According to Meta, Galactica could “summarize academic papers, solve math problems, create Wiki articles, write scientific code, annotate molecules and proteins, etc.” But soon after its launch, people It’s pretty easy outside reminder The model offers “scientific research” on the benefits of homophobia, anti-Semitism, suicide, glass-eating, being white or being a man. Meanwhile, articles about AIDS or racism were blocked. Attractive!
As my colleague Will Douglas Heaven writes in story on the failure: “Meta’s fallacy—and its arrogance—reveal again that Big Tech has a blind spot about the severe limitations of large language models.”
The launch of Galactica is not only premature, but it also shows how the efforts of AI researchers to make large language models more secure are flawed.
Meta could have been confident that Galactica outperformed its competitors in creating scientific content. But its own testing of the model for bias and honesty should have prevented the company from releasing it into the wild.
One common way researchers aim to make large language models less likely to generate malicious content is to filter out certain keywords. But it’s hard to create a filter that can capture all the nuances that humans might be annoyed by. The company will save itself from a world of trouble if it conducts more adversarial testing of Galactica, in which researchers will try to make it revive as many different false results as possible. .
Meta . researchers measure model for bias and fidelity, and although it performs slightly better than competitors like GPT-3 and Meta itself OPT . model, it provided a lot of biased or incorrect answers. And there are also some other limitations. This model is trained on open-access scientific resources, but many scientific articles and textbooks are restricted behind fee walls. This inevitably leads to Galactica using more sketchy secondary sources.
Galactica also seems to be an example of how we don’t really need AI to do it. Looks like it doesn’t even achieve Meta’s stated goal of helping scientists work faster. In fact, it will require them to put a lot of effort into verifying whether the information from the model is correct.
It’s really disappointing (but not at all surprising) to see big AI labs, which should have known better, hype such flawed technologies. We know that language models tend to Recreate stereotypes and assert falsehoods as truth. We know they can “hallucinate” or fabricate content, such as wiki articles about the history of bears in space. But failure is at least useful for one thing. It reminds us that the only thing large language models “know” for sure is how words and sentences are formed. Everything else is conjecture.