OpenAI's chief scientist, Jakub Pachocki, warned that AI is growing faster than designed, with systems now surpassing human intelligence in key areas. This is the company's first major safety update since 2023, when early results from the RLSlow project hinted at the potential for scalable reasoning models.
OpenAI reported significant progress in AI capabilities, with systems now able to operate computers, collaborate with humans, and perform complex research tasks. These developments have transformed computer security and introduced new risks.
The advancements are driven by increased computational power and new algorithms, with AI systems now exhibiting behaviors that are difficult to fully understand. Pachocki emphasized that the intelligence produced by scaling deep learning is not directly comparable to human intelligence.
"AI is grown more than designed - it is, to first degree, the product of repeating a straightforward optimization step many times on a hard-to-imagine amount of compute," said Jakub Pachocki, Chief Scientist at OpenAI. This results in an incredibly complex system that works through abstract concepts and can simulate facets of human behavior.
The announcement follows three years of rapid progress in AI, with OpenAI highlighting the need for extreme caution as the pace of development continues. Pach, as he is known, expressed concerns that no one is prepared for the consequences of a continued rapid rise in machine intelligence.
OpenAI did not say when further scaling will be halted, and raised the open question of how to ensure AI systems continue to hold human values regardless of their environment. The company will continue to seek technical solutions to alignment and monitoring, and may unilaterally withhold further scaling as needed.
Source: openai
