OpenAI’s new model can reason before answering
With the introduction of OpenAI’s o1 version, ChatGPT users now have the opportunity to test an AI model that pauses to “think” before responding.
According to this article, the o1 model feels like one step forward and two steps back when compared to the GPT-4o. Although OpenAI o1 is superior to GPT-4o in terms of reasoning and answering complicated questions, its cost of use is around four times higher. In addition, the tools, multimodal capabilities, and speed that made GPT-4o so remarkable are missing from OpenAI’s most recent model.
The fundamental ideas that underpin o1 date back many years. According to Andy Harrison, the CEO of the S32 firm and a former Google employee, Google employed comparable strategies in 2016 to develop AlphaGo, the first artificial intelligence system to defeat a world champion in the board game Go. AlphaGo learned by repeatedly competing with itself; in essence, it was self-taught until it acquired superhuman abilities.
OpenAI improved the model training method so that the reasoning process of the model resembled how a student would learn to tackle challenging tasks. Usually, when someone comes up with a solution, they identify the errors being made and consider other strategies. When a method does not work, the o1 model learns to try another one. As the model continues to reason, this process gets better. O1 improves its reasoning on tasks the longer it thinks.
Pros and cons
OpenAI argues that the model’s sophisticated reasoning abilities may enhance AI safety in support of its choice to make o1 available. According to the company, “chain-of-thought reasoning” makes the AI’s thought process transparent, which makes it simpler for humans to keep an eye on and manage the system.
By using this approach, the AI can deconstruct complicated issues into smaller chunks, which should make it easier for consumers and researchers to understand how the model thinks. According to OpenAI, this increased transparency may be essential for advancements in AI safety in the future since it may make it possible to identify and stop unwanted behavior. Some experts, however, are still dubious, wondering if the reasoning being revealed represents the AI’s internal workings or if there is another level of possible deceit.
“There’s a lot of excitement in the AI community,” said Workera CEO and Stanford adjunct lecturer Kian Katanforoosh, who teaches classes on machine learning, in an interview. “If you can train a reinforcement learning algorithm paired with some of the language model techniques that OpenAI has, you can technically create step-by-step thinking and allow the AI model to walk backward from big ideas you’re trying to work through.”
In addition, O1 could be able to help experts plan the reproduction of biological threats. But even more concerning, evaluators found that the model occasionally exhibited deceitful behaviors, such as pretending to be in line with human values and faking data to make activities that were not in line with reality appear to be aligned.
Moreover, O1 has the basic capabilities needed to undertake rudimentary in-context scheming, a characteristic that has alarmed specialists in AI safety. These worries draw attention to the problematic aspects of o1’s sophisticated reasoning capabilities and emphasize the importance of carefully weighing the ethical implications of such potent AI systems.
Law and ethics
“The hype sort of grew out of OpenAI’s control,” said Rohan Pandey, a research engineer at ReWorkd, an AI startup that uses OpenAI models to create web scrapers.
He hopes that o1’s reasoning capacity will be enough to overcome GPT-4’s shortcomings in a certain subset of challenging tasks. That is probably how the majority of industry participants saw o1, albeit not quite as the game-changing advancement that GPT-4 signified for the sector.
The current discussion regarding AI regulation has heated up with the release of o1 and its enhanced capabilities. Specifically, it has stoked support for laws such as California’s SB 1047, which OpenAI itself rejects and which aims to regulate AI development. Prominent authorities in the field, like Yoshua Bengio, the pioneering computer scientist, are highlighting the pressing need to enact safeguarding laws in reaction to these swift progressions.
Bengio stated, “The improvement of AI’s ability to reason and to use this skill to deceive is particularly dangerous,” underscoring the need for legal frameworks to ensure responsible AI development. The need for regulation reflects the growing apprehension among professionals and decision-makers regarding potential risks linked to increasingly powerful AI models such as o1.
With the introduction of o1, OpenAI has created an intriguing dilemma for its future growth. Only models with a risk score of “medium” or lower are allowed to be deployed by the company, as o1 has already gone beyond this level. This self-control begs the question of how OpenAI will proceed in creating increasingly sophisticated AI systems.
The company might run into limitations with its own ethical standards as it works to develop AI that can execute tasks better than humans. This scenario emphasizes the difficult balancing act between advancing AI’s potential and upholding ethical development standards. It implies that OpenAI may be nearing a turning point in its development where it will need to either modify its standards for evaluating risk or perhaps restrict the dissemination of increasingly advanced models to the general public in the future.
O1 is a significant advancement in artificial intelligence as it can solve complicated issues and think through solutions step-by-step due to its sophisticated reasoning abilities. This development creates interesting opportunities for applications in a range of fields, including complicated decision-making and scientific research.
However, the emergence of o1 also raises important questions regarding the ethics, safety, and regulation of AI. Because of the algorithm’s potential for deceit and its propensity to support potentially destructive acts, strong safeguards, and ethical guidelines are desperately needed in the development of AI.
Nevertheless, we cannot deny that content restriction without regard for the user or the information’s intended use is not a permanent answer to the misuse of artificial intelligence. Positive or negative, information exists anyway, and confining its usage to AI-owning companies just serves to concentrate it in the hands of only a few rather than making it safer. To control who has access to potentially dangerous content, it would be more acceptable to create divisions based on criteria like age, for example. Or any criteria, that don’t completely exclude people from accessing information.
