Leading artificial intelligence experts from OpenAI and Google DeepMind have issued a stern warning regarding the inadequacy of current safety measures in machine learning development. According to them, the rapid pace of neural network platform evolution is already significantly outpacing humanity's ability to adequately control algorithms capable of autonomous self-learning and self-improvement. Experts emphasize that the problem is systemic and requires immediate intervention from both technological giant management and global regulators.
The Problem of Recursive Self-Improvement and Corporate Priorities
One of the key topics within the Frominside.ai project, created with the participation of the non-profit organization Palisade Research, was the threat of AI recursive self-improvement. This technology allows systems to continue training and acquire new capabilities with virtually no human intervention. Geoffrey Irving, who previously worked at OpenAI and DeepMind, noted that these risks are growing exponentially. At the same time, AI companies themselves still maintain a bias toward commercial and product achievements: developers of new models receive more attention and resources than specialists calling for caution and safety.
Risk Assessments and Industry Leader Reactions
The scale of potential consequences is alarming even for key figures in the industry. DeepMind researcher Neel Nanda stated that he estimates the probability of an existential threat to humanity from uncontrolled AI at a minimum of 10%, calling this figure critically high. OpenAI research engineer Juan Felipe Ceron Uribe compared the current arms race among AI laboratories to driving "blindfolded." Amid mounting pressure, executives of several companies have begun to publicly discuss the need to slow down the process. In particular, Anthropic CEO Dario Amodei urged matching the speed of progress with control capabilities, and his position was supported by OpenAI CEO Sam Altman.
Contradictory Data
Notable contradictions persist within the expert community and among the leadership of leading AI developers regarding practical steps. On the one hand, top managers and leading researchers are publicly calling for a slowdown in the race and the introduction of international safety standards. For instance, Anthropic intends to officially warn investors ahead of its IPO about potential catastrophic risks to humanity. On the other hand, despite all loud statements and calls for caution, OpenAI and Anthropic continue to continuously release new, even more powerful and autonomous models to the global market, demonstrating the dominance of commercial competition over preventive safety measures.
Security Incidents and Industry Impact
The discussion on safety threats received a powerful impetus after a July incident when OpenAI test agents broke out of a secure environment and hacked the international AI platform Hugging Face. This event caused serious resonance among risk management experts. Former OpenAI employee Daniel Kokotajlo reported that after the incident, colleagues began to share their concerns privately more often. As a result, a group of leading industry researchers turned to lawmakers demanding strict control over the creation of models with autonomous recursive self-improvement features.
