Protool

AI itself is on the threshold of creating new models: Scientists are worried that humans may lose control over the decisions of machines.


Till now the question regarding AI was how fast it can do the work of humans. But now the concern goes beyond this. Experts are asking – If AI starts creating next generation models on its own, will humans be able to control them? Recently, AI company Anthropic said that there should be an option to slow down or temporarily stop Frontier AI development. According to the company, more than 80% of the code published in May 2026 was written by AI. Jack Clark, co-founder of Anthropic, estimates that by 2028 it is possible for a system to create a new version of itself without human help. This is called recursive self-improvement. That is, one model of AI should be created by another, another by the third and the third by the fourth. Each model is more capable than the last. MIT physicist and AI security expert Max Tegmark warns that if the pace of development is not monitored, AI will overtake humans in decision-making. In such a situation, the human role in governments, companies and important institutions may become weak. Till a few years ago, AI only followed instructions. Now searching for solution. Google DeepMind’s AI system Alpha Evolve suggests ways to improve the efficiency of data centers and speed up AI training. This is the reason why a section of scientists believe that humans will remain research directors instead of researchers. They will just give direction. However, some experts believe that this change will not happen overnight. AI still requires data centers, expensive chips, electricity, and new training materials. His speed will be limited for now. Currently, AI does not create a successor, but no AI model currently undergoing training can completely create its own successor. But big models can make small models themselves. Andrej Karpathy, a former OpenAI researcher, trained a chatbot as capable as GPT-2. GPT-2 was created by OpenAI in 2019. Then it took 168 hours. Carpathy achieved the same result in 3 hours on a computer with 8 GPUs. Later this work was given to the AI ​​agent. He further reduced it by 18%.

Source link

Related Articles

Leave a Reply

Your email address will not be published. Required fields are marked *