According to a new analysis by the research institution Epoch AI, rapid advancements in reasoning-based artificial intelligence models appear poised for a significant slowdown in the near future, possibly within just one year.
These models, spearheaded by OpenAI's "o3" model, have demonstrated remarkable results in recent AI benchmarks, particularly in tasks related to mathematics and programming.
The key differentiator in their performance lies in their ability to consume greater computational power to solve complex problems.
However, this superiority is counterbalanced by a noticeable slowness in task execution compared to traditional models.
Development Mechanism of Reasoning-Based Models
Typically, these types of models are built in two stages.
The first stage begins with training a conventional model on vast amounts of data.
Subsequently, the second stage employs what is known as "reinforcement learning" to provide the model with feedback, helping it improve its performance when tackling difficult problems.
Until recently, significant computational power had not been extensively utilized during this second phase.
Nevertheless, the situation is beginning to change, as OpenAI has confirmed using computational power equivalent to ten times what was used in training the previous "o1" model. It is presumed that the majority of this increase was directed towards the reinforcement learning phase.
In this context, OpenAI researcher Dan Roberts revealed that the company's future plans prioritize reinforcement learning; indeed, its computational power usage might even surpass that used during the initial model training.
Despite this push towards enhancing model performance through greater computational power consumption, Epoch AI's analysis indicates the existence of limits that cannot be easily overcome.
These limitations are not solely technical but also encompass the exorbitant costs associated with research and development.
Josh You, an analyst at the institution and author of the report, clarifies that while the performance of traditional models doubles annually, the improvement resulting from reinforcement learning multiplies at a faster rate, up to ten times every three to five months.
Regardless of this, the rate of improvement in reasoning models is expected to converge with the general trajectory of AI models by 2026.
Economic and Operational Hurdles Threaten Continued Progress
On another note, high operational costs are considered one of the most prominent challenges.
The continued need for substantial investments in research could pose an obstacle to the future expansion of these models.
Consequently, if sufficient resources are not available to support experimentation and refinement stages, sustaining this progressive path might become difficult.
You added that rapid growth in computational capacity is a fundamental element in improving the performance of reasoning models, making it essential to closely monitor this factor in the upcoming period.
Growing Concern Within the Industry
The report also suggests that the potential slowdown could stir widespread concern within the AI industry, especially given the massive investments already poured into this field.
Furthermore, previous research has shown that these models, despite their high capabilities, exhibit a greater tendency towards "hallucination" or producing inaccurate answers than some traditional models.
In conclusion, the analysis from Epoch AI does not negate the significant progress achieved by reasoning models. Instead, it highlights fundamental challenges that could limit their ability to continue at the same pace.
This situation may present the AI industry with a genuine test: how can innovation be maintained while controlling costs, without sacrificing model quality?
