Reasoning AI Models: A Current Trend

A Surge in Reasoning AI Development

A notable resurgence in AI reasoning capabilities is currently underway. Following the introduction of OpenAI’s o1, a model specifically designed for reasoning, numerous other AI laboratories are now actively developing and releasing their own reasoning-focused algorithms. DeepSeek, a research firm backed by quantitative analysts, previewed its DeepSeek-R1 in early November. Simultaneously, Alibaba’s Qwen team presented what they describe as the first openly accessible alternative to o1.

Drivers Behind the Innovation

What has spurred this rapid development? One key factor is the ongoing pursuit of innovative methods to enhance generative AI technologies. Recent reports indicate that simply increasing the scale of models – a “brute force” approach – is no longer delivering the same level of improvement as before.

Furthermore, there's significant competitive pressure within the AI sector to sustain the current rate of advancement. The global AI market was valued at $196.63 billion in 2023, with projections estimating a substantial rise to $1.81 trillion by 2030.

OpenAI’s Claims and Skepticism

OpenAI asserts that reasoning models are capable of addressing more complex problems than their predecessors, representing a significant leap forward in generative AI. However, this perspective isn’t universally shared.

Ameet Talwalkar, an associate professor specializing in machine learning at Carnegie Mellon, acknowledges the initial reasoning models are “quite impressive.” Nevertheless, he expressed reservations about accepting definitive claims regarding the ultimate potential of these models.

Talwalkar cautioned that AI companies may be inclined to present optimistic forecasts concerning the capabilities of future technology iterations due to financial motivations. He emphasized the importance of the broader AI research community avoiding uncritical acceptance of marketing claims and instead prioritizing verifiable outcomes.

Challenges: Cost and Power Consumption

Reasoning models present two primary challenges: they are expensive to operate and require substantial power.

For example, OpenAI’s API pricing structure charges $15 for analyzing approximately 750,000 words with o1 and $60 for generating the same amount of text. This represents a sixfold increase in cost compared to OpenAI’s GPT-4o, a model not specifically focused on reasoning.

While o1 is available within ChatGPT with certain limitations, OpenAI also offers a premium o1 tier – o1 pro mode – at a yearly cost of $2,400.

Guy Van Den Broeck, a computer science professor at UCLA, stated that the cost of utilizing large language model reasoning is not decreasing.

The Mechanics of Reasoning and Associated Costs

The high cost of reasoning models stems from their intensive computational demands. Unlike many AI systems, o1 and similar models actively verify their own work during processing. This self-checking mechanism helps mitigate common errors but also extends the time required to reach a solution.

OpenAI anticipates future reasoning models potentially “thinking” for extended periods – hours, days, or even weeks. While acknowledging increased usage costs, the company believes the potential benefits, such as breakthroughs in battery technology or cancer treatment, could justify the expense.

Current Limitations and Reliability

The practical benefits of today’s reasoning models are not yet fully apparent. Costa Huang, a researcher and machine learning engineer at Ai2, points out that o1 is not consistently accurate in performing calculations. Furthermore, social media reports indicate instances of errors within o1 pro mode.

Huang notes that these reasoning models are specialized and may exhibit subpar performance in broader, more general domains. He suggests that some limitations will be addressed more quickly than others.

Van den Broeck contends that reasoning models do not engage in genuine reasoning, thereby restricting the scope of tasks they can successfully complete. He argues that “true reasoning” should apply universally, not just to problems present in a model’s training data, and that overcoming this limitation remains a significant hurdle.

Future Outlook and Concerns

Given the strong market incentives, it is reasonable to expect that reasoning models will continue to improve. Investment in this area extends beyond OpenAI, DeepSeek, and Alibaba, with venture capitalists and founders in related fields increasingly focusing on the potential of reasoning AI.

However, Talwalkar expresses concern that major AI laboratories may restrict access to these advancements.

He explains that competitive pressures lead these labs to maintain secrecy, which hinders the research community’s ability to fully understand and contribute to these developments. While he anticipates progress as more individuals engage with this field, he believes that most models will likely be developed and offered by large industrial labs like OpenAI, given the financial incentives involved.