Sejal Sharma is IE’s AI columnist, offering deep dives into the world of artificial intelligence and its transformative impact across industries. Her bimonthly AI Logs column explores the latest trends, breakthroughs and ethical dilemmas in AI, delivering expert analysis and new insights. To stay informed, subscribe to our AI Logs newsletter for exclusive content.
Looking at some of cinema’s greatest and most chilling villains, HAL 9000 from Stanley Kubrick’s 1968 Marvel 2001: A space odyssey easily ranks in the top 10.
HAL is one of the main depictions of an artificial general intelligence (AGI), a human-level or higher intellect in an artificial being. We never see him, but his words are eerily ubiquitous in the film. “I’m sorry, Dave. I’m afraid I can’t do that,” HAL responds to Dr. David Bowman’s request to reenter the spacecraft, paving the way for Bowman’s slow and suffocating demise.
HAL’s calm but increasingly menacing tone evokes a sense of dread, reminiscent of the time when Sir Anthony Hopkins played Hannibal Lecter in 1991 and was inspired by HAL.
Fast forward to today, HAL’s chilling presence feels uncomfortably relevant in the race of big tech companies to develop AGI. With industry leaders predicting its arrival in just a few years, the question becomes: are we building the next HAL? Or is true AGI still a distant dream, overshadowed by challenges?
While these questions have yet to be answered, some top leaders in the AI industry claim that AGI will be realized by 2025 or 2026.
In a recent interview, when Y Combinator CEO Garry Tan asked OpenAI CEO Sam Altman what he’s excited about in 2025, Altman said, “AGI… that’s what I’m most excited about in my life “, without addressing whether OpenAI has done so. have already developed it in their laboratories or not.
While venture capitalists and tech bros heading major companies are inherently optimistic about the future of their products, Altman’s prediction could have been swept under the rug if OpenAI hadn’t been the fastest growing AI company in the world. But another AI CEO who made a similar prediction has added fuel to all this AGI fire.
Dario Amodei, CEO of Anthropic – the maker of AI chatbot Claude and a competitor to OpenAI’s ChatGPT – also recently said that AGI can be expected in 2026 or 2027.
Commenting on the rapid progress of generative AI, Amodei explained: “We are starting to reach the PhD level, and last year we were at the bachelor’s level and the year before that we were at the high school student level… If you look at the The rate at which these capabilities are increasing, you think we’ll be there in 2026 or 2027.”
Millions of dollars depend on these expectations. The size of the global AI market has increased from $124 billion in 2022 to $184 billion in 2024. The industry is expected to be worth $826 billion by 2030.
But Amodei was cautious, also saying that many things can go wrong and derail AGI’s development. For example: “We may no longer have any data. We may not be able to scale clusters as much as we want. Maybe Taiwan will get blown up or something and we won’t be able to produce as many GPUs as we want,” he said facetiously.
Is AI showing signs of slowing down?
Too bad Amodei didn’t bet on things going wrong (although that would have been a bet against himself), but in the last few months alone we’re seeing signs that things are moving… rather slowly.
One of the things that the AI industry has really believed in is that the more data an AI model will train, the more efficient it will become. By that logic, super-intelligent AI would be the result of the availability of big data, accelerated increases in computing power, and advances in machine learning. But some AI industry pioneers seem to contradict this.
Ilya Sutskever, co-founder of Safe Superintelligence and OpenAI, said the big gains from training AI on massive amounts of data are starting to slow.
“The 2010s were the era of scaling, now we’re back in the era of wonder and discovery. Everyone is looking for the next thing,” Sutskever said Reuters. “Scaling right is now more important than ever.”
In a separate report, some insiders said The information that OpenAI’s Orion excels at language tasks, but stumbles in coding – an area where expectations are sky-high. Another one Bloomberg The report claims that the jump from GPT-4 to Orion seems much less impressive than the big steps from GPT-3 to GPT-4, raising eyebrows.
This trend is not unique to OpenAI. Other major players such as Google and Anthropic are also facing challenges with their latest models. These difficulties are leading to discussions about the need for new approaches to training AI that go beyond simply increasing model size and computing power.
A big part of the problem is also the AI industry’s Achilles heel: data. OpenAI is running low on new training data and has assembled a special team to look for more. Meanwhile, they try solutions like refining models after training, but that only hides the rifts.
Waving on mathematical reasoning
Another thing that shows that AI may not be where we thought it was is their performance in mathematical reasoning. Mathematical problems are widely used as benchmarks because they require an understanding of rules, the ability to manipulate them flexibly, and creativity in finding solutions – qualities often associated with human-like intelligence.
A study published four days ago introduced a new benchmark called FrontierMath – a test of hundreds of highly advanced mathematical reasoning problems – to evaluate AI models. Researchers tested six top AI language models, namely OpenAI’s o1-preview, o1-mini and GPT-4o, and Anthropic’s Claude 3.5 Sonnet, xAI’s Grok 2 Beta and Google’s Gemini 1.5 Pro.
The results were disappointing: none of the models could solve even two percent of the problems, much worse than their performance on other math tests such as GSM8K or AIME, where they almost maxed out their potential.
They also took a closer look at the few problems that each model managed to solve. Even then, the results were inconsistent. For example, only one model (o1 preview) correctly solved an issue every time it was tested.
There is no one true definition of AGI
AI pioneer Nils Nilsson’s 2005 prediction that “achieving true human-level AI would necessarily imply that most tasks that humans perform for pay could be automated” highlights the challenge of defining AGI, because it raises questions about its social implications.
There will be a lot of debate when (and if) we finally reach AGI. And without a clear definition, researchers may disagree about whether a system truly qualifies as AGI.
This ambiguity was recently highlighted by Fei-Fei Li, the so-called ‘godmother of AI’ and a key figure in the development of AI. She said she doesn’t really know what AGI means.
“I come from academic AI and was trained in the more rigorous and evidence-based methods, so I don’t really know what all these words mean,” says Li. “To be honest, I don’t even know what AGI means. Like people say you know it when you see it, I guess I haven’t seen it.
Others, like Altman and Amodei, define AGI more concretely, likening it to working with a highly intelligent human colleague. However, these definitions vary, and some researchers still question whether these AI systems meet the necessary criteria.
One thing is certain: scaling up will be essential. OpenAI outlines a layered approach to AGI, with five levels of AI intelligence ranging from simple chatbots to fully autonomous systems capable of running organizations, which it calls AGI.
With the launch of the o1 model, OpenAI is currently at level 2: the level of reasoners. And early next year, according to reports, it is expected to reach Level 3 after the launch of an autonomous AI agent codenamed ‘Operator’.
However, given the pace of progress and recent reports of AI models showing signs of slowing, Altman’s prediction to achieve AGI by 2025 seems ambitious. How can AGI be achieved by jumping two levels in a year, especially when the current trajectory suggests that AI’s rapid progress may face obstacles?
Ultimately, as we continue to develop AI models, we may end up in a situation similar to that of HAL 9000: a machine that appears to have intelligence, yet is somewhat far removed from true human-like understanding. And given the many different definitions of AGI, is there a chance that we might not recognize it when we finally see it?
