecture,” co-founder and CEO Austin Cheng said in a recent video interview with VentureBeat.

A new architecture beyond traditional Transformers

Traditional GPT style model dependencies autoregressive Method that generates predictions by sequentially constructing previous outputs.

While this approach is effective for general tasks, it struggles with multi-step reasoning and complex problem solving.

“Current models are trained using autoregressive methods. The advantage of this is that the model is more likely to converge on general tasks,” Zheng explained. “So it sounds really smart, so it can solve a lot of different tasks. It generalizes really well, but it's really hard for them to solve a sub-, complex, long-term, multi-step task. It's very difficult. That's where the hallucinations come in,” Zheng said.

Sapient's answer is a novel model architecture inspired by neuroscience and mathematics that mixes transformer components with recurrent neural network structures and mimics the way the human brain works.

“The model will always evaluate solutions, evaluate options, and give itself a reward model accordingly,” Zheng said. “And the model can also keep calculating something in a loop until we get the right solution. This way, our agents will be able to deploy into an environment in an enterprise or production environment and continue to learn and improve themselves through trial and error. , and learn to become an expert on your existing code base.

This design enhances the flexibility and power of Sapient models, allowing them to handle a variety of tasks accurately and reliably.

It also pits them against a new generation of inference models OpenAI and its o1 seriesalso Other Chinese competitors.

Perform well in benchmarks and more

The company's innovation is reflected in benchmark performance.

“The first benchmark we used was actually Sudoku,” Zheng told VentureBeat. “Currently, our model is the best performing neural network on the market when it comes to solving Sudoku problems – achieving 95% accuracy without the use of intermediate tools and data.”

According to Zheng, while other leading models require training intermediate steps to solve popular number-sequencing puzzles, Sapient only provides the model with unfinished Sudoku boards, rules and final solutions, and must deduce on its own how to solve them through trial and error.

Likewise, Sapient's models performed well on tasks such as 2D navigation and complex mathematical problem solving, consistently outperforming competing methods.

Training these models is another area where Sapient differentiates itself. “Unlike traditional models that require large amounts of high-quality, step-by-step data, our method only requires question-and-answer pairs. This greatly lowers the barrier to training complex models,” Zheng said.

By leveraging synthetic data, Sapient reduces reliance on curated data sets, creating scalable and efficient training pipelines.

Practical applications: From code to robots

Sapient's initial focus is on real-world applications, starting with enterprise coding and robotics.

Its autonomous coding agent aims to revolutionize the way enterprises manage their software development and maintenance needs.

The company has deployed autonomous AI coding agents in Sumitomo's enterprise environment to learn the company's code base and eventually begin maintaining and contributing code.

Sapient aims to provide similar services to other enterprise customers, which Zheng describes as “intelligent and tailored AI staff and AI software engineers that can help them maintain, update and evolve their existing technology stack.”

unlike cognitive germanPowered by GPT-4o, Sapent believes its coded AI agents will be able to work autonomously, without the need for anyone to guide the process or troubleshoot issues, except for a supervisor to check the work before going live.

The company is also pushing the boundaries of embedded artificial intelligence, designing models that enable robots to interact, learn and adapt on the fly.

“Only a few startups are committed to understanding the environment, planning options and tasks, and understanding what tasks are possible—and continuously improving themselves in terms of understanding the environment, understanding the problems, and understanding use cases,” Zheng pointed out. “This will be our main focus over the next 1-2 years.”

global perspective

Sapient not only differentiates itself through technology, but also through its global and inclusive approach.

“Outside of China, there are very few AI startups at the base model level that are truly led by Asian founders,” Zheng points out. “We really want to position ourselves as an international, research-led organization. And we want to be one of the first, very few Asian-led international research organizations that are solving really, really challenging problems. problem, and we see that this goal is bearing fruit.

The company, which has an office in Singapore and plans to open one in the Bay Area, is building an artificial intelligence research lab to bring together diverse perspectives and talent.

Its team embodies this ethos, consisting of scientists and engineers from leading institutions such as DeepMind, Anthropic and Microsoft AI.

This diversity, coupled with strong partnerships with Japanese investors such as Sumitomo Group, makes Sapient a unique player in the global AI ecosystem.

For individuals and businesses

Sapient's long-term vision is ambitious, targeting technology that delivers equally useful results for individuals and businesses.

“The ultimate goal is to build a truly universal agent that can actually solve everyday tasks for our users – a 'full agent solution' that serves as a personal assistant and solves all your tasks… That's what we This is true for the technical goals and directions.

This includes future public-facing products such as autonomous coding agents and universal personal assistants.

Currently, Sapient is focused on perfecting its technology and delivering enterprise-grade solutions. Pricing models are still being explored, but may include licensing fees, subscription fees, or task-based fees tied to successful completion.

As Sapient expands its business and capabilities, it remains a company to watch in the rapidly growing field of artificial intelligence.