TML 成立7个月首发声：揪出大模型随机元凶，开源方案终结 LLM 推理乱象

Core Insights - The AI industry faces a long-standing technical challenge of output randomness, where the same input can yield different results [1][5] - Thinking Machines Lab (TML), founded by former OpenAI CTO Mira Murati, aims to address this issue and enhance AI reliability [1][3] Company Overview - TML was established in February 2025, four months after Murati left OpenAI, and has raised $2 billion in seed funding, achieving a valuation of $12 billion [3] - The company has not yet released any products but has attracted significant capital interest from major investors including a16z, NVIDIA, AMD, and Cisco [3] Team Composition - TML's team consists of 30 members, with two-thirds coming from OpenAI, including key developers of ChatGPT [4] - AI safety expert Andrew Tulloch joined TML after declining a $1.5 billion rehire offer from Meta [4] Research Focus - TML's mission is not to create stronger AI models but to bridge the gap between AI capabilities and human needs [5] - The core cause of AI output randomness has been identified as batch processing differences rather than "random seed" settings [6][7] Technical Innovations - TML introduced a "batch-invariant kernel" solution to ensure consistent results regardless of data size or grouping [10] - Initial tests showed that previous AI systems could produce up to 80 different answers for the same question, while TML's new approach ensures identical outputs for the same input [10] Performance and Industry Impact - Although the new solution initially slowed AI computation speed by nearly 50%, optimizations have made the performance loss acceptable [12] - This technology is particularly valuable in high-stakes industries like healthcare and finance, where inconsistent AI outputs can lead to critical errors [12] Industry Perspective - TML's approach contrasts with other companies focused on expanding model sizes, instead prioritizing stability and transparency in AI decision-making [15] - The research aims to demystify AI processes, making them more predictable and reliable for societal integration [15][16]