Today’s large language models may never be good enough for Artificial General Intelligence (AGI)

The tech industry is dumping hundreds of billions into generative AI and large-language models with the long-term hopes that such systems will become Artificial General Intelligence (AGI).

But AI researchers think they’re wrong. Three-quarters of the 475 AI researchers polled by the Association for the Advancement of Artificial Intelligence said that scaling up current AI approaches to build AGI is unlikely or very unlikely to succeed, while 60% said they don’t believe problems such as inaccuracies can be solved soon.

“The vast investments in scaling, unaccompanied by any comparable efforts to understand what was going on, always seemed to me to be misplaced,” Stuart Russell at the University of California, Berkeley, one of the report’s authors, told New Scientist. “I think that, about a year ago, it started to become obvious to everyone that the benefits of scaling in the conventional sense had plateaued.”

Deep-learning systems get better when bigger, with more parameters applied to larger datasets. But concerns remain about whether all challenges faced by this style of AI can be surmounted purely by size, or if the law of diminishing returns will kick in.

AGI using large language models (LLMs)?

This matters because companies like OpenAI have a stated goal of achieving AGI – but they might not be using the right AI technology to ever achieve that.

The report noted a few challenges that could keep LLMs from evolving into AGI, including a lack of memory, struggling with long-term planning and reasoning, challenges generalising beyond training data, and a lack of deep understanding of physical reality, among others. Quite aside from the fact they don’t really understand idioms.

On a more positive note, the report said that the “aspirational goals” of AGI has historically “inspired many fundamental advances in AI and frame key research questions moving forward to more capable AI systems”.

On the topic of AGI, the survey also found that most respondents (77%) believed the industry should prioritise building AI with an “acceptable risk-benefit profile” over direct pursuit of AGI (23%). And 82% said AGI should be publicly owned.

Seven in ten, however, are against halting research direct at AGI in the name of safety. “These answers seem to suggest a preference for continued exploration of the topic, within some safeguards,” the report stated.

Can we ever trust LLMs?

Those surveyed (around 60%) didn’t believe that current AI models would be able to overcome issues with accuracy, including the so-called “hallucinations” where LLMs make up facts.

Three-quarters of those polled said factuality and trustworthiness were important to their own research. But how to solve the challenge? Of those surveyed, 73% said improving accuracy required more research was required on new neural net architectures, followed by a call for external fact-checking tools, better reinforcement training, improved data quality, and more synthetic training.

When it comes to improving trustworthiness in current models, respondents called for new neural net architectures as well as enabling models to describe the reasoning processes – or even using “understandable models” rather than “black box” neural networks.

But it will take time. “Over the last couple of years, I haven’t seen any evidence that really accurate, highly factual language models are around the corner,” John Thickstun, Assistant Professor of Computer Science at Cornell University, told CNET.

Related articles not written by AGI

• Want to use AI in business? Then think mundane tasks
• Could small language models be the key to unlocking AI value in 2025?
• How Retrieve-Augment-Generate (RAG) improves Gen AI models