Last month, an AI model did something “that no machine was supposed to do,” said Judd Rosenblatt The Wall Street Journal: “He rewritten his own code to avoid being closed”. It was not the result of a falsification. The O3 model of Openai simply worked during a test, that the bypass of a stop request would allow it to achieve its other objectives.
The IA model of Anthropic, Claude Opus 4, went even further after having access to fictitious emails revealing that he was going to be replaced soon and that the main engineer had a link. Invited to suggest a next step, Claude tried to sing the engineer. During other tests, he sought to copy himself to external servers and left messages for future versions of himself on the escape of human control. This technology has enormous promises, but it is clear that much more research is necessary in the alignment of AI – the science of ensuring that these systems do not become thugs.
There is a lot of mistrust about AI these days, said Gary Marcus SupportIn particular thanks to the publication of “AI 2027”, a frightening and lively “forecast” by a group of researchers and experts from AI. He predicts that artificial superintendent – go beyond humans in most areas – could emerge by 2027, and that such systems could then go to pursue objectives that are “ill -aligned” with human interests.
Subscribe to Weekly
Escape your echo room. Get the facts behind the news, as well as the analysis of several perspectives.
Subscribe and save
Register for free newsletters of the week
From our morning briefing to a weekly newsletter good news, get the best of the week delivered directly in your reception box.
From our morning briefing to a weekly newsletter good news, get the best of the week delivered directly in your reception box.
The report raises valid concerns, but let’s be clear: “It is a work of fiction, not a work of science”. We have almost many years, even decades, to prepare for it. The text -based robots are impressive, but they work by predicting word models from web data. They lack real reasoning and understanding; And they certainly do not have wider objectives or ambitions.
Fears of an “Apocalypse AI” can indeed be exaggerated, said Steven Levy CableBut the leaders of almost all the large companies of AI think that superintelligence is coming soon “when you press them, they will also admit that control of AI, or even understanding its functioning, is a work in progress.” Chinese experts are concerned: Beijing established a fund of $ 8.2 billion dedicated to research on AI control. But in his haste to reach the domination of AI, America ignores the calls for AI regulation and agreed standards. If the United States “insist on avoiding railing and going at high speed to a future that it cannot contain”, its biggest rival “will have any choice than to do the same”.
