The boundaries of artificial intelligence (AI) are continually being tested and expanded. A new AI model, known as 03, has shown significant improvements over its predecessor but at a substantial financial cost. The model employs a technique called Test-Time Compute, which allows it to take more time to generate a range of solutions, thus providing better answers to complex questions. However, 03 does not yet qualify as a general artificial intelligence (AGI).
Despite the high costs, the method appears to be effective. In an ARC-AGI benchmark test, which compares artificial intelligence with human intelligence, the model performed surprisingly well. By the end of December 2024, François Chollet, the developer of the test, announced on Twitter that the model achieved 87.5 percent in the benchmark. This score is nearly three times higher than the previous model, o1, which scored 32 percent.
The improvement comes at a significant price. Chollet noted that at full capacity, the system could cost “thousands of dollars per task.” The less powerful version of o3 costs around 20 US dollars per task and is slightly less accurate at 75.7 percent.
The scaling of AI performance has sparked discussions within the industry about whether improved AI models will eventually hit a limit due to their increasing demand for computational power and training data. This test seems to challenge that notion on a technical level.
However, as the cost per task continues to rise, questions arise about who can afford such improvements and whether a financial ceiling will eventually be reached. Even the less accurate version incurs relatively high costs at 20 dollars per task, especially when compared to the predecessor model or the monthly 20 dollars (19.34 euros) for ChatGPT Plus.
A mini-version of o3 is planned for release in January 2025. Currently, o3 is not available to the public, but researchers can apply for early access.
The continuous evolution of AI technology presents both opportunities and challenges. The advancements in AI models like o3 demonstrate the potential for improved performance and capabilities. However, these advancements also bring to light issues related to cost and accessibility, which could limit the widespread adoption and application of such technologies.
As AI technology progresses, it will be crucial to address these challenges to ensure that the benefits of AI can be realized by a broader audience. This involves finding ways to make AI more cost-effective and accessible while maintaining high performance standards.
The future of AI holds promise, but it also requires careful consideration of the economic and ethical implications. As AI models continue to advance, it is essential to balance innovation with practicality to maximize the positive impact of AI on society.