'AI responses less random': Mira Murati's lab drops major research
Mira Murati's Thinking Machines Lab—which has raised $2 billion in seed funding and is now valued at $12 billion—just dropped its first blog post (September 10, 2025), sharing new research on making AI responses less random and more repeatable.
This tackles a big headache with today's large language models: you can ask the same question twice and get totally different answers.
The lab believes it can make AI responses more consistent
Researcher Horace He says the culprit is how NVIDIA GPU kernels handle things behind the scenes when you prompt an AI.
By taking tighter control of this process, the lab believes they can make AIs give consistent answers every time—a pretty bold move since most people think unpredictability is just part of how AI works.
If this works, it could make training AIs way easier
If AIs become more predictable, it could make training them way smoother and help businesses customize models without all the usual chaos.
The lab wants to keep their research open and share updates regularly—which is refreshing compared to bigger companies that keep their work secret.
Their 1st product is coming soon, and it's for researchers
Thinking Machines Lab is gearing up to launch its first product for researchers and startups.
Their open approach might help demystify how these complex systems work—and potentially make advanced AI tools easier for everyone to use.