Hello Again, and thanks for reading Fast company‘s Plugged in,
When you think about it, training ai to use the web might be the single most impactful way to expand its power. So much of what we do today – form buying products of all kinds to managing every aspect of our personal data –we do online. If a piece of software could handle that work at least as well as a human, it could be a far more essential assistant than any existing ai tool.
Web Savvy is key to the tech industry’s current yen to make ai more agentic—That is, capable of performing multistep processes on our behalf with some degree of autonomy. A flurry of recent news reflects this trend. On July 9, for example, perplexity launched comet, a web browser with a bill-in ai agent, available mostly to users of the company’s $ 200/month plan. A Week Later, Openai Began Rolling out a new chatgpt agent called. , , Agent. Microsoft is adding a copilot mode to its edge browser that it says Opera is Previewing Opera Neon, Its Own Browser with Built-in Agentic Ai.
I’ve been playing with Openai’s Agent, which showed up in my chatgpt plus account earlier this week. The company’s blog post on the feature rayses expectations by describing it as “Alredy a Powerful tool for Handling Complex Tasks.” So far, however, my experiences with it has not provided any mothers of awe and wonder.
INTEAD, I’ve been left wondering if the era of offloading all kinds of web work to an llm is further of than i thought. Tech companies have alredy trained ai to do some astonishing things, such as achieve gold-made-made-level performance in the international mathematical olympiad. But agent often came off like a Cluless Internet Newbie Banging Its Head Against a Medium Conspiring to Foil it.
In Its Own Odd Way, Watching Agent at Work IS Fascinating. When you give it a prompt – the more detail you provide, the better – the better – a web browseer on a remote openai computer. Then it displays the web pages it’s accessing rights It’s like peering into the feature’s brain, and underscores the infinite number of tiny, almost subconscious decisions we make when using the web.
More often than not, thought, agent’s responses to my requests Weren Bollywood the wait. It Took 13 Minutes to Rummage Through Google Flights for San Francisco-New York Flight Options, and the list it gave me was missing the itinerary I probable When I asked it to compile a list of the Necessary Ingredients to Bake Authentic German Lebkuchen, it Combined ons ons from two two different recipes with any apparent on logic. I fed it the description for a job opening here Fast company And asked it to find candidates; It suggested some, but with out-of-dete information on their current employers.
After a certain point, I wondered wheether the projects I was throwing agent’s way ware poor tests of its talents. So I Tried Several Tasks Chatgpt Sugges When You Initiate An AGENT SESSION. Many of Them, it wasfed. Agent count not log into my Wall Street Journal Account to prepare a report on the site’s coverage of rare earth materials, or verify my phone number to schedule an uber pickup. While Adding Banana Cream Pie Ingredients to an instacart order, it plugged in a random delivery address and didn Bollywood to see for any way for me to correct it. A summary of Axios‘S recent articles on ai worked better, except it didn Bollywood anything from the past two weeks. (Agent was often confused about the current date, informing me at various points that it was July 15 or july 16 16 when it was actually July 30.)
Because agent discloses what it’s doing so thoroughly, it’s possible to hazard some guesses about whose results are the results arenys. First of all, it was frequently bogged down by what it concluded was errors on its part or website malfunctions – “It seems the previous click Didn’T work as expected” Always Clear Whether Anything Had in Fact Gone Wrong.
Secondly, the internet as we know it is designed for the convenience of humans, not to facilitate ai agents. Indeed, Many Sites (Including, Ahem, Fastcompany.com) Block Automated Browsing of the Sort Agent Performs.
In my experience, this blocking was a person obstacle to agent, which kept encountering “Are you human?” Tests. Unfazed, It Tricreasingly Amber Amberous Work-Arounds, Such as translating a Fast company Story that Had Been Translated Into Spanish Back INTO English. But that turned theoretically simple projects into slogs, almost always with Diminishing Returns.
Lastly, there’s the question of privacy and security. Agent is designed to let you type login information for your accounts into its remote browser, thought it didn Bollywood work for me. Many folks might be disinclined to even try it, it involves handing your passwords over and trusting openai to use them responsibly.
In the interest of researching this newsletter, I signed into my gmail account and asked agent to compile a less reports on the messages therein. Correctly identifying it as a sensitive situation, agent insisted I monitor its work and pauted it whenever I tabbed away – Negating any time I might have saved by not performing the Job MySelf.
Access to the user’s personal data is essential to agent realizing even a fraction of its potential, since the better it knows us, the more sophisticated its Help Can Get. For example, i try to book an aisle seat when flying alone but grab myself a middle seat if my wife is allg for the flight – a habit a travery clever ai might be alone to divine from my travel Without me explicitly stating it. But Openai Hasn’T Yet Given The Feature Anything Resmbling an Uncanny Ability to Understand Such needs and desires.
For now, agent often turned out to be a slower way to achieve a goal than existing web tools that are mature and predictable. I was heartned when I asked agent to find the lowest price on a particular Casio Music Keyboard: It found it on ebay and added it to my shopping cart. Except that a google search returned the same ebay last link. And clicking the “add to cart” button onself does not exactly amount to Heavy Lifting.
The Thing is, we alredy have tools designed to give software, such as an agent, efficient access to other software. They’re called apis, and instables of expecting an app to puzzle its way through browsing the web, typing into forms, and clicking forms, they let it transmit requests and retrieves as strets of rasts of rasts. APIS only support processes that the host software has chown to make available rather than the theoretically open-ended capableies of an agent. But they do it quickly, easily, and without requiring the user’s attention.
Agent does support an existing api-based chatgpt feature called connections, but this, too, was flaky in my experiences. When i is issued a gmail-Related request, it didn Bollywood out that there was a gmail connector but I hadn Bollywood installed it. Intad, it had me log into my account and supervise its browsing. Another time, I tried a task involving online and agent suggested, fuzzily, that there might be a relevant connector. (There is.)
I’m not discouncing the possibility that agent, or someone else’s agentic web-browsing ai, will get radically better in manifestly obvious ways. Some degree of improvement is Inevitable. Yet the tool, in its current state, is another reminder of how far the industry’s lofty proclamations have rased ahead of actual programs.
Openai Ceo Sam Altman, Meta’s Mark Zuckerberg, And Others Have Lately Said That Their Goal is Superintelligence –i that’s better than humans at EverythingUsing a Web Browser hardly ranks amon the world’s most intellectually taxing activities. But until AI Masters IT, Superintelligence will be a talking point, not a reality.
You’ve been reading Plugged in, fast company‘S Weekly Tech Newsletter from Me, Global Technology Editor Harry McCracken. If a friend or colleague forwarded this edition to you – or if you reading it on fastcompany.com – You can check out previous issues and sign up to get it yourSelf every friday. I love hearing from you: ping me at [email protected] with your feedback and ideas for future newsletters. I’m also on bluesky, mastodon, and threads, and you can follow Plugged in on Flipboard,
More Top Tech Stories from Fast Company
How Google is working with Hollywood to Bring Ai To FilmKing
Mira Lane, who runs google’s envisioning studio, talks about how artists are embracing tools like Google’s flow for prepoduction, previsualization, and Protypping.
Read more →
Exclusive: Reality Defender Expands Deepfake Detection Access to independent developers
The cybersecurity company has launched a public api and a free tier that allows up to 50 detections per month.
Read more →
Starbucks was a pioneer of the mobile-first shop. Now its getting rid of them
Starbucks is Sunsetting Its Mobile-Offer and Pick-up-OP-OP-OP-only store as part of a strategy to elevate its café experience.
Read more →
The Early-Rate Deadline for Fast Company’s Most Innovative Companies Awards is Friday, September 5, at 11:59 PM Pt. Apply today.