OpenAI Launches GPT-4 With New Capabilities But Keeps Architecture Secret
OpenAI releases GPT-4, a more capable multimodal AI model, but breaks from past transparency by hiding its technical details due to safety and competitive concerns.
OpenAI officially launches GPT-4, its next-generation large language model that shows significant improvements over the previous GPT-3.5 system. The new model features multimodal capabilities that allow it to process both text and image inputs, though it still only produces text as output. Additionally, GPT-4 scores higher on academic and professional benchmarks, maintains better guardrails against disallowed content, and successfully processes much longer blocks of text.
Despite detailing these impressive capabilities, OpenAI keeps the underlying architecture and training methods a closely guarded secret. The company only reveals that GPT-4 relies on a pre-trained Transformer-style model fine-tuned using Reinforcement Learning from Human Feedback. This marked departure from the detailed transparency of the GPT-3 release stems from a highly competitive landscape and the safety implications associated with massive AI models.
OpenAI co-founder and Chief Scientist Ilya Sutskever admits that the company's previous approach to sharing research is "flat out" wrong, yet the organization readily makes GPT-4 available to the public. Developers access the system through the company's API, while everyday users utilize it via paid ChatGPT subscriptions. Experts note that this shift forces researchers and policymakers to seriously balance the need for equitable access with the critical necessity of preventing the proliferation of potentially dangerous systems.