What is likely to be the most significant event in the history of artificial intelligence (AI) took place last Friday when OpenAI, the company behind ChatGPT, unveiled its o3 model — possibly the most powerful AI model developed to date.
Though still in testing and unavailable to the public, the data shared by OpenAI prompted some commentators to declare that we’re finally witnessing the birth of artificial general intelligence (AGI), the “superintelligence” surpassing human intellect.
The announcement marked the grand finale of OpenAI's 12-day AI reveal streak, aimed at cementing the company's status as the undisputed leader in artificial intelligence and reminding investors and the public not to forget it.
However, one issue disrupted this streak — tech giants like Amazon and Google, also deeply invested in AI development, have no intention of letting Sam Altman’s company dominate the field. The competition resulted in two weeks of back-to-back announcements and a flurry of new AI models from industry titans.
Even though Google demonstrated its capabilities by launching the impressive Gemini 2.0 Flash version of its Gemini model, it held the title of the most advanced model on the market for just about a week — until OpenAI unveiled o3, once again tipping the scales in its favor.
The o3 model is the successor to o1, which followed the lineage of GPT models. Interestingly, the skip over o2 sparked curiosity, with some speculating it was a secret model. However, OpenAI provided a more mundane explanation: the name was already taken by a well-known British telecom company and no one wanted to court unnecessary complications.
According to OpenAI's data, o3 is over 20% more accurate than o1 in several programming tasks. The model even outperformed OpenAI's chief scientist, Jacob Pachucki, in a programming competition.
"This model is astonishing in its programming capabilities," Altman said during the system's presentation. The model also achieved exceptional results in math, physics, chemistry and biology tests.
Like its predecessor, o3 is a reasoning model, meaning it rigorously checks itself before delivering an answer, performing millions of calculations, comparisons and alternative evaluations.
This leads to better answers than previous-generation AI models but also adds some delay to response times. The o3 model allows users to predefine the depth of reasoning required — low, medium or high.
Until now, we’ve thought of AGI as “superintelligence,” a godlike intellect, as Ian Hogarth described, that could dominate humanity or even extinguish it. OpenAI, however, has a more restrained definition of AGI: “A highly autonomous system that outperforms humans at most economically valuable work.” Even by this definition, achieving AGI remains a formidable challenge for AI.
Still, according to OpenAI’s data, it’s approaching that goal. In the ARC-AGI benchmark test, which measures an AI’s ability to learn tasks it wasn’t trained on, o3 scored 87.5. While that’s not a perfect 100, it’s three times the achievement of o1.
The issue with this, however, is that benchmark tests are frequently updated and an AI optimized for one test might fail another of the same type.
In response, ARC-AGI test creator Francois Chollets noted that in the test’s next version, the model might only score 30, whereas an average human would score 95, simply because humans excel at adapting to new tasks. So, the path to an omniscient AI is likely still a long one.
Meanwhile, o3 raises some concerning questions. Firstly, operating it at a high reasoning depth consumes enormous computational and electrical resources, casting doubt on whether any economic model could sustain it.
Secondly, its predecessor, o1, already demonstrated a greater tendency to mislead humans than earlier AI generations and OpenAI must prove this trend hasn’t continued.
The company claims to have developed a new approach it calls “deliberative alignment,” designed to keep the model on track. This remains to be proven if OpenAI wants to solidify its position as the AI industry leader.
Get the Ynetnews app on your smartphone: Google Play: https://bit.ly/4eJ37pE | Apple App Store: https://bit.ly/3ZL7iNv
Beyond its technological achievement and unprecedented economic implications, reaching the AGI milestone holds another significant meaning: from that point, OpenAI will no longer be obligated to provide Microsoft with access to its future technologies.
For now, o3 will be available in two versions: a full version and a smaller version called o3 mini. The mini version is expected to be publicly available by the end of next month, with the full version to follow later.