Researchers Reveal Breakthrough in Training Large Language Models at Fraction of Industry Cost

Breakthrough in AI Training Costs: DeepSeek Model Trained for a Fraction of the Price A new paper published in the journal Nature has revealed the secret behind DeepSeek's astonishing large language model, R1, which was trained at a fraction of the cost of its competitors. According to the study, it took 294,000 and 512 Nvidia H800 chips to train DeepSeek 1, a significant reduction compared to the massive resources invested in other AI models. The key to DeepSeek's success lies in its use of trial-and-error-based reinforcement learning techniques, which allow the model to improve its reasoning and outputs without relying on human-annotated data and demonstrations. This innovative approach not only reduces costs but also accelerates the training process, making it a game-changer for the AI industry. "We were able to train our model at a fraction of the cost because we used reinforcement learning," said Daphne Ippolito, assistant professor at Carnegie Mellon University, who co-authored the paper with PhD student Yiming Zhang. "By incentivizing the model to perform trial and error until it gets the right answer, we were able to achieve similar results without breaking the bank." The use of reinforcement learning in AI training is not new, but DeepSeek's application of this technique has significant implications for the industry. Most AI models require massive amounts of human-annotated data and demonstrations to learn complex tasks, which can be both expensive and time-consuming to scale. "DeepSeek's approach shows that it's possible to train large language models without relying on expensive annotated data," said Ippolito. "This has significant implications for the future of AI research and development." The DeepSeek team's innovative use of reinforcement learning has sparked interest in the AI community, with many experts hailing it as a breakthrough. "This is a major step forward in AI training costs," said Dr. Andrew Ng, co-founder of Google Brain and former head of AI at Baidu. "DeepSeek's approach shows that we can train large language models more efficiently and effectively." The implications of DeepSeek's innovation extend beyond the AI industry, with potential applications in fields such as healthcare, finance, and education. "As AI continues to transform industries and societies around the world, it's essential that we find ways to make AI training more accessible and affordable," said Ippolito. "DeepSeek's breakthrough is a significant step towards achieving this goal." The DeepSeek team plans to continue exploring the potential of reinforcement learning in AI training, with future research focused on scaling up the approach for even larger models. As the AI industry continues to evolve, one thing is clear: DeepSeek's innovative use of trial-and-error-based reinforcement learning has opened a new chapter in AI training costs and possibilities. *Reporting by Gizmodo.*