OpenAI is on the verge of introducing significant upgrades to ChatGPT. It is reported that the ChatGPT Pro, priced at $200 per month, has officially launched. Although no new features or models have been introduced yet, users can expect to experience the new "Strawberry" model within the next two weeks. This highly anticipated AI model will not only provide robust technical support for ChatGPT but will also draw industry attention with its unique reasoning capabilities.
The core feature of the "Strawberry" model lies in its ability to mimic human thought processes. It can take 10 to 20 seconds to "think" and retrieve information before providing an answer, aiming to enhance the efficiency of AI computing power utilization and generate more accurate content.
However, it is important to note that the "Strawberry" model initially only supports text processing and does not have the image and audio processing capabilities of GPT-4.
In fact, this technology of extending reasoning time and adaptive adjustment is not pioneered by OpenAI. Google DeepMind has already conducted in-depth research in this field and published relevant papers. Researchers have found that through "computation at test time" technology, the performance of large language models can be significantly improved, breaking through the limitations of current models in training datasets and inference computing resources.
The optimization strategies for the "Strawberry" model are mainly divided into two types. The first is based on a dense, process-oriented validation reward model, which requires the model to not only output results but also provide logical reasoning processes, particularly suitable for complex mathematical and logical reasoning tasks.
The second strategy dynamically adjusts subsequent responses based on previously generated content, continuously optimizing output quality through multiple rounds of iteration. The "computationally optimal" strategy proposed by researchers aims to select the most appropriate test-time computation method based on specific circumstances, significantly enhancing computational efficiency.
However, the "Strawberry" model also faces some challenges. Although it performs excellently in reducing errors and hallucinations, the 10 to 20-second response time may affect user experience. Some testers have feedback that these slightly more accurate answers do not seem to compensate for the longer waiting time.
Additionally, the advanced capabilities may lead to higher computational resource consumption, and the cost of use may also increase. To balance user experience and resource consumption, OpenAI may set a limit on message sending frequency and consider launching higher-priced packages to provide faster response speeds.