Thinking Machines Lab, led by ex-OpenAI team, claims to fix AI model unpredictability by addressing GPU processing flaws.
The startup's solution ensures consistent AI outputs but results in a 20-50% performance delay.
The research highlights a shift towards transparency, contrasting with OpenAI's increasing secrecy.
Thinking Machines Lab aims to leverage this consistency for more efficient model customization using reinforcement learning.
The lab faces the challenge of converting its scientific breakthroughs into products that justify its $12 billion valuation.
Mira Murati’s secretive startup, Thinking Machines Lab, has published its first piece of research detailing a fix for one of AI’s most fundamental quirks: its unpredictability. By addressing a flaw in how GPUs process information, the ex-OpenAI team says it can make AI models produce identical, deterministic results for the same prompt every time.
Ghost in the machine: The startup claims the randomness that makes models give different answers to the same question isn’t from minor rounding errors, but from a "lack of batch invariance." This means the way AI models process information in groups can slightly alter calculations, creating tiny differences that snowball into noticeably different outputs. The team found that rewriting specific operations to force them to execute in the same order solves the problem.
Predictability has a price: The fix works, but it comes with a trade-off. In an experiment where a standard model produced eighty different answers over a thousand runs, the lab's new approach gave the exact same answer every time. That consistency, however, came with a performance delay of 20 to 50 percent, a price the researchers believe is acceptable for many scientific and enterprise applications.
A jab at the old boss: The research was published on the lab's new blog, "Connectionism," part of a pledge to share its work openly—a pointed contrast to the increasing secrecy at OpenAI. The work also connects to the company's business goals; as The Information previously reported, Thinking Machines plans to use reinforcement learning to customize models, a process that becomes more efficient with the consistency this new research provides.
Murati’s lab has proven it can tackle foundational AI challenges. Now comes the hard part: turning that science into products worth its staggering valuation. Elsewhere in the AI arena, the competitive landscape is heating up as Microsoft, OpenAI’s biggest partner, is reportedly turning to rival Anthropic to power some features in its Office software. In a case of coincidental naming, a completely different company called Thinking Machines Data Science recently became OpenAI’s first official services partner in the Asia-Pacific region.
