OpenAI Expands GPT-4 Vision Access Despite Ongoing Safety Questions
OpenAI opens GPT-4 with vision to all developers through the new GPT-4 Turbo API. Independent researchers praise the model's accuracy but highlight unresolved safety and privacy concerns.
OpenAI officially expands access to GPT-4 with vision, allowing all developers to integrate the image-understanding AI into their applications through the newly launched GPT-4 Turbo API. Previously, this multimodal capability remains limited to ChatGPT premium subscribers, users of the Be My Eyes accessibility app, and a select group of red teamers due to privacy and abuse concerns. Now, the company expresses confidence in its safety mitigations and prepares to roll out the technology to the wider development community in the coming weeks.
Despite this broader release, questions linger regarding the actual safety of the model. OpenAI recently publishes a whitepaper detailing the system's limitations, including problematic tendencies like discriminating against certain body types. However, critics point out that this internal research lacks the impartial perspective of independent, outside evaluators, raising doubts about how thoroughly the company addresses these significant flaws.
Fortunately, external red teamers like University of Pennsylvania professor Chris Callison-Burch and PhD student Alyssa Hwang provide some independent insight into the model's performance. After testing the system since July on tasks ranging from visual question-answering to analyzing fine art, the researchers report highly accurate and incredibly detailed results. While their early impressions highlight impressive technical capabilities, the broader tech community continues to watch closely to see if OpenAI effectively manages the remaining ethical and safety risks.