Sapient Intelligence accelerates autonomous AI-led research and development

Artificial general intelligence research company Sapient Intelligence has launched PRAXIST, an autonomous AI research and development (R&D) system that takes on complex technical problems and independently tests and validates potential solutions.

While AI has rapidly accelerated productivity in content generation and other well-defined workflows, R&D remains inherently difficult to automate and scale, Sapient Intelligence said. Breakthroughs require testing multiple hypotheses, learning from successes and failures, and continuously determining which paths to pursue. Organisations must address this all while balancing specialist expertise, time, cost, and infrastructure.

PRAXIST, currently in beta, independently explores which technical approach can best achieve a measurable objective. Users define the goal, parameters, and budget; PRAXIST then autonomously experiments and evaluates the strongest solution.

The approach has demonstrated promising results in controlled internal evaluations, Sapient Intelligence revealed. On 75 challenging Kaggle competitions from MLE-Bench, a benchmark designed to test how well AI systems tackle complex, real-world machine learning problems across, PRAXIST achieved the highest-level result in 49 competitions at an approximate recorded model cost of US$3,000. In contrast, the 34 highest-level results for Claude Code cost approximately US$38,000 under the same evaluation conditions.

“Unlike general coding agents, which are built primarily to execute a specific task, PRAXIST is designed for long-horizon R&D, where the problem-solving approach itself may need to be discovered and adapted,” said William Chen, Co-Founder at Sapient Intelligence.

“A conventional research team is ultimately constrained by the number of experiments its researchers can realistically run and evaluate. PRAXIST is designed to provide organisations with the ability to augment their existing teams with additional research capacity, enabling them to explore problems with a breadth and speed that would otherwise require significantly greater specialist resources.”

PRAXIST deploys multiple autonomous research peers to explore approaches in parallel, test hypotheses, investigate failures, and validate promising results. According to Sapient Intelligence, similar systems typically use a tree-like search to accomplish the research. Each candidate under consideration 'inherits' mechanisms, constraints and evidence from a parent, and if the evaluation has a weak outcome, then the entire 'branch' is pruned even if some of the data is valuable. It is common for only one branch to be presented as the best path to the desired outcome.

PRAXIST, on the other hand, uses a generation-layered research graph that preserves what every experiment teaches. This approach allows later generations to combine mechanisms, evidence, and constraints across different lineages, including failed attempts. Rather than searching for the single best branch, PRAXIST constructs stronger solutions from discoveries made across branches.

Use cases

Sapient Intelligence is focused initially on sectors with intensive R&D requirements or highly measurable or easily-simulated optimisation problems, including AI, engineering and robotics, manufacturing, finance, and health and drug discovery. 

For organisations with limited AI or machine learning capabilities, PRAXIST can provide an AI research layer alongside existing domain expertise; for sophisticated teams, it can augment existing capabilities and increase the scale and breadth of R&D. PRAXIST can also operate with proprietary data in private or customer-controlled environments, giving organisations greater control over their research infrastructure and intellectual property.

A traditional robotics company, for example, can define problems, objectives, and constraints from an engineering perspective while PRAXIST conducts the AI research, experimentation, and validation needed to develop solutions, effectively serving as an in-house AI research team.

In a partner-provided rocket-landing simulation, PRAXIST improved baseline to 100% within 12 hours, demonstrating a proof-of-concept capability at technology readiness level (TRL) 3. 

In an industrial robotic SLAM problem, a partner team achieved 9.37 cm of accumulated error after several months of development. PRAXIST reduced the error to 5.01 cm within three days. When measurements are taken many times, a minimal error, such as the measurement tool being off by a fraction of a millimetre, can add up to an amount that can be reported in centimetres. 

Over time, Sapient Intelligence aims to extend PRAXIST beyond traditional R&D into broader business optimisation, from inventory and sales to logistics and shipping.

“AI has mastered executing what we know. The next frontier is discovering what we don’t,” said Jin Li, Chief Scientist of PRAXIST.

“PRAXIST is our first step toward making autonomous discovery a practical capability for organisations tackling complex problems. As we continue to develop the platform, we will look to expand beyond traditional R&D. Our ambition is to give organisations a fundamentally greater capacity to explore what is possible, enabling existing teams to pursue more experiments, approaches, and innovations while keeping human expertise and judgment at the centre.”