As artificial intelligence companies face mounting challenges in developing more powerful language models, some of the world’s most influential AI scientists and investors say a change in training techniques could reshape the industry’s competitive landscape. Leaders at top AI firms, including OpenAI, are now exploring new approaches that mimic human reasoning, moving away from the “bigger is better” philosophy that has defined the field. ‘’
A dozen scientists, researchers, and investors told Reuters they believe these emerging techniques — exemplified by OpenAI’s new “o1” model — could be transformative. By enabling models to “think” more like humans, AI labs are looking to achieve more with existing resources rather than simply adding data and computing power. OpenAI, however, declined to comment.
The AI boom, spurred by the success of OpenAI’s ChatGPT in 2022, has fueled both innovation and stock valuations across tech companies. But as AI pioneers like Ilya Sutskever, co-founder of OpenAI and now head of Safe Superintelligence (SSI), observe, “the limits of scaling up” have become clear. Sutskever, once a leading advocate for massive data-driven model development, now says, “The 2010s were the age of scaling; now we’re back in the age of wonder and discovery.”
AI model training — a high-stakes process involving hundreds of costly chips and months of processing time — has been plagued by delays and equipment failures, sources report. Additionally, vast energy demands and the finite supply of readily available training data have stymied attempts to surpass the performance of existing models like GPT-4.
To address these issues, scientists are developing “test-time compute,” which allows models to process multiple possibilities in real-time during their inference phase, mimicking human decision-making. This approach can allocate extra computational power to complex tasks like mathematics and coding, with promising early results. Noam Brown, an OpenAI researcher, explained at a recent TED AI conference, “Having a bot think for just 20 seconds in a poker game offered the same performance boost as scaling up the model by 100,000 times.”
OpenAI’s o1 model, launched in July, embodies this philosophy, integrating human-like problem-solving capabilities with curated feedback from PhDs and industry experts. This hybrid training approach builds on top of “base” models like GPT-4 and is expected to enhance future iterations of large language models.
Other top AI players, such as Anthropic, xAI, and Google DeepMind, are also investing in these techniques, hoping to refine real-time model inference. “By the time people catch up, we’re going to be three more steps ahead,” said Kevin Weil, OpenAI’s chief product officer, at a conference in October.
The shift toward inference-based training methods could significantly impact AI hardware demand, historically dominated by Nvidia’s training chips. Venture capitalists from firms like Sequoia and Andreessen Horowitz, which have invested billions in the AI sector, are re-evaluating their strategies as inference-based “clouds” gain importance. Sonya Huang, a partner at Sequoia, highlighted that these “inference clouds” may soon replace pre-training clusters.
While Nvidia, recently crowned the world’s most valuable company, leads the training chip market, it could face competition in the emerging inference sector. Nvidia’s CEO Jensen Huang acknowledged this shift last month, saying, “We’ve now discovered a second scaling law, and this is the scaling law at the time of inference,” referencing the surging demand for their newest inference-focused chip, Blackwell.
As AI labs pivot from sheer scale to strategic advancements in inference, the industry is once again on the brink of transformation — redefining how models learn and operate and reshaping the underlying technology that powers them.
Related Readings:
