Thinking Machines Lab Tackles AI Model Consistency with New Research
Thinking Machines Lab Tackles AI Model Consistency with New Research
Thinking Machines Lab, the new venture founded by former OpenAI CTO Mira Murati, is aiming to solve a challenge that has long puzzled the AI community: making large language models (LLMs) deliver consistent, repeatable answers.
Why Consistency in AI Matters
Most people who have used AI chatbots like ChatGPT have noticed that asking the same question multiple times often yields different answers. This randomness, or nondeterminism, is a widely acknowledged trait of today’s AI models. While variety can be entertaining, it presents real challenges for researchers, businesses, and developers who rely on predictable AI outputs.
Inside the Research: Unpacking AI Randomness
On September 10, 2025, Thinking Machines Lab published its inaugural research blog post, “Defeating Nondeterminism in LLM Inference”, authored by researcher Horace He. The post explores what causes AI models to behave unpredictably, pinpointing the process by which GPU kernels—the core programs running on chips like those from Nvidia—are orchestrated during AI inference (the step after you hit "enter" in a chatbot).
- He argues that inconsistencies arise from the way these kernels are managed.
- By introducing more control and precision at this layer, it may be possible to make AI responses reproducible.
Potential Impact for Businesses and AI Training
Deterministic AI models could transform enterprise applications, research workflows, and even how AI is trained. For instance, reinforcement learning—where models are improved through reward feedback—could become more effective if responses are consistent, reducing noisy training data and improving reliability.
Thinking Machines Lab has indicated its ambition to use reinforcement learning to tailor AI models for business needs, potentially giving enterprises more control and confidence when deploying AI solutions.
Commitment to Open Research
The lab’s new blog series, Connectionism, marks a pledge to share research, code, and insights with the public. This open approach sets Thinking Machines Lab apart in an industry where leading companies have become increasingly secretive about their breakthroughs.
Mira Murati has announced that the lab’s first product—intended to support researchers and startups building custom models—will launch in the coming months. While details remain under wraps, the focus on reproducibility hints at a tool or service designed to set a new standard for reliability in AI.
What’s Next?
With a $2 billion seed round and a team of top AI talent, Thinking Machines Lab is well positioned to address one of the most pressing technical challenges in AI today. Their commitment to open research and solving non-determinism could have ripple effects across the industry, especially for businesses that demand consistency from their AI systems.
