Thinking Machines Lab, led by former OpenAI Chief Technology Officer Mira Murati, is focused on solving one of AI’s most persistent challenges: inconsistent responses.
Today’s AI models are known for nondeterminism. This means that when users ask the same question multiple times, the model may deliver different answers.
For average users, the change may appear minor. However, for businesses, researchers, and developers, this inconsistency poses a serious problem.
Reliable results are essential in scientific research, customer service, and enterprise applications.
And without consistent answers, trust in AI wanes. Thinking Machines Lab wants to address this issue directly.
Research Approach
In its first blog post, titled “Defeating Nondeterminism in LLM Inference,” the lab outlined its strategy.
Researcher Horace He argued that the main source of randomness lies in how GPU kernels are executed.
These small programs, which power Nvidia’s chips, often run in slightly different ways during inference, the step after a user submits a query.
This variation produces unpredictable results. By carefully controlling how these kernels are orchestrated, He believes AI models can be made more deterministic.
In other words, the same question should always yield the same answer.
Determinism
Consistency could improve more than just user trust. It also has implications for reinforcement learning (RL).
RL trains AI systems by rewarding correct answers. Yet when answers change with each attempt, the training data becomes noisy.
More consistent responses would make RL training smoother and more efficient. This research also ties to Thinking Machines Lab’s long-term strategy.
The company has told investors that it plans to use reinforcement learning to customize models for businesses. Determinism could make this process faster and more reliable.
Future Plans
Murati has said that the company’s first product will arrive in the coming months. She described it as a tool for researchers and startups building custom models.
It is not yet clear whether reproducibility research will be part of that product. Still, the announcement reflects the lab’s priorities.
The blog post also launched a new series called “Connectionism.” Through this series, the lab promises to share research, code, and insights with the public.
This open stance contrasts with the industry trend. Companies such as OpenAI began with open research but later became more guarded as they grew.
Whether Thinking Machines Lab maintains transparency over time remains to be seen.
