Sam Altman, the CEO of OpenAI, is a self-professed techno-optimist and accelerationist. He believes that Artificial Intelligence (AI) technologies will usher in a bright future for humanity and that we should do whatever we can to bring about this bright future as soon as possible. We should not put any obstacles in the way of AI progress (https://www.nytimes.com/2023/11/20/podcasts/hard-fork-sam-altman-transcript.html). Ironically, by being opaque and unforthcoming about its AI models, it is OpenAI itself that is curtailing progress in AI research.
Progress in scientific research is made possible by the sharing of methods, data, and results within the scientific community. If every researcher had to reinvent the wheel, scientific progress would grind to a near halt. For, without the sharing of methods, data, and results, there can be no peer review. Peer review is the process by which scientists evaluate the research done by other scientists to determine whether it makes a worthwhile contribution to the growing body of scientific knowledge. In other words, transparency is critical to establishing scientific results that endure the test of peer review in the scientific community. That is what makes science successful. Scientific progress is made possible by openly sharing methods, data, and results in the scientific community where rigorous peer review can occur.
Now, when OpenAI released its latest foundation model, GPT-4, it provided what seemed like a scientific article about the AI model (https://cdn.openai.com/papers/gpt-4.pdf). On closer inspection, however, it becomes clear that the document says nothing about the methods, procedures, and data used to build and train the AI model. In fact, the report states, “Given both the competitive landscape and the safety implications of large-scale models like GPT-4, this report contains no further details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar” (https://cdn.openai.com/papers/gpt-4.pdf).
An article containing such a disclaimer instead of detailed descriptions of methods, data, and results would never get accepted for publication in a scientific journal. Unlike OpenAI’s report on GPT-4, every report of scientific results published in a scientific journal has a methodology section in which experimental designs, data collection procedures, and data analysis methods are described in detail (https://plos.org/resource/how-to-write-your-methods/).
OpenAI is notoriously opaque about its AI models. In fact, Stanford University’s Center for Research on Foundation Models (CRFM) put together a transparency index for AI foundation models, which evaluates one hundred different aspects of transparency, including how a tech company builds its models, how the models work, and how they are deployed. On its GPT-4 model, OpenAI scored 48%, which is “unimpressive” (https://hai.stanford.edu/news/introducing-foundation-model-transparency-index).
Opaqueness and secrecy are the characteristics of pseudoscience, not science. Distinguishing between science and pseudoscience is a difficult philosophical question. But there are clear examples of both. The philosopher of science, Karl Popper, identified astrology as a clear example of pseudoscience (https://www.npr.org/sections/13.7/2017/05/08/527354190/what-is-pseudoscience). On the other hand, astronomy is a clear example of science. While astronomers must explain how they obtain their results, astrologers simply pronounce their daily horoscopes without having to explain how they came up with them. That is why, according to the philosopher of science, Thomas Kuhn, the science of astronomy is being revised continuously in light of new observational and experimental evidence, which is examined and reviewed by other astronomers, whereas astrology has remained pretty much unchanged. In other words, unlike astrology, astronomy makes progress as a science.
Secrecy and opaqueness allow OpenAI to make extraordinary claims about its AI models with impunity and no accountability. For example, its GPT-4 report states that “new applications built on top of generative models [like GPT-4] will often solve more complex tasks than the model on its own” (https://cdn.openai.com/papers/gpt-4.pdf). Such extraordinary claims cannot be subjected to peer review because OpenAI does not share methods and data with the community of AI researchers. Without a proper evaluation of such claims through peer review, OpenAI’s talk of AI accelerating the pace of technological progress is mere speculation with no evidence to support it.
OpenAI hires data scientists and describes itself as “an AI research and deployment company” (https://openai.com/careers/data-scientist-product). If it wants to claim the privilege of doing “research” in the scientific sense of the word, OpenAI should also accept the responsibilities that come with that, chief among them is the responsibility to openly share methods, data, and results with other researchers. If it is sincere about its mission to benefit all of humanity through progress in AI research (https://openai.com/about), OpenAI cannot act like astrologers by keeping its methods and data secret from the rest of the AI community. If it wants to claim the honor of doing science, OpenAI should play by the rules of science. But if all that matters to Altman and OpenAI is profiting from its AI models, then all the optimistic talk about a bright future ushered by progress in AI research rings hollow.