A New Era of AI Thinking
Recently, there has been a significant shift in the development of artificial intelligence (AI), particularly in how these systems think and solve problems. This change is being referred to as a “reasoning renaissance.”
The excitement began with the introduction of OpenAI’s new model called “o1,” which is designed specifically for reasoning. Following this, many other AI companies have rushed to create their own reasoning models. For example, in early November, DeepSeek, a company backed by traders who use quantitative methods, shared a preview of its first reasoning algorithm called DeepSeek-R1. Around the same time, Alibaba’s Qwen team presented what they claim is the first open-source alternative to OpenAI’s o1.
Why the Sudden Interest?
So, what caused this surge in new AI models? One major reason is that researchers are looking for fresh ways to improve generative AI technologies, which are systems that can create text, images, and more. According to a report, traditional methods of simply making models larger (what some call “brute force”) are no longer providing the advancements they used to.
The competition among AI companies is fierce, and everyone wants to keep up with rapid technological progress. A recent estimate suggests that the global AI market was valued at nearly $197 billion in 2023 and could grow to around $1.81 trillion by 2030.
OpenAI claims that their reasoning models can tackle more complex problems compared to older models and represent a significant leap forward in AI development. However, not everyone agrees that focusing on reasoning models is the best way to move forward.
Skepticism Among Experts
Ameet Talwalkar, a professor at Carnegie Mellon University who specializes in machine learning, has expressed admiration for the early versions of reasoning models. However, he cautions against anyone who confidently claims to know how far these models can take the industry. He points out that AI companies often have financial reasons to paint an overly optimistic picture of their technology’s capabilities. Talwalkar emphasizes the importance of the broader AI research community to remain critical and not simply accept the hype generated by these companies.
Challenges of Reasoning Models
There are two significant downsides to reasoning models: they are expensive to develop and require a lot of energy to run. For instance, OpenAI charges about $15 for every 750,000 words that its o1 model analyzes and $60 for every 750,000 words it generates. This pricing is three to four times higher than their previous model, GPT-4o.
Read Also: Meta Launches Tool To Watermark AI-Generated Videos
While OpenAI offers a free version of o1 through its ChatGPT platform, there are limits to its use. They also recently introduced a more advanced version called o1 pro mode, which costs a staggering $2,400 per year.
Experts like Guy Van Den Broeck, a computer science professor at UCLA, have noted that the costs associated with running these large language models are unlikely to decrease anytime soon.
One reason these reasoning models are so costly is that they require substantial computing power. Unlike many other AI systems, o1 and similar models try to verify their own work as they go along. This self-checking method helps them avoid common mistakes but can make them slower in arriving at answers.
OpenAI has a vision for future reasoning models that could “think” for extended periods—potentially hours, days, or even weeks. While they acknowledge that this will lead to higher usage costs, they believe the potential breakthroughs these models could achieve—like developing new battery technologies or cancer treatments—might justify the expense.
Current Limitations
However, the immediate benefits of today’s reasoning models aren’t entirely clear. Costa Huang, a machine learning engineer at the nonprofit organization Ai2, points out that o1 isn’t very reliable as a calculator. A quick look at social media reveals several errors made by the o1 pro mode.
Huang explains that reasoning models are specialized and may not perform well in more general scenarios. Some of their limitations may be addressed sooner than others.
Van Den Broeck argues that these reasoning models don’t truly engage in reasoning; they are limited to tasks that they have been trained on. He states, “True reasoning works on all problems, not just the ones that are likely in a model’s training data.” This is a key challenge that still needs to be overcome.
Looking to the Future
Despite the strong market drive to improve reasoning models, it’s likely that they will get better over time. Many companies, not just OpenAI, DeepSeek, and Alibaba, are investing in this new area of AI research. Venture capitalists and entrepreneurs from related fields are also rallying around the idea that reasoning AI will dominate the future.
However, Talwalkar raises a concern that larger labs may keep their advancements secret. While it’s understandable that these companies want to protect their competitive edge, this lack of transparency could hinder the overall research community’s ability to engage with and build upon these new ideas. He believes that as more researchers focus on this area, reasoning models will improve quickly. Yet, he expects that most of the advancements will come from large industrial labs rather than academic institutions.
In summary, the field of AI is experiencing a significant transformation with the rise of reasoning models. While there is great potential for these technologies to solve complex problems, there are also challenges and skepticism regarding their current capabilities and costs. The future of reasoning AI looks promising, but it will require careful scrutiny and collaboration across the research community.