Singapore startup Sapient is entering the global enterprise AI race

Rational intelligenceSingapore’s first modular AI startup has announced the successful closing of its seed funding round, raising $22 million at a valuation of $200 million.

Backed by prominent investors including Vertex Ventures, Sumitomo Group, and JAFCO, the company hopes to blaze a trail in AI development, addressing what it sees as fundamental flaws in GPT-style models.

“The goal of the startup, in fact, is to create a new generation of basic model architectures to solve complex, long-horizon inference tasks that are a real challenge for solving large language models (LLMs), especially GPT architectures.” said co-founder and CEO Austin Cheng in a recent interview with VentureBeat via video chat.

New architectures go beyond traditional transformers

Traditional GPT style models are based on Autoregressive Methods that generate predictions by sequentially building on previous outputs.

Although effective for general tasks, this approach has difficulty with multi-step reasoning and complex problem solving.

“With the current models, they are all trained using the autoregressive method, however, the benefit is that it is easier for the model to converge on the general task,” Cheng explained. “So they seem really intelligent, so they can solve a lot of different tasks. They have a really good generalization ability, but it’s very difficult for them to solve complex, long-term, multi-step subtasks. And that’s the kind where hallucinations come in,” Cheng said.

Sapient’s answer is a new modular architecture inspired by neuroscience and mathematics, mixing transformer components with recurrent neural network structures and simulating how the human brain works.

“The model will always evaluate the solution, evaluate the options, and give you a reward model based on that,” Cheng said. “The model can also calculate something iteratively until it reaches the correct solution. With this, our agent will be able to deploy in an enterprise or production environment, constantly learning and improving ourselves through trial and error and learning to be experts in the existing code base.”

This design supports the flexibility and power of Sapient models, enabling them to handle a wide range of tasks with precision and reliability.

It also puts them up against a new generation of inference models OpenAI and its o1 seriesbesides Other Chinese competitors.

Excellence within and beyond standards

The company’s innovations are reflected in record performance.

“The first standard we use is actually Sudoku,” Cheng told VentureBeat. “Currently, our model is the best-performing neural network for solving Sudoku on the market – with up to 95% accuracy without the use of intermediate tools and data.”

According to Zheng, while other leading models needed to practice the intermediate steps of solving a common numerical order puzzle, Sapient only provided the model with incomplete Sudoko boards, rules, and final solutions, and had to deduce on its own how to solve it. Through trial and error.

Likewise, Sapient models excel at tasks such as 2D navigation and solving complex mathematical problems, consistently outperforming competing methods.

Training these models is another area where Sapient differentiates itself. “Unlike traditional models that require huge amounts of high-quality step-by-step data, our approach only needs question-answer pairs. This greatly reduces the barrier of training complex models,” Cheng said.

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

Practical applications: from code to robotics

Sapient’s initial focus is on real-world applications, ranging from enterprise coding to robotics.

Its independent coding agents aim to revolutionize how companies manage their software development and maintenance needs.

The company is already deploying an autonomous AI coding agent in Sumitomo enterprise environments to learn about the company’s code base and eventually begin maintaining and contributing to it.

Sapient aims to provide a similar service to other enterprise customers, which Cheng describes as “intelligent, tailored AI employees and AI software engineers who can help them maintain, update, and grow their existing technology stacks as well.”

Unlike Devine perceptionPowered by GPT-4o, Sapent believes its crypto-AI agents will be able to operate autonomously – without any human guidance of the process or troubleshooting, except for supervisors to check the work before it goes live.

The company is also working on developing embodied artificial intelligence, designing models that enable robots to interact, learn and adapt in real time.

“There are only a few startups that are working on understanding the environment, as well as planning options and missions, understanding what kind of missions are possible – and also continuing, improving themselves by understanding the environment, understanding the problem, and understanding the use cases,” Cheng noted. “This will be our main focus over the next year or two.”

Global vision

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

“There are very few seed-level AI startups outside of China that are actually led by Asian founders,” Cheng noted. “We really want to position ourselves as an international research-oriented organization. But we also want to be one of the first few Asia-led international research organizations working on really hard problems, and we are seeing that pay off as well.”

With its offices in Singapore and plans for the Gulf region, the company is building an AI research lab to bring together diverse perspectives and talent.

Its team reflects this spirit, and includes scientists and engineers from leading organizations such as DeepMind, Anthropic, and Microsoft AI.

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

Targeting individuals and institutions

Sapient’s long-term vision is ambitious, targeting technology that can be applied with outcomes that are equally beneficial to individuals and organizations.

“The goal will ultimately be to build a truly generalized agent that can actually solve the everyday tasks of our users — a ‘one-stop agent solution’ for a personal assistant and to solve all your tasks… This is where we are in terms of our technology goal as well as our direction,” Cheng said.

This includes general future products such as standalone coding agents and general-purpose personal assistants.

Currently, Sapient is focused on improving its technology and providing enterprise-class solutions. Pricing models are still being explored but may include licensing, subscription fees, or task-based fees associated with successful completion.

As Sapient expands its operations and capabilities, it remains a company to watch in the rapidly evolving AI landscape.