Artificial Intelligence: What comes after ChatGPT?
There have been plans, promises and announcements for decades. The goal is a machine that can think like a human – or better. Since the end of 2022, technical developments appear to have accelerated dramatically. Now an breaking latest news, a general artificial intelligence, seems within reach. In the current issue, MIT Technology Review attempts to take stock of developments.
Advertisement
However, what exactly an breaking latest news should be is not even clearly defined. When AI researchers Shane Legg and Ben Goertzel popularized the term breaking latest news in the early 2000s, they wrote somewhat vaguely of an AI that would need to be able to “perform a range of cognitive tasks that humans can solve.” . They did not explain further what kind of people and what kinds of tasks they meant. Today, Goertzel sells his esoteric visions of the future on the event platform Singularity.net. Shane Legg co-founded DeepMind and now works as the Google subsidiary’s “Chief breaking latest news Scientist”. Together with seven of his colleagues, he published a paper in November 2023 that was intended to advance the seemingly endless debate about definitions and criteria. In it, Legg and his team list nine different definitions of the term.
Is the breaking latest news already there or not yet?
Only one thing seems certain at the moment: the community agrees to disagree. Blaise Agüera y Arcas of Google Research and Peter Norvig of Stanford University, for example, wrote in an essay for the journal Noema as early as 2023 that breaking latest news “is already here.” Large language models could solve a variety of different tasks and learn new skills from the context and with just a few examples. Both require human-like intelligence.
If you believe AI critics like the cognitive researcher Gary Marcus or the computational linguist Emily Bender, large language models, on the other hand, are as dumb as bread. They function purely statistically and do not understand a single bit of the language they are processing. No trace of internal world models.
AI expert and complexity researcher Melanie Mitchell from the Santa Fe Institute, on the other hand, argues that there is no way to know that. There would be at least “some understanding” in these models – “in certain contexts” they could apply existing knowledge to new situations, “but in other cases not”. Since we simply don’t know enough about the inner function of the models, the question can hardly be answered meaningfully at the moment.
After the guesswork surrounding OpenAI’s possible breakthrough in the field of general artificial intelligence, it’s time to consider the possible consequences: What will it mean if we one day have to deal with a human-like artificial intelligence? Highlights from the magazine:
Political agenda
And Timnit Gebru, one of the most prominent critics of the AI plans of the large Silicon Valley companies, suspects that there is no technical agenda behind the project, but rather a political one: Together with other authors such as Émile P. Torres, Gebru draws a historical line from the American eugenicists via the transhumanists to the leading minds of OpenAI, which was never about the future and well-being of humanity as a whole, but rather about sorting out everything useless and superfluous.
What current AI systems are undoubtedly still missing, almost all experts agree, is the ability to learn independently. However, there is some pretty exciting research on this.
Icelandic AI researcher Kristinn Thórisson, who has been researching breaking latest news for 30 years, wants to develop AI agents that are completely autonomous – starting with only basic knowledge about learning itself and then changing their own programming over time. His role model: a child’s learning. Back in 2008, Thórisson and his team developed a demonstrator for these capabilities: the Autocatalytic Endogenous Reflective Architecture (AERA).
This “is a cognitive architecture – and a blueprint – for constructing agents with a high degree of operational autonomy,” he writes. She starts with “a small amount of designer-specified code – a seed,” and then modifies her own programming. He now wants to further develop the model with fresh research funding.
Autonomous agents develop unknown recipes
Pierre-Yves Oudeyer from the University of Bordeaux calls such “developmental AIs” that learn “continuously and self-motivatedly” “autotelic” – giving themselves meaning. However, his goal is not AGIs – Oudeyer and his team are more interested in modeling learning. The researchers coupled autotelic software agents with large language models: The team released software agents that can independently explore their environment into a virtual kitchen that, in addition to some food, furniture and tools, also contained a virtual cookbook with a few short recipes . In fact, the agents coupled with a large language model were able to develop some new, previously unknown recipes by combining and rearranging what they had learned over time. However, the researchers complain that the use of large, commercial language models is currently far too expensive, and the models are too opaque to experiment on a large scale.
“Making predictions about something that has never been seen and may not even exist, whether it is extraterrestrial life or superintelligent machines, requires theories based on general principles,” writes Melanie Mitchell in a recent guest post for Science. “Ultimately, the meaning and consequences of ‘breaking latest news’ will be resolved not through media debates, court cases, or our intuitions and speculations, but through long-term scientific exploration of these principles.”
Read MIT Technology Review here:
(wst)
To home page