Media Roundup: GPT-4’s Big Debut

(Ebru-Omer/Shutterstock)

Just two years ago, nobody outside of a small group of AI researchers and practitioners knew what a “large language model” was. But thanks to the launch of ChatGPT in November, everybody and her cat is an LLM expert now. So when OpenAI finally took the veil off GPT-4 this week, you just knew it was going to be a big deal.

The media response to the official unveiling of GPT-4 this week has delivered on the hype, just as advertised. If GPT-4 were a celebrity (and mind you, it hasn’t declared itself to be alive–at least not yet), it might be bigger than Elvis.

“It’s basically mind-blowing,” one early GPT-4 tester, Miðeind COO Linda Heimisdottir, told Sifted.eu. “It feels like the sky’s the limit.”

GPT-4 “stunned many users” in early tests, CNN reports. Whether it’s passing the bar exam at the 90th percentile, explaining a joke, or creating a functional website from a drawing, the new AI “promises to blow previous iterations out of the water, potentially changing the way we use the internet to work, play and create,” writes CNN’s Samantha Murphy Kelly.

GPT-4 can analyze up to 25 pages of text, nearly 10x the amount of its processor. It scored a 1410 on the SAT, 150 points higher than its predecessor. While GPT3.5 resembled the intelligence of a sixth-grader, GPT4 sounds more like “a smart 10th-grader,” Keith Peiris, co-founder of AI startup Tome, told the Wall Street Journal.

While GPT-4 hasn’t (yet) tried to end New York Times columnist Kevin Roose’s marriage, as the GPT-powered Bing chat mode infamously did earlier this year, the new AI still gave Roose that “dizzy and vertiginous feeling” that he experienced earlier. “The more time I spend with A.I. systems like GPT-4,” he writes, “the less I’m convinced that we know half of what’s coming.”

GPT-4 is good, but not good enough to run a nuclear power plant, one commentator said (BESTWEB/Shutterstock)

Not knowing what an LLM will say next is part of the fun, as hallucinations are a feature of all LLMs. While GPT-4 appears to completely make things up less frequently than GPT-3.5 (30% error rate) or ChatGPT (20% error rate), it continues to be a concern with the new model. “It’s going to be a long time before you want any GPT to run your nuclear power plant,” Oren Etzioni, professor emeritus at the University of Washington, told Wired.

That tendency to make things up and be creative can be harnessed to boost human performance and creativity, reports Lifewire’s Sascha Brodsky. “GPT-4 has an uncanny ability to accomplish tasks that, until recently, would have seemed like science fiction,” Brodsky writes.

That extra productivity might not be embraced by all, according to Bloomberg’s Parmy Olson, who writes that GPT-4 “could turn work into a hyperproductive hellscape.” Have 1,000 emails to read by 4 p.m.? GPT-4 to the rescue!

Several media cites and AI experts chided OpenAI for not sharing more details about how GPT-4 was trained.  “The 98-page paper introducing GPT-4 proudly declares that they’re disclosing *nothing* about the contents of their training set,” Nomic AI Vice President of Information Design Ben Schmidt tweeted.

OpenAI appears to be concerned that opening GPT-4 to the world will give competitors a leg up, but some experts are concerned that preserving secrecy gives a place for biases to fester and grow. “Researchers are excited about the AI,” writes Katharine Sanderson in a piece for Nature, “but many are frustrated that its underlying engineering is cloaked in secrecy.”

The folks at OpenAI acknowledge there’s room for improvement with its AI. “It’s not perfect,” Greg Brockman, president and co-founder of OpenAI, reporetdly said during the announcement, which was streamed live on YouTube. “But neither are you.”

Related Items:

GPT-4 Has Arrived: Here’s What to Know

ChatGPT Brings Ethical AI Questions to the Forefront

Has Microsoft’s New Bing ‘Chat Mode’ Already Gone Off the Rails?