'Salad shooter': Why two Google tech luminaries don't think AI chatbots are ready for prime time

‘Salad shooter’: Why two Google tech luminaries don’t think AI chatbots are ready for prime time

Posted on

Less than a week after Google LLC stumbled when it introduced its experimental artificial intelligence chatbot Bard, two Google luminaries today explained why they think AI chatbots — including OpenAI LLC’s wildly popular ChatGPT — aren’t really ready for prime time.

Vint Cerf, Google’s chief internet evangelist, and John Hennessy (pictured), chairman of Google parent Alphabet Inc. both spoke at VC firm Celesta Capital’s invite-only TechSurge Global Deep Tech Summit conference today at the Computer History Museum in Mountain View, California. Although they had varying views on the current and future utility of AI chatbots and the generative AI that powers them, it’s clear they think a lot more work needs to be done for them to be used both safely and efficiently.

Now, you might say, “Well, they’re both with Google, so why wouldn’t they express doubts about OpenAI’s and Microsoft’s work?” And there may be something to that. Bard’s reception had to sting, given Google’s long corporate leadership in AI.

But you really can’t dismiss either Cerf, widely known as the Father of the Internet, or Hennessy, the former Stanford University president and co-founder of MIPS Computer Systems Inc., whose RISC chip technology now underlies virtually all processors. They’ve been around the block.

Cerf (below) was especially critical of ChatGPT, providing some good reasons for his skepticism. He cited an example of asking ChatGPT for his own bio, which “got a bunch of things wrong.” That’s because it works by generating plausible next words and phrases rather than applying a deep understanding of the facts.

“It’s like a salad shooter,” he said. “It mixes [facts] together because it doesn’t know any better.”

Worse, he said, the fact that ChatGPT is supposedly based on factual material doesn’t matter. “You can’t tell the difference between an elegantly expressed lie and a fact,” he said. “We are a longer way away from the awareness we want. The best applications right now are writing fiction.”

Cerf was asked if there are any better applications for generative AI. “Snake oil,” he replied quickly. “There’s an ethical issue here. Engineers like me should be responsible for trying to find a way to tame some of these technologies so they’re less likely to cause trouble.”

That’s not to say there won’t be any useful applications, he added. The key, he believes, is to “find things to do where these tools work well,” such as the AlphaFold protein structure database created by Alphabet’s Deep Mind AI unit.

One way to get to useful services such as precision medicine, he said, is to figure out how to create AI that has context to understand things in real space, like a two-year-old who can understand the function of a table and also understand that function works on a chair seat or a lap as well.

Indeed, other speakers at the TechSurge conference suggested paths to doing that, in particular training AI models on very specific sets of data. “Every company wants their own ChatGPT model but trained on their own proprietary model,” said Kunle Olukotun, professor of electrical engineering and computer science at Stanford. Olukotun thinks that may be facilitated more by open models than by proprietary ones such as OpenAI’s.

For instance, Northwestern Mutual is already using AI models in underwriting, said Christian Mitchell, executive vice president and chief customer officer for the insurance giant. “We can almost make a decision on underwriting even more you apply,” if provided the right data, he said. The firm is even mulling a generative AI that could listen to an avatar in a metaverse setting and advise what services might save a person money, he added, somewhat ominously. “It’s a total game changer for our business,” he said.

For his part, Hennessy sounded a somewhat more optimistic note about AI, saying he admires the quality of its natural language responses. “It’s awakened in everybody a sense that maybe the singularity” — in this context, the time when artificial intelligence exceeds human intelligence — “is closer than we thought,” he said. More specifically, instead of current predictions that could occur in 4o to 50 years, he thinks it could come 10 to 20 years sooner.

Closer to today, he thinks useful business intelligence applications using generative AI could come as soon as a year or two — though he said it’s not truly ready yet for mainstream applications, at least without more guardrails.

That said, Hennessy thinks there’s a lot more work to do for AI models to be both useful and safe, because currently the chief flaw is that ChatGPT, for one, is “confident even when it’s wrong.”

And that’s one reason Google has been hesitant to turn its AI technology into a similar product, Hennessy noted. “You don’t want to put a system out that either says wrong things or sometimes says toxic things,” he said. “We in the tech industry have to be a little more careful about the situation we create out in civil society.”

Paradoxically, perhaps, making AI better may take a lot of human help. “We’re going to have to rethink how we’re going to train them,” he said, perhaps not by training them on the entire web but on a more curated subset, even more than OpenAI has already done with ChatGPT. Moreover, he said, “we’re going to have to figure out how to make these models much more resilient over time.”

If that can be accomplished, he added, generative AI services can be used effectively as an “amplifier of human ability.” Perhaps not coincidentally, that echoes the “augmentation” that Doug Englebart, who famously demonstrated the mouse, hyperlinks, graphical interfaces and other technical ideas way back in 1968, envisioned as computing’s ultimate contribution to human advancement.

Photos: Robert Hof/SiliconANGLE

Show your support for our mission by joining our Cube Club and Cube Event Community of experts. Join the community that includes Amazon Web Services and Amazon.com CEO Andy Jassy, Dell Technologies founder and CEO Michael Dell, Intel CEO Pat Gelsinger and many more luminaries and experts.

Source link

Leave a Reply

Your email address will not be published. Required fields are marked *