Blog

Jul 27, 2023

Down the Chatbot Rabbit Hole

Steven Levy Further generations of humans—or robots—might one day look back on

Steven Levy

Further generations of humans—or robots—might one day look back on this week as the tipping point in the way that computers and people interact. On Monday, CEO Sundar Pichai announced Google's new chatbot, dubbed Bard, based on its previously disclosed AI bot LaMDA. (It also reportedly made a $400 million investment in the large-language-model startup Anthropic.) A day later, Microsoft unveiled a new version of search engine Bing, powered by OpenAI's breakaway hit ChatGPT. In barely more time than it takes to fulfill a query, artificial-intelligence-powered systems became a critical component to search, the internet's most powerful application.

Prepare yourself for endless discussion of the implications. But I had already tumbled into that rabbit hole after pondering a less-heralded beta product soft-launched last December and opened to the public a week ago. It is a chatbot called Poe, produced by, of all companies, Quora, a 14-year-old social network that helps users find answers to questions by tapping the knowledge of other users. Like Quora itself, you type in your question and wait for the answer. But Poe, which allegedly stands for Platform for Open Exploration and is not a reference to the writer of the macabre, provides its responses using text-generation algorithms like ChatGPT and Anthropic's Claude. With no need for a human to ponder the query and respond, the answers come instantly.

This struck me as a weird pivot for a social network. But when I contacted Adam D’Angelo, Quora's cofounder and CEO, he pointed out that even when he attended high school, working on projects with classmate Mark Zuckerberg, he was aswim with the possibilities of AI. "That's what I was really excited about," says D’Angelo, who went on to join Zuckerberg's startup Facebook. When he left his CTO post there in 2009 to start Quora, using other people to answer questions was kind of a fallback because AI hadn't advanced enough to do so. "Getting AI to work at that time was really, really hard," he says. "But there was just this huge untapped potential of connecting people with other people over the internet. So instead of worrying about creating this artificial intelligence before it was ready, why not just let people access all the other intelligence that's out there?"

It turned out to be a pretty good idea. While Quora never became a juggernaut like Facebook, it has over 300 million monthly users, D’Angelo says, and in late 2021 it was widely reported that pre-pandemic the company was preparing an IPO with a possible valuation of $4 billion. Though the recent advertising downturn led Quora to lay off some workers late last month, D’Angelo says the service is getting more questions than ever, and he expects the flagging ad market to rebound.

But as a board member of OpenAI, ChatGPT's progenitor, he saw firsthand the field's dramatic advances and sensed an opportunity. By providing a front end to multiple bots, perhaps Quora could simplify access to the wellspring of AI knowledge. Their conversational responses would appear in the same vein as the human answers provided on Quora itself. So his team secured access to OpenAI's bot and Anthropic's chatbot Claude—he won't share the terms—and built Poe.

Quora's move tells us a lot about the depth of the changes AI is forcing on the world right now. In case the symbolism is lost on you, let me bop your head with it: A company whose very foundation was built upon connecting humans with each other to share knowledge is now pursuing a model where people turn not to each other, but to robots for their answers.

Jeremy White

Kate Knibbs

WIRED Staff

Stephanie McNeal

I tried Poe, typing in a query paying tribute to the recently deceased and instantly mourned guitarist and singer of the 1970s band Television. "Why was Tom Verlaine such a good guitarist?" I asked. The answer came in seconds.

I was blown away by how quick and how detailed the answer was. (I posted the same query to Quora, but after two days, no one has answered.) Poe cited Verlaine's use of "dissonant chords, unconventional chord progressions and intricate fingerpicking." It also highlighted his use of effects pedals. In my followup question asking what pedals he used, Poe quickly spit out a list, describing the effect of each.

I had gotten my answer, and it was accurate. (Poe does warn that "this bot may make inaccurate statements," a caveat that humans seldom mention about themselves.) But like many AI responses, it failed to engage my imagination. Compare this to the appreciation of Verlaine in The Guardian, which talked about how each note in the song "Spiritual" "reveals itself like a discrete shooting star, with a mysterious raga-like complexity." Or the blog post from Wilco guitarist Nels Cline that spiced his appreciation of Verlaine's playing with his personal recollections on meeting one of his inspirations.

D’Angelo readily agrees that for many queries, Poe's answers might fall short. But he also believes that the systems Poe draws on will get closer and closer to the best humans could provide. "You've got to keep in mind using these products that this is AI today," he says. "But even in three months the AI is going to get better. In six months, it's going to be better than that."

One thing that would help those answers improve is using the corpus of responses in Quora's database. But D’Angelo promises that he won't do that without the creators’ permission. "The starting point is giving people the ability to opt out," he says. "Everything else can be negotiated or built on top of that."

Even getting permission to tap Quora's organic archive, of course, won't help AI systems answer every question. But he speculates that when a future chatbot is stumped, the robot itself might do what Quora users have been doing for over a decade—go to a human for an answer. "I don't expect to ever get to a point where AI knows everything. We may get to a point where you go to AI, and then AI will, in turn, go to some people to better understand or help address your question."

All of this turns the idea of a social network on its head. There is a social element to Poe: People can choose to share their exchanges with the AI systems, in kind of a mass training exercise in learning how to best prompt these machines to provide good answers or even provocative discussions. It's a step toward our new mode of ongoing collaboration with AI models, which is the antithesis of social networking. I can only imagine the discussions at Meta, D’Angelo's onetime employer, which has invested billions in AI. The basis of the original Facebook was to keep in contact with the actual people in your network. No matter how good AI gets at generating human-esque content, it can't be your mother or your college roommate.

Jeremy White

Kate Knibbs

WIRED Staff

Stephanie McNeal

I ask D’Angelo whether creating Poe might actually contribute to the demise of Quora's original mode. Right now Quora is the leader in connecting people with questions with those who deliver great answers. Poe will be only one of a zillion contenders to provide a gateway to big AI systems. "I hope we can get to be a leader in this too," he says. "It's an early market, and I hope we can help people navigate that with an easy-to-use product."

I suspect that D’Angelo, who is a very smart tech founder, isn't saying the obvious. Generative AI is one of those hairy humongous super-shifts that will overwhelm those who ignore it. There once might have been a world where the most expedient answers came from social networks. But as quoth the raven: nevermore.

Adam D’Angelo isn't fibbing when he says that AI was an early interest for him. Over 20 years ago, as a high school student at the boarding school Exeter, he created a music product with a form of AI built in. His friend, classmate, and coinventor was Mark Zuckerberg. I wrote about the project, a music app called Synapse, in my book, Facebook: The Inside Story.

Exeter held a reception in New York City for incoming students. Zuckerberg found himself chatting with another rising junior, a gangly kid with a similarly low-key demeanor whose name was Adam D’Angelo. Like Zuckerberg, D’Angelo was a suburbanite transferring to the tony boarding school after topping out at his public high school. They had something else in common. When Zuckerberg asked D’Angelo what he was interested in, the answer was one golden word: programming. Zuckerberg was thrilled—none of his public high school friends shared his passion for building things on the computer and now the first person he met at Exeter was a lot like him …

Jeremy White

Kate Knibbs

WIRED Staff

Stephanie McNeal

Exeter students are required to create a senior project before graduation, and Zuckerberg was casting around for one, listening to tunes on his computer, when the playlist he had set up went silent after the final song played. He recruited D’Angelo to partner with him in creating a personalized virtual DJ they called Synapse. Both were big fans of an online music player called WinAmp, and they decided Synapse would ape WinAmp's functions while providing a personalized playlist.

Though both Zuckerberg and D’Angelo were utter novices in artificial intelligence, they boasted about the AI in SynapseAI, even calling the code that determined the playlist "the brain." Synapse would suggest songs to you based on what you had listened to before. D’Angelo, the more accomplished programmer, focused on building the brain, while Zuckerberg created the front end … The pair presented Synapse as their senior project, to kudos from their instructors, who were especially impressed with D’Angelo's AI component.

Zuckerberg [had] a lot of hopes on Synapse, which he no longer saw as just a class project but something that might catch on in the outside world. D’Angelo would have been fine leaving it as a class project, preferring to concentrate on his studies at the college he had chosen, the California Institute of Technology. "Caltech is, like, hard—you have to do work," says D’Angelo. "Harvard, honestly, it's not that much work. So I think he had a lot more time."

Todd asks, "Will post-quantum cryptography make Y2Q become the new Y2K?"

Thanks for the question, Todd. Before I get to the answer, however, I must upbraid you. In this newsletter only one person gets to deploy the wordplay, and that is me. I do accept that your question is acceptable shorthand for asking whether quantum computers, which are slowly becoming more than interesting lab demos, will make it trivially easy to break the cryptographic systems we use to protect our privacy and our transactions. One of the most popular encryption methods, RSA, uses a mathematical technique called factoring to protect its keys, and a well-designed attack on a powerful quantum-based system would drastically reduce the time it took to break the code.

Jeremy White

Kate Knibbs

WIRED Staff

Stephanie McNeal

There's a lot of debate about whether such computers will come in the next few years or even decades. But when it comes to protecting secrets, it's a bad idea to bet that a plausible breakthrough won't happen. Cryptography is about assuming the worst and protecting against it. That's why the US National Institute of Standards and Technology has, since 2016, been soliciting candidates for new, quantum-resistant cryptographic standards. This, and other efforts in that direction, ideally will make Y2Q, as you call it, as much of a non-event as Y2K.

You can submit questions to [email protected]. Write ASK LEVY in the subject line.

The NFL just televised its best athletes playing touch football. Next year, ditch the stadium and stage the game in the street, pausing for traffic. (Fortunately, tackling will be permitted in the Super Bowl. Go birds!)

Hands-on with the ChatGPT-enhanced version of Microsoft's search engine Bing.

Hands-on with a bunch of vibrators. Find the best one for you!

Jeremy White

Kate Knibbs

WIRED Staff

Stephanie McNeal

Inside the bone-breaking, eye-shattering, brain-concussing world of "non-lethal" crowd control tech.

We’re seeing a little progress on right-to-repair, but companies are still fighting laws that would lock in that right. Wondering if I can fix that vibrator without voiding the warranty.

If you buy something using links in our stories, we may earn a commission. This helps support our journalism. Learn more.