When chatgpt debuted two years ago, it demonstrated a clear departure from an ecosystem of assistants and chatbots that came before it. For the first time, we could envision a world of any input to any output at a time when Siri and Alexa required -and still do- clear instructions to trigger certain responses. They’re trained to handle a handful of actions and a handful of search queries, but they operate within defined parameters. The limitations of those assistants relegated them to the role of alarm setters and music players, but not much else.
“play Espresso by Sabrina Carpenter” is a plausible, certain input -if you’re a 30-year old man pretending to be a 18 year old girl- that the assistant can structure into an action to send over to Spotify to play a song. But if you try “I’m in the mood for a hit pop song, what do you got?”. That’s an uncertain input that would throw off the assistant. What the AI breakthrough showed us is that no input is too difficult to decipher.
But two years since that AI breakthrough, the consensus is that the future is “Agentic”. That’s a world where AI can process uncertain inputs and produce uncertain outputs. Building that digital publication gave me a peek into what that future might look like.
When creating this digital publication, I knew what I wanted to do and the tools I wanted to use when building this digital publication. So I gave the code generation agent a laundry list of things to do and the tools to use.
”I want to you to work on my backyard garden. dig the soil, plant the seeds and water the area when done. Please use shovels” I prompted the agent. “Here’s your garden, I used shovels to dig the soil. Let me know if you need anything else” the agent responded. “Oh great! wait… what’s that excavator doing there?” I exclaimed. Turns out the agent resorted to using a combination of shovels and excavators to dig. Did I really need an excavator to dig a backyard garden? Probably not, but I’m no gardening expert. I have a beautiful garden and that’s all that matters. And most people wouldn’t mind either, as long as the task gets done.
The conundrum of choice is not new to tech. One incarnation of it is Yelp’s decade long feud with Google. User runs a local search, gets the google “OneBox” at the top and yelp’s link at the bottom. Yelp cries foul alleging anti-competitive behavior. Google’s argument is that it serves the better results to their users. But who’s to say? Google does. Either way, the output is certain, google search processes the query, the search ranking algorithm churns out the results based on criteria set by Google for rank search results.
Because after all, do users care if there’s a 10% chance their search for a local pizza shop around the corner does not show enough reviews in Google’s “OneBox”? Probably not, but at the off chance that they do, they can still scroll down and click on that Yelp link. Yelp’s argument? Users don’t scroll.
Yelp is right. Google has been optimizing their search engine for decades to surface the most relevant information to a user’s query at the top in the fastest time possible. Google was never in the business of wanting users to stay on its service. The point is to find the information as fast as possible and jump off. Every Google optimization serviced that, disincentivizing users from going past the first page to not scrolling. And now with AI overviews, users are less likely to click on links.
Google processes billions of queries a day and there’s a wide spectrum of things that users “google”, but they largely fall into two major categories; factual and subjective.
“what’s the weather like tomorrow?” is a factual query. Google would look up that information on the Weather Channel based on your location and then serve you the result ”what’s the best burger place in town?” is a subjective query. Google would typically surface review sites, forum sites and other content generated by people. The AI overview iteration of that would be the summary of what these sources say, so a user would get 5-10 names of burger joints to try out. And that's just search, it controls what users get to see, but not what they decide on.
Now try another query with an action-oriented agent and the output is much less certain. “Order me a healthy breakfast” you would ask the agent, and then 30 minutes later you’re enjoying a McDonald’s fruit and maple oatmeal with all the added sugar that your heart desires. Yum!
Agents tend to resort to defaults, leaving you with little room for research and discovery. They’re about getting tasks done. And the less you know about something, the more likely you’ll end up with a suboptimal default, or in that case, McDonald’s.