Mira Murati’s new company, Thinking Machines, claims a fix for one of the most persistent issues in large language model behavior: nondeterministic outputs. The team says it has engineered batch-invariant kernels for key operations so that models return identical responses when run at temperature zero. By removing variation tied to how inputs are grouped or processed, the method is positioned as a way to deliver deterministic outputs on demand, a foundational requirement for trustworthy Artificial Intelligence systems.
The company frames determinism as a practical necessity rather than a theoretical nicety. Nondeterministic behavior makes it difficult to reproduce results, verify model changes, or maintain consistent behavior across runs. Thinking Machines’ batch-invariant kernels are presented as a direct solution to this, asserting that the same prompt will yield the same answer at temperature zero regardless of batch composition. That consistency can simplify debugging, streamline comparisons during model iteration, and make it easier to establish a stable baseline for evaluation.
According to the company’s positioning, the benefits are especially pronounced in research, audits, and safety-critical Artificial Intelligence. In research workflows, deterministic outputs enable repeatable experiments and clearer attribution of performance changes to specific interventions. For audits and compliance, guaranteed identical responses support traceability and verifiable record keeping. In safety-critical contexts, where predictability is paramount, identical outputs at temperature zero help reduce operational risk and strengthen the case for deployment under strict reliability constraints.
Thinking Machines characterizes this advance as a potential shift in how the industry approaches inference standards. If batch-invariant kernels become a reference approach, deterministic, temperature-zero inference could move from a best effort practice to an expected baseline for production systems. The company suggests that the breakthrough may redefine expectations for deterministic, trustworthy Artificial Intelligence, signaling a move toward systems that are more reproducible, easier to audit, and better aligned with safety and governance requirements.
