AI researchers hype check AI claims, doubt current models will achieve AGI

The singularity is near…ly wrong about the date?

The gap between AI hype and technical reality is widening, with most AI researchers now deeply skeptical that current approaches will lead to artificial general intelligence. A new survey reveals that the tech industry’s long-held belief that simply scaling up existing models will produce human-level AI capabilities is losing credibility, even as companies prepare to spend trillions on AI infrastructure. This shift marks a significant departure from the optimism that has characterized the generative AI boom since 2022.

The big picture: Approximately 76% of AI researchers surveyed believe scaling current approaches is “unlikely” or “very unlikely” to achieve artificial general intelligence (AGI), according to a new report from the Association for the Advancement of Artificial Intelligence.

• This represents a marked shift from the “scaling is all you need” philosophy that has dominated the tech industry since the beginning of the generative AI boom in 2022.
• Despite this growing skepticism, tech companies are collectively planning to invest an estimated $1 trillion on data centers and chips in the coming years to support their AI initiatives.

Behind the plateau: Recent advances in transformer models, which form the backbone of today’s AI systems, have shown diminishing returns despite being trained on ever-increasing volumes of data.

• The latest releases of these models have demonstrated only incremental improvements in quality, suggesting a performance plateau.
• Stuart Russell of UC Berkeley noted that “about a year ago, it started to become obvious to everyone that the benefits of scaling in the conventional sense had plateaued.”

The perception gap: A striking 80% of survey respondents indicated that current public perceptions of AI capabilities don’t align with reality.

• “Systems proclaimed to be matching human performance – such as on coding problems or mathematics problems – still make bone-headed mistakes,” explains Thomas Dietterich from Oregon State University.
• These AI systems remain valuable as assistive tools for research and coding but are far from replacing human workers, according to researchers.

Alternative approaches: Companies have begun exploring other paths to improvement beyond simple data scaling.

• Arvind Narayanan of Princeton University points to “inference-time scaling,” where models use more computing power and take longer to process queries, but cautions this is “unlikely to be a silver bullet” for reaching AGI.

Defining the goalpost: The tech industry lacks consensus on what constitutes AGI, further complicating progress assessments.

• Google DeepMind defines AGI as a system that outperforms humans on cognitive tests, while Huawei suggests AGI requires a physical body to interact with its environment.
• Microsoft and OpenAI have taken a commercial approach, reportedly considering AGI achieved only when OpenAI develops a model generating $100 billion in profit.