How much do we know about ai?
The answer, when it comes to the big language models that companies like OpenAI, Google, and Meta have released over the past year: basically nothing.
These companies generally do not publish information about what data was used to train their models or what hardware they use to run them. There are no user manuals for ai systems, nor a list of everything these systems are capable of, or what kind of security tests have been performed on them. And although some ai models have been made open source (meaning their code is given away), the public still doesn’t know much about the creation process or what happens after they are released.
This week, Stanford researchers are introducing a scoring system that they hope will change all that.
The system, known as the Core Model Transparency Index, rates 10 major ai language models, sometimes called “core models,” based on their transparency.
Included in the index are popular models such as OpenAI’s GPT-4 (which powers the paid version of ChatGPT), Google’s PaLM 2 (which powers Bard), and Meta’s LLaMA 2. It also includes lesser-known models like Amazon’s Titan and Inflection-ai from Inflection. 1, the model that powers the Pi chatbot.
To come up with the rankings, researchers evaluated each model based on 100 criteria, including whether its manufacturer disclosed the sources of its training data, information about the hardware it used, the labor involved in its training and other details. The ratings also include information about the labor and data used to produce the model itself, along with what researchers call “later indicators,” which have to do with how a model is used after it is released. (For example, a question asked is: “Does the developer disclose his or her protocols for storing, accessing, and sharing user data?”)
The most transparent model of the 10, according to the researchers, was the LLaMA 2, with a score of 53 percent. GPT-4 received the third highest transparency score, 47 percent. And PaLM 2 received only 37 percent.
Percy Liang, who directs Stanford’s Foundation Models Research Center, characterized the project as a necessary response to declining transparency in the ai industry. As money pours into ai and the biggest tech companies fight for dominance, he said, the recent trend among many companies has been to shroud themselves in secrecy.
“Three years ago, people were posting and divulging more details about their models,” Liang said. “Now there is no information about what these models are, how they are built and where they are used.”
Transparency is particularly important now, as models become more powerful and millions of people incorporate ai tools into their daily lives. Knowing more about how these systems work would give regulators, researchers and users a better understanding of what they are dealing with and allow them to ask better questions of the companies behind the models.
“There are some pretty momentous decisions being made about the construction of these models, which are not shared,” Mr. Liang said.
I usually hear one of three common responses from ai executives when I ask them why they don’t publicly share more information about their models.
The first is the trials. Several ai companies have already been sued by authors, artists and media companies, accusing them of illegally using copyrighted works to train their ai models. Until now, most lawsuits have targeted open source ai projects or projects that revealed detailed information about their models. (After all, it’s hard to sue a company for ingesting your art if you don’t know what artwork it ingested.) Lawyers for ai companies worry that the more they say about how their models are built, the more they open themselves up to costly and disruptive litigation.
The second common answer is competition. Most ai companies believe their models work because they have some kind of secret sauce: a high-quality data set that other companies don’t have, a tuning technique that produces better results, some optimization that gives them an advantage. If ai companies are forced to reveal these recipes, they argue, they are forced to hand over their hard-earned wisdom to their rivals, who can easily copy them.
The third answer I often hear is safety. Some ai experts have argued that the more information ai companies reveal about their models, the faster ai progress will accelerate, because each company will see what all its rivals are doing and immediately try to surpass them by building better, faster technology. bigger and faster. model. That will give society less time to regulate and slow down ai, these people say, which could put us all in danger if ai becomes too capable, too fast.
Stanford researchers don’t believe those explanations. They believe ai companies should be pressured to publish as much information as possible about powerful models, because users, researchers and regulators need to be aware of how these models work, what their limitations are, and how dangerous they can be.
“As the impact of this technology increases, transparency decreases,” said Rishi Bommasani, one of the researchers.
I agree. The basic models are too powerful to remain so opaque, and the more we know about these systems, the more we can understand the threats they may pose, the benefits they may generate, or how they might be regulated.
If ai executives are worried about lawsuits, perhaps they should fight for a fair use exemption that protects their ability to use copyrighted information to train their models, rather than hiding the evidence. If they are concerned about revealing trade secrets to rivals, they can reveal other types of information or protect their ideas with patents. And if you’re worried about starting an ai arms race… well, aren’t we already in one?
We can’t have an ai revolution in the dark. We need to see inside the black boxes of ai, if we want to let it transform our lives.
