#44: Are We Ready for Everything to be Automated?
New developments in AI pose the question
If you’ve been tuning into Relentlessly Curious for a while, you’ve likely picked up that I’m a big fan of AI. I’m always looking for ways AI can make me more efficient and effective, and I get genuinely excited by emerging use cases.
However, OpenAI’s latest ChatGPT release made me stop in my tracks. If you have a paid subscription to ChatGPT, you now have access to an agent that can pretty much do your job and chores for you. It can do all this because it interacts with websites designed for humans, the front-end experiences we use every day. This new agent represents the most advanced mimicking of human behavior on the internet to date. There’s a demo linked above if you want to see it in action.
“ChatGPT agent” (new release) can handle tasks like choosing a recipe for a curated dinner and then buying the ingredients on your behalf. No more ‘Oh, if I just had that one ingredient,’ excuses.
Also, it can build a slide deck based on your notes, past meetings (scans your calendar), documents, and spreadsheets. Speaking of spreadsheets, ChatGPT agent can scour the web for various data points (think public SEC filings) and build a financial model based on the data it collects. I’ll pause: how many corporate jobs have I endangered in this paragraph so far?
So, the gist is, ChatGPT’s latest advancement puts OpenAI’s models on a tier even closer to human capabilities (check out ChatGPT agent’s model exam performance versus other models and humans in the link above). In When AI Buys Your Groceries, I broke out a scenario where AI helps you decide what to cook for dinner tonight and orders the groceries on your behalf. I discussed this scene as hypothetical, suggesting we were 12 to 18 months away.
Well, I wrote that article in May and it’s July. Two months later, OpenAI highlights in their release how their unified agent capabilities allow you to “design and book entire dinner parties”. We’re pretty much there, as a partnership between OpenAI and Instacart suggests. Two months!
Clearly, AI is advancing at a blistering pace. If agents are executing complex human tasks end-to-end (like synthesizing disparate information from many sources into a cohesive presentation), Big Tech will disrupt specialized industries like niche client services and private market investing.
Here’s the thing though: ChatGPT agent and future developments from the likes of the LLM creators must not only execute a wide range of tasks, but also do so with a high degree of accuracy.
That’s difficult, especially since more specialized tasks require nuance. I believe we’ll see a lag in adoption on end-to-end agent execution compared to AI that “helps” you or automates parts of a complex project (but not all of it).
First off, humans place a disproportionate amount of value on things they partially created themselves. Sound familiar? This is a cognitive bias called the IKEA effect.
The name says it all. That bookshelf you built after spending an afternoon sprawled on the living room floor, engrossed by your hammer, screws, and the instruction manual provides a feeling of accomplishment. More than if West Elm just showed up and dropped it exactly where you wanted it.
Although ChatGPT requests permission before taking actions of consequence”, the agent makes it clear that it’s doing the work for you. OpenAI even advertises that you can walk away from your laptop after entering a prompt and expect a notification on your phone when the task is complete. You’re fully out of the loop. Impressive, but it’ll take some time until everyone wants that scenario.
We’re already seeing strong adoption where AI makes you better but doesn’t do the entire job for you. Think about how Cursor helps software engineers edit their code or fill gaps in logic. Or how Granola helps automate the note-taking process so you can stay focused on work calls. These are examples of AI that assist with complex tasks but stop short of doing everything. It makes you feel empowered, not relegated to the bench.
Another issue is that the technology still isn’t good enough yet to do the entire job for you. Ever let ChatGPT edit your writing or draft a document for you? Even with highly detailed prompts, you’re going to end up with “em” dashes (—), emojis, and an overly positive tone. Candidly, I use ChatGPT to edit my Relentlessly Curious pieces, but I always read it over out loud and perform the final edits myself. Otherwise, I lose my writer’s voice (something ChatGPT can’t quite replicate) and risk taking a brand hit if my writing looks like sloppy AI output.
For writing, AI is an assistant, not the replacement. It speeds up the editing process and cleans up low-hanging fruit grammar mistakes of which my 6th grade English teacher would be appalled. On that note, memorizing grammar nuances has become a lost art. An art I’m glad is lost if we’re being frank.
From my experience leveraging ChatGPT for editing, I’ve gained a keen eye for what type of content is AI-generated and what clearly comes from the human brain. I must admit that I have a bias against writing that clearly looks AI-generated (I’m talking filled with emojis, templatized structure, and the consistently positive tone). The LLMs will take you 80% of the way there, but you need to put in the remaining 20% of clean-up and originality work. Otherwise, you come off as lazy and unoriginal.
We’re more likely to see the adoption of tools built on top of these models rather than the raw models themselves. I’m talking about the AI that gets built off the mighty LLMs like the aforementioned Cursor or Granola. The application layer generally leverages both LLM insights, as well as their own proprietary logic that’s trained on private data sets. The combination of these insights gives them a better shot at handling edge cases and outliers than the base LLMs can.
To wrap up, I encourage everyone to give ChatGPT agent a try. Automation of complex tasks has the potential to massively boost productivity across society. But that doesn’t mean your job or your daily routine changes overnight.