For those of you who thought 'agentic' was obscure
AI excels at generating impressive content, often surpassing paid professionals. So why can’t it handle basic online chores like ordering a book or booking a flight? And what’s the outlook for tackling far more consequential agentic tasks?
On his latest trip to North America, my chronically accident prone friend was thrown to the floor by a savage jolt from a defective hotel iron. Fortunately there was no harm done and he did consequently get a completely free stay in the hotel’s largest suite.
As I’d serendipitously mentioned extreme ironing in this recent post, I’d like to send him a memento of his experience.
Stack rank the ease of execution for these options to acquire said memento.
- Buy a book on Extreme Ironing
- Create an image of a man ironing his kilt on The Inaccessible Pinnacle, the most challenging mountain summit in Scotland
- Scale The Inaccessible Pinnacle yourself carrying an ironing board and then take a selfie.
For most readers I suspect the stack rank in increasing order of difficulty for these tasks would be as they are listed above. And maybe the difficulty scores out of 10 might be something like 2,5 and 9 (unless my readership includes any hardcore climbers or extreme ironists.
It’s sort of obvious why AI cannot yet master the third option, although I’d wager that within 5 years this will be within the grasp of an AI powered humanoid robot. But it’s much less intuitive why AI can manifestly create the cover image for this post, while until a few weeks ago AI still struggled to buy a book from amazon.
Let’s start by separating AI into two categories. On the one hand Poietic AI digitally makes stuff for the user including long form text, Sophisticated images, lifelike voice and even studio grade video. On the other hand Agentic AI digitally does stuff for the user by interacting with external services like eCommerce websites, airline reservation systems, thermostats or nuclear reactor control rods.
My new best friend Open AI’s O3 model (available from the model selection menu in ChatGPT Pro) coined ‘Poietic AI’ from the Greek poiēsis (“the act of making”) . We’re deliberately avoiding ‘Generative AI’ (already used for Large language Models such as GPT - Generative Pre-trained Transformer), and we’re also avoiding ‘Creative AI’ because of all the baggage around the slippery question of what uniquely constitutes creativity.
So thank you O3 – as far as I can tell, this is the first documented use of Poietic AI as the counterpoint to Agentic AI.
[side note: I honestly thought Intel would forever top the Confused Product Numbering leader board until OpenAI invented the model numbering sequence 2, 3, 3.5, 4o, 4.1 , O3]
And now here is the rest of the post, written entirely by O3 after reading earlier Open Eyed Man posts as a style guide.
Poietic AI versus Agentic AI — why the maker beats the doer (for now)
1. The internet is a T-Rex–sized buffet of training data
Poietic models dine on public text, Reddit memes, cat videos, stock-music loops and every selfie ever posted. Action logs, on the other hand, are slim pickings. We don’t publish the minute-by-minute telemetry of our kettles, Roombas or marital negotiations, so Agentic AI lacks the high-fidelity sequences it needs to learn how to act. Result: the poets get fat on data; the agents stay on rations.
2. Feedback doesn’t require a crash helmet
When a language model hallucinates that Marc Andreessen invented the trombone, nobody dies. When a delivery-bot hallucinates the location of the stairs, somebody sues. Low-risk correction loops let Poietic AI iterate at Silicon-Valley speed, while Agentic AI must tiptoe through regulatory minefields, safety cases and, occasionally, my toes.
3. Business models favour content over conduct
Marketers, studios and overworked bloggers (hello) will pay today for infinite copy and concept art. By contrast, insurers, city councils and aerospace auditors insist your autonomous widget work perfectly before they green-light a purchase order. One revenue stream gushes; the other trickles behind compliance dams.
4. Human psychology loves a shiny bauble
A synthetic Drake track or a photorealistic clip of me disco-dancing with my guide dog triggers instant “wow.” A dishwasher that politely rearranges the plates? Worth applause, sure, but not exactly TikTok gold. Attention — the currency of our age — skews the hype cycle toward Poietic breakthroughs.
So what?
Agentic AI will have its day; nobody wants to spend retirement navigating kerbs with a white cane version 2.0. But until labelled behavioural data, safety tooling and public trust catch up, Poietic AI remains the sprinter in lighter shoes. The smart money banks on the makers: models that write the joke, draw the cartoon and lip-sync the punch-line while their agentic siblings are still filling in risk-assessment paperwork.
In other words, the orchestra is already playing; the stagehands will join us after they’ve completed a thorough safety audit. Until then, enjoy the sonata — Poietic-style.
End note: For transparency although all the words in the second half came from O3 I did delete a couple of complete sentences for brevity. FYI the actual prompt I used to elicit the Poietic classification and all the text was: “you are a blogger writing about the reasons why AI that creates content (text, images, video, audio, voice) is developing more rapidly than agentic AI. Propose a word for the AI which is of the content creation type rather than the agentic type. I propose not using Creative or Generative as these both have different connotations. Now write the post, which should make serious points but using language and a composition style similar to other posts in my blog - The Open Eyed Man”. O3’s output feels noticeably closer to my style than a similar experiment with ChatGPT-4o only 3 months ago – impartial views welcome.