Excellent Performance in SAT and Bar Exam
MS Plans to Integrate into Bing Search Engine
Limitations in Distinguishing False Information
OpenAI, the developer of the artificial intelligence (AI) chatbot 'ChatGPT,' released 'GPT-4' on the 14th (local time), an AI tool with performance closer to human levels than previous models.
According to major foreign media, OpenAI unveiled GPT-4, a large-scale AI language model (LLM), on the same day.
OpenAI emphasized that GPT-4 demonstrated human-level performance, scoring in the 90th percentile on the U.S. bar exam simulation and 89% on the SAT college entrance exam. The AI chatbot 'ChatGPT,' which applied the previous language model GPT-3.5, had scored in the 10th percentile on the bar exam simulation.
Although OpenAI did not disclose specific numbers, experts estimate that GPT-4 has around 100 trillion parameters, which correspond to synapses in the human brain. Compared to the 175 billion parameters applied in GPT-3.5, this represents a significant performance improvement. OpenAI stated that, considering safety and competition issues, it will not disclose technical details such as the hardware used for the language model or the training data for the model.
However, OpenAI explained that it used Microsoft's (MS) cloud service Azure to train the GPT-4 model. MS plans to integrate GPT-4 into its new search engine Bing.
Nonetheless, GPT-4 was also confirmed to have limitations in distinguishing false information, similar to previous models. OpenAI said it conducted stress tests before the GPT-4 release to identify such issues. The ChatGPT equipped with GPT-3.5 was also criticized for potentially spreading fake information by presenting falsehoods as facts.
They expressed concerns, stating, "GPT-4 can generate harmful content such as biased hate speech and provide information necessary for illegal activities like chemical weapon development."
Regarding whether GPT-4 can act with free will without human intervention, OpenAI stated, "We tested whether AI can operate autonomously, but it appears that it cannot do so yet."
© The Asia Business Daily(www.asiae.co.kr). All rights reserved.
