ChatGPT came out on November 30, 2022. Less than four years from today. Since then, the technology has changed fundamentally. AutoGPT followed in 2023, MCP was introduced in 2024. Within a highly compressed timeframe, the industry moved from simple chatbots to fully autonomous Multi-Agent systems. Since the Dartmouth Conference established AI as a research field in 1956, there have been several periods of unprecedented growth and also lengthy periods of stagnation („AI winters“). However, the technological progress has never been as dynamic as it is now, thanks to technological progress, intense competition, and also high levels of investment.
The tone of the discussion around AI has shifted fundamentally as well. The hype and optimism in AI has always been there from the beginning. Users laughed about the fault-ridden AI video of Will Smith eating Spaghetti, among many other shortcomings of GenAI outputs. But progress did not stop, and soon the discussion changed. Geoffrey Hinton, Nobel Prize laureate and globally known for his work on Neural Networks, was the first one to sound the alarm bell as early as May 2023 in an interview with the New York Times. Many of the aspects of the current discussion were already present back then: the alignment problem, bad actors misusing the technology, loss of control, insufficient guardrails.
What followed were letters to slow down AI development throughout the last years, heightened media attention, several cyberattack occurences (e.g. the OpenAI–Hugging Face Incident) and some AI researchers leaving their companies and warning about the dangers of the products their former employer is developing, with Jacob Coxon quitting Anthropic at the beginning of September being the last prominent case. He emphasized the danger of self-improving systems and warned that AI „‚could kill us all by the end of the decade“. Just a few days later, the CEOs of Anthropic and OpenAI, called for a slowdown in the development of frontier models. Within three years, the broader discussion shifted from Will Smith eating Spaghetti to AI ending humanity. This just goes to show how fast things have been changing.
What now? Are we in for another „AI Winter“, but this time a voluntarily decided one? Not likely, for three main reasons. First, AI companies still need to prove that they can actually build and sustain profits and find a successful business model, combined with IPOs planned in the near future. Second, AI is a main driver of global competetiveness and the risk of falling behind weighs larger than the associated security risks, at least for many players in the market. Third, the potential benefits of the technology are so high (e.g. in drug discovery) that there is a real interest here to keep speed high and actually save human lives through progress in medicine and other related areas. The incentives to actually slow down are limited.
What is needed now is a phase of closing the capability-safety gap, where we need to align AI capability growth with guardrail development. Based on the alignment problem and possible AGI/ASI in the near future, this will remain challenging, but not outright impossible. Next, AI safety and security needs to be at the forefront of AI research again, and in a broad coalition of industry, regulators, academic research, and practical experts in the field. More focus is needed, faster professionalization, more ressources. And third, we need to shift the tone of the debate again. OpenAI already warned about GPT2 back in 2019 and with every new model, safety warnings are now part of the overall communication. All discussions are driven by the survival of humankind, it sometimes seems. We need another tone shift where the discussion is focused on the concrete risks of the technology, and the concrete governance mechanisms to address them, be it in 2026 or in a possible future with AGI, ASI, and self-improving supermodels. That means, a discussion rooted in the here and now as well as in the future, but based on evidence, technological expertise, and effective governance. While the race will not slow down, a balance can be achieved when the development of AI governance speeds up.