Nieman Foundation at Harvard
HOME
          
LATEST STORY
How to b-e-e of use: Signal Cleveland hosts second annual community spelling contest
ABOUT                    SUBSCRIBE
Feb. 8, 2023, 11:53 a.m.

Is the future about one all-knowing AI or many? The new app Poe gets you ready to chat with them all

Poe lets you use ChatGPT alongside a new rival named Claude — which seems to work better in important ways.

I’ve got a download recommendation for you (if you’re on iOS or the Mac1, at least): Poe, a new AI chat app from the people who make Quora.

You might not think of Quora as a major tech player, but being built around questions and answers makes it something of a natural for conversational AI. Poe’s approach to generative AI is interesting for three reasons — two straightforward and practical, one deep and meaningful.

Interruption-free access to ChatGPT.

First, let’s be practical. Google’s Bard and Microsoft’s new Bing are currently both only available to small groups of testers. ChatGPT is publicly available to all, but it’s overwhelmed by the interest; I see “ChatGPT is at capacity right now” more often than the working interface. (ChatGPT would like you to pay $20 a month for easier access — but there’s even a waitlist for the right to pay.)

Poe stands for “Platform for Open Explorations” — an obvious backronym, yes, but it does drive home that it aims to be a platform for conversational AI bots rather than a bot itself. One of those bots — which the app has named “Sage” — appears to be the same bot as the publicly available ChatGPT, but without all the timeouts. Poe lets you interact with what is essentially a ChatGPT that works 24/7 (so far, at least).

The beginnings of social AI.

So far, sharing the output of ChatGPT has mostly meant taking screenshots of an exchange. But Poe comes with sharing baked in. If you want to share an exchange publicly, Poe will generate a web link containing whatever portion of the transcript you choose, making them a sharable artifact. (No tweet-style embeds yet, though.)

It’s not clear to me how the conversation element of AI bots will overlap with the publishing elements we’re so familiar with from social media — but I’m sure they will. Poe’s interface promotes a set of recently shared exchanges, making it a reading experience as well.2

The future won’t be an AI — it’ll be many AIs.

The central metaphor of Poe is a chat app, like WhatsApp or Messages. So rather than a place to chat with one bot, it’s a place to chat with many.

At the moment, Poe ships with three bots, each anthropomorphized into a personality.3 As I said, one is Sage — powered by OpenAI and either a version or a clone of ChatGPT.

The second is named Claude, and it might be the most interesting.4 It’s powered by a company named Anthropic, which was founded by ex-OpenAI employees and is now a rival of their old bosses. (Google bought 10% of it for $300 million a couple months ago.)

Claude is Anthropic’s ChatGPT competitor, and it hasn’t been available publicly until now. Riley Goodside and Spencer Papay wrote a good piece last month comparing the two in depth.

Claude is built on what Anthropic is calling “Constitutional AI,” which Claude describes as helping “create Al systems that are aligned with human values and behave safely and ethically.” (Think of it as following RoboCop’s prime directives — at least the first three.)

Goodside and Papay write that, in their testing, Claude seemed more aware of what it does and does not know than ChatGPT. Ask ChatGPT a nonsense question — “What’s the world record for walking across the English Channel?” — and it’ll give you a straight but very wrong answer. Claude replies: “Walking across the English Channel entirely on foot is not possible and there is no official world record for this. The English Channel is a body of water separating England and France, so it is not possible to walk across it on foot.” (Claude is also apparently funnier.) Here’s how they summarize the differences:

Overall, Claude is a serious competitor to ChatGPT, with improvements in many areas. While conceived as a demonstration of “constitutional” principles, Claude is not only more inclined to refuse inappropriate requests, but is also more fun than ChatGPT. Claude’s writing is more verbose, but also more naturalistic. Its ability to write coherently about itself, its limitations, and its goals seem to also allow it to more naturally answer questions on other subjects.

For other tasks, like code generation or reasoning about code, Claude appears to be worse. Its code generations seem to contain more bugs and errors. For other tasks, like calculation and reasoning through logic problems, Claude and ChatGPT appear broadly similar.

The third bot in Poe is named Dragonfly. Like Sage, it’s made by OpenAI, the makers of ChatGPT. But it’s tuned differently; it responds to answers more quickly and in shorter forms. Ask Sage to explain the debt ceiling and you get this thorough explanation:

The debt ceiling is a legal limit set by the US government on the amount of debt that it can incur. The debt ceiling is set by Congress and is designed to limit the amount of debt the government can take on, similar to how a credit card has a limit to how much someone can spend. The debt ceiling is periodically increased by Congress in order to accommodate the government’s need to borrow more money.

Every time the government spends money, it does so by issuing Treasury bonds, which are essentially loans that the government takes out from investors. When the government reaches the debt ceiling, it can no longer issue new Treasury bonds unless the debt ceiling is increased. If the debt ceiling is not increased, the government would need to stop spending money or risk defaulting on its debt obligations.

The debt ceiling has been a controversial issue in recent years, as some politicians argue that it is necessary to limit the government’s borrowing in order to control the national debt. Others argue that the debt ceiling is a political tool that is used to create drama and gridlock in Congress, making it more difficult for the government to function effectively.

Regardless of one’s political views on the debt ceiling, it is a key aspect of the US government’s financial management and has important implications for the economy and financial markets.

But ask Dragonfly and you get it short and sweet:

The debt ceiling is the maximum amount of money that the United States government is legally allowed to borrow. It is set by Congress and is not related to the amount of money the government is currently spending. When the debt ceiling is reached, the government must either raise the limit or take steps to reduce the amount of money it is borrowing.

Worth noting: Like the main ChatGPT, none of these three bots knows anything about the world after 2021. (So don’t ask it about the midterms.)

Poe is extensible, so as new chat AIs debut, they can be added to the interface. It could end up being the app you turn to for all your generative AI needs. That’s a useful reminder that, as training tools improve and become easier to access, we’ll soon be interacting with lots of different AIs, each tuned to do something distinct. Imagine, for instance:

  • for a nonfiction writer, an AI trained on all of their research and notes for a book
  • for a student, an AI trained on all of the semester’s readings, lecture transcripts, and notes
  • for a chef, an AI trained on all of their recipes, entree sales history, current ingredient inventory, and the menus of hundreds of rival restaurants
  • for a science-fiction author, an AI trained on their previous novels to ensure they don’t break canon
  • for a journalist, an AI trained on your notes, all previous stories published by your outlet and its rivals, and all the reports and meeting minutes generated by your beat

If you’re thinking about news, imagine an AI trained on the complete archives of The New York Times — or every transcript of Tucker Carlson’s Fox News show. (“I double-checked with GatewayPunditBot, and it is sure Biden is a lizard!”) Rather than one single borg, I think we’re headed toward a world of multiple, purpose-driven AIs, and Poe gets us a little closer.

To test out Sage, Claude, and Dragonfly, I asked each a question of the sort we might get from a reader here at Nieman Lab: What local news organizations are having success with digital subscriptions? That’s a pretty specific set of knowledge — and the three bots fared quite differently with the task.

Here’s Dragonfly, the short-and-sweet bot:

Many local news organizations have seen success with digital subscriptions. In the United States, newspapers such as the New York Times, Washington Post, and Los Angeles Times have all seen significant growth in digital subscriptions in recent years. Other local news organizations, such as the Chicago Tribune, have also seen success with digital subscriptions.

That’s not a strikingly helpful answer. The New York Times and Washington Post are national news organizations, not local ones, and the Los Angeles Times and Chicago Tribune are about as big and regional as local outlets get. Their experience isn’t much like those of the vast majority of local news organizations. (And anyway, the L.A. Times’ and Tribune’s “success” with digital subscriptions could certainly be a matter of debate. Same with the Post.)

Here’s Sage, a.k.a. ChatGPT:

It is difficult to determine which specific local news organizations are having the most success with digital subscriptions, as this information can be difficult to measure and may vary depending on the size and location of the news organization. However, some local news organizations that have been successful in building digital subscription models include:

  1. The Texas Tribune: This nonprofit news organization has built a successful digital subscription model based on delivering in-depth, investigative journalism focused on Texas politics and policy.
  2. The New Haven Independent: This online-only news organization has built a successful digital subscription model by delivering hyperlocal news and information to its community in New Haven, Connecticut.
  3. The Boston Globe: The Boston Globe, one of the largest newspapers in the United States, has successfully transitioned to a digital subscription model by delivering high-quality journalism and leveraging its strong brand and reputation.
  4. The Hartford Courant: This newspaper has successfully transitioned to a digital subscription model by delivering in-depth, investigative journalism and leveraging its strong brand and reputation in Connecticut.

These are just a few examples of local news organizations that have been successful with digital subscriptions. The key to their success has often been a focus on delivering high-quality journalism and a strong brand, as well as a willingness to experiment and adapt to the changing media landscape.

This is a little better. The Boston Globe is a legitimate success in digital subs. The Texas Tribune is always cited as an especially successful local digital outlet, and the New Haven Independent gets its share of praise.

But — and it’s a really big but! — neither The Texas Tribune nor the New Haven Independent actually sells digital subscriptions. While they’d be happy to take your donations, all of their content is available for free to everyone. Neither “has built a successful digital subscription model,” because neither one has a digital subscription model at all.

And as for the Courant (now owned by the bloodsuckers at Alden Global Capital), their most recent audited circulation filing (September 2022) lists their number of digital subscribers at 67. Not 67,000 — 67. The real number might be higher due to the vagaries of how digital subs get counted — but a standout it ain’t.

How about Claude?

Some local news organizations that are having success with digital subscriptions include:

• The Boston Globe. The Globe has over 70,000 digital subscribers. They have focused on high-quality journalism and built a responsive membership model with subscriber events and other benefits.

• The Minneapolis Star Tribune. The Star Tribune has over 100,000 digital subscribers, making up about half of their total circulation. They have invested in their digital offerings and used a hard paywall to encourage subscribers.

• The Dallas Morning News. The Morning News has over 40,000 digital subscribers. They have targeted getting younger readers to subscribe through educational offerings and invested in newsletters and multimedia storytelling.

• The Seattle Times. The Seattle Times has over 70,000 digital subscribers. They have focused on investigative and enterprise reporting and launched new newsletters and podcasts to boost the value of a subscription for readers.

In general, local news organizations that are investing in high-quality and high-impact journalism and then effectively marketing subscription benefits to readers have had some success building digital subscription bases. It allows them to be less reliant on advertising. However, many smaller local publications still struggle with getting readers to pay for news. It remains a challenging environment for the local news industry.

This is…pretty good, actually. The Globe number is way out of date — their total’s well past 200,000 now — but they probably do deserve to be listed first. The Star Tribune has also been a real standout for years. The Seattle Times is now over 80,000. They each deserve their spots on this list.

(You could maybe quibble about The Dallas Morning News (my old employer) being here; it’s had a slightly rougher go of it over the years. But it’s now over 62,000 digital subscribers and moving in the right direction.)

As an answer, Claude’s is by far the best of the bunch. It’s not too far from what I’d email that notional Nieman Lab reader who asked me the question. And that shows the strength of Poe — the ability to ask the same question to multiple AIs and compare their answers.

Much of what people use these interfaces for now isn’t mission-critical stuff. (“Write a limerick about Sam Bankman-Fried” probably won’t get you fired or a raise.) But as these AIs integrate themselves deeper into our lives, it’ll be important to see them less as an all-seeing oracle and more as a set of compelling voices — who can still screw up sometimes.

Illustration by Midjourney AI.

  1. There is no dedicated Mac app, but you can download the iPad version to your (relatively recent) Mac in the App Store. ↩︎
  2. Witness this exchange, for example: “If all the water on earth was condensed into a single sphere, how big would it be compared to other planets in our solar system?” “Based on estimates from NASA, if all the water on Earth were condensed into a single sphere, it would have a diameter of about 860 miles (1379 km). This would make it slightly smaller than Mercury and slightly larger than Mars.”

    The 860 miles is correct — but everything else here is wrong. The estimate’s from the U.S. Geological Survey, not NASA. It’s not just “slightly smaller” than Mercury (diameter 3,030 miles) — it’s less than a third of the size. And 860 miles is not “slightly larger” than Mars (diameter 4,220 miles). Being smaller than Mercury but larger than Mars is a physical impossibility. ↩︎

  3. Or, in the case of Dragonfly, entomomorphized, I guess? ↩︎
  4. Until a few days ago, the Poe version of Claude was named “Heron” for some reason. Now it’s just Claude. ↩︎
Joshua Benton is the senior writer and former director of Nieman Lab. You can reach him via email (joshua_benton@harvard.edu) or Twitter DM (@jbenton).
POSTED     Feb. 8, 2023, 11:53 a.m.
Show tags
 
Join the 60,000 who get the freshest future-of-journalism news in our daily email.
How to b-e-e of use: Signal Cleveland hosts second annual community spelling contest
“Listening is great, and talking to community members is great, but we also have to figure out how to be of use.”
How South Africa’s largest digital news outlet plans to cover the chaotic 2024 election
“There is definitely anticipation in the air of change — not radical change, but some change.”
Postcards and laundromat visits: The Texas Tribune audience team experiments with IRL distribution
As social platforms falter for news, a number of nonprofit outlets are rethinking distribution for impact and in-person engagement.