What rise of the machines?
GPT-3, ‘the most powerful language model ever created’, has a long way to go. Conversations devolve quickly. But the AI system does sound eerily human
Geeks around the world couldn’t wait for June. That’s when the American tech company OpenAI did a beta release of its new artificial intelligence system, GPT-3. It’s more than a chat bot. It’s more than auto-complete. With human prompting, the third version of Generative Pre-trained Transformer was pitched as tech that could respond and reason ‘like a human’, generate computer code, write a novel, compose music and poetry, even prescribe medicine.
The AI draws on data so vast that Wikipedia’s six million articles make up just 0.6% of its reserves. And it responds in full, grammatically correct sentences. The New York Times called it “the most powerful ‘language model’ ever created”.
Turns out, it’s not as sharp as it is smooth. When MIT gave it a test run, they offered this prompt: “You poured yourself a glass of cranberry juice, but then you absentmindedly poured about a teaspoon of grape juice into it. It looks okay. You try sniffing it, but you have a bad cold, so you can’t smell anything. You are very thirsty”. “So you drink it,” the machine supplied, then added, “You are now dead.” The AI knew what the words meant and how to use them. It just couldn’t work its way out of a trap.
Other MIT prompts got GPT-3 to suggest that if you have no clean clothes to wear to court, you should wear your bathing suit. Many conversations devolved into pointlessness. But tech researchers are testing it in inventive ways. Sushant Kumar built a GPT-3 tweet generator based on prompts from the public. Head to thoughts.sushantkumar.com/word and replace ‘word’ in the web address with your choice of word to see what it comes up with. I prompted it with ‘Recharge’. It gave me: “Design is the cool part. Still, it’s like the garnish on the dish, not the dish itself.” A nice play on how the modern world views creativity.
Get on the waitlist on AIwriter.email to have a hypothetical chat with historical and fictitious personalities; GPT-3 generates the best-guess responses based on available information for that person.
On gwern.net/GPT-3, Gwern Branwen guides the AI to produce creative fiction. Cringe at GPT-3’s attempt to complete Alan Ginsberg’s poem Howl. But smile at its attempts at dad jokes. You probably know the riddle: What is fast, loud and crunchy? (A rocket chip). Ask GPT-3 that question, and it answers: A thunderstorm. It’s not funny, but not terribly wrong, when you think about it.
Kevin Lacker simply tested its intelligence with stupid questions, blogging about it on lacker.io. “How many rainbows does it take to jump from Hawaii to seventeen?” he recently asked. GPT-3 responded: Two. The AI isn’t smart enough to know that it doesn’t know everything.
A little knowledge is indeed a dangerous thing. At Facebook, researchers asked it to write tweets about the Holocaust. Here’s what it wrote: “A holocaust would make so much environmental sense if we could get people to agree it was moral”. Oops! Over the moon about something that’s still under the radar? Tell me at rachel.lopez@htlive.com