Brunella Tipismana Urbano: “Does Pangram discriminate against those who don’t write in English? No. Does it discriminate against writers of colour or with disabilities? Also no. But perhaps its reliability has not been examined in literary fiction? No, it has. Again, this program isn’t infallible: Pangram performs much worse when examining short texts, and it’s easy to outsmart with a little effort. But it remains the most reliable detection tool available to us. OpenAI and Anthropic might, in a few months, improve their language models to the point of making them undetectable—but for now we do live in a world where it is possible to detect most AI-generated text. Denying this is like denying that we have technologies like airplanes, computers or vaccines. Airplanes sometimes crash, computers sometimes short-circuit, and vaccines sometimes fail to immunize a tiny sliver of the population. But these are generally reliable and necessary technologies: the same is true for Pangram.”
Pangram is very good indeed.
Terence Tao: “I think going forward, the idea of focusing on individual problems, that currency has basically been hyperinflated. We now have machines that can just brute-force, basically solve massive numbers of the old problems that we used to use as landmarks, and we have to really switch to some sort of math 2.0. I think instead of dealing with individual problems, we may start creating huge datasets of thousands or millions of problems and try to sweep them out.”
Links
Previous
Previous
Practising Failure
Next
Next