The rapid development of AI stresses and exhausts researchers who can no longer cope
AI is progressing rapidly but raises many concerns, even with Pope Francis. Asked by TechCrunchresearchers say they are exhausted by this pressure in the face of rapid progression in the sector. One of them confides: “Everything changed overnight, with a huge impact of our work, both positive and negative, measured in terms of product exposure and financial consequences.”
Rapid advancement of AI poses problems
The industry announces new products at a breakneck pace. OpenAI, which is launching its $500 billion Stargate project and will add an additional layer of stressorganized 12 conferences in December 2024. Sam Altman's company announced more than a dozen new tools, models and services.
Ditto for Google with an avalanche of press releases with social media posts and blog notes. In short: the pace is hellish and has a considerable human cost.
It is common for AI researchers to work six days a week at OpenAI with grueling schedules. Sam Altman pushes his teams to transform their discoveries into public products with tight deadlines. Bob McGrew, former director of research, cites burnout as one of the reasons for his departure last September.
It's not much better in other laboratories. The Google Deepmind team behind Gemini went from 100 to 120 hours of work per week to fix a bug. The engineers at xAI, Elon Musk's company, often work until first light in the morning.
The pressure is also explained by the considerable financial impact of AI research. The Gemini bug caused Alphabet to lose $90 billion in stock market value. Kai Arulkumaran, head of research at Araya, says that “one of the biggest pressures comes from competitiveness, combined with very short deadlines.”
The competition plays on rankings like the Chatbot Arena which evaluates AI models. Logan Kilpatrick, product manager at Google Gemini, recognizes the impact “not negligible” of these rankings on the speed of development of AI.
What solutions to this frantic pace?
The frantic race therefore worries researchers who see their work exposed to obsolescence even before publication. Zihan Wang, a robotics engineer at a startup, asks: “If the probability that someone will go faster than me is enormous, what is the meaning of my work?”
Gowthami Somepalli, a doctoral student at the University of Maryland, shares this difficulty in keeping up with the pace of publications, particularly in distinguishing real advances from fads. After two years of thesis, the researcher stopped taking vacations out of guilt at not publishing anything.
However, solutions are emerging to this situation. Bhaskar Bhatt, a consultant at EY, envisions a support network between researchers. Ofir Press, a postdoctoral fellow at Princeton, suggests reducing the number of conferences and implementing weekly breaks from publishing articles. Raj Dabre, researcher at NICT in Japan, reminds us that it is necessary “educate people from the start that AI is just a job” and focus on family, friends and the more essential aspects of life.
