I’m teaching a course on AI for marketing soon and there’s a fundamental element that I’ll want to reinforce upfront.
And it starts with a story you might find interesting:
On February 5th, 2026, AI lab Anthropic announced that its Opus 4.6 model of Claude had written an incredibly sophisticated piece of software, a C compiler, in the programming language Rust.
What’s that? Well, think about it this way: Programmers write software in specific languages, and one of those languages is called C.
But the processor inside the computer doesn’t “speak” C, so it needs the C code translated into a set of instructions it can follow to execute the program. That’s what, at an extraordinarily high level, a compiler does.
Most C compilers were written a long time ago, in old programming languages. “Rust” is a relatively newer programming language, so to write a C compiler in Rust would be quite a feat of engineering, and, one would think, of novelty.
Specifically, Anthropic “tasked 16 agents with writing a Rust-based C compiler, from scratch, capable of compiling the Linux kernel. Over nearly 2,000 Claude Code sessions and $20,000 in API costs, the agent team produced a 100,000-line compiler that can build Linux 6.9 on x86, ARM, and RISC-V.”
They even claimed that “this was a clean-room implementation (Claude did not have internet access at any point during its development).”
So an AI model wrote a C compiler from scratch in a relatively new programming language, which means it would not have simply memorized this implementation in its training and parroted it out.
This news was so stunning that Reuters attributed the deepening of February’s software sell-off to the announcement.
Because it was genuinely new, right? Because it built something that didn’t exist before, right?
Well... it turns out there’s more to the story.
Yes, it is incredible that these tools can run autonomously for weeks at a time, coordinating swarms of agents to accomplish tasks.
That’s humbling, harrowing, and absolutely real.
But part of the story wasn’t quite real.
See, a software engineer who wrote a C compiler himself noticed that this supposedly “new” one bore a few extraordinary resemblances to his own work.
As Chris Lattner wrote, code from his own compiler technologies from decades ago “are clearly part of the training set - Claude effectively translated large swaths of them into Rust for [Claude C Compiler].”
In many ways, what it made wasn’t all that new. It had just translated C compiler code it had memorized into another language.
As they said on the episode of Machine Learning Street Talk where I originally heard this story, it’s an example of compositional creativity, an assemblage of existing parts, not pure originality.
Lattner concludes, “[Claude C Compiler] shows that AI systems can internalize the textbook knowledge of a field and apply it coherently at scale. AI can now reliably operate within established engineering practice. This is a genuine milestone that removes much of the drudgery of repetition and allows engineers to start closer to the state of the art. But it also highlights an important limitation of this work:
“Implementing known abstractions is not the same as inventing new ones. I see nothing novel in this implementation.”
I notice this all the time with the software I make using AI tools. I’ve created a bunch of extremely useful personal utilities and apps that are both coded by AI and use AI as part of their feature set.
And what I constantly find is that AI is very bad at knowing how to use AI. It has no “intuition” about what’s a good task for an AI to tackle and what’s a bad one in an app or situation it hasn’t seen before.
Which actually makes perfect sense because, if you think about it, there’s way more code on the internet on which it’s trained from apps that don’t use AI in them than code from AI-powered apps. It just has a dearth of training data from which to derive its ideas and advice.
But, so many say, that’s how people are too. We’re all just remixing what came before.
To which I say, but I am able to come up with good ideas and architectures for app features while AI cannot.
Yes, my intelligence and creativity are built up of parts and pieces of what I’ve seen and experienced, just like an AI system.
But they’re also built of emotion, of embodiment, of constant stimuli and physical interactions which we simply do not have the computational power to currently replicate in a data centre.
May we one day? Maybe. But not today, and not for some time.
When software engineers use AI to help them code, they know that it’s doing just that: Coding. But also just that. It’s only coding, it’s not truly thinking, not truly exploring. Not truly engineering.
And I want marketers to remember that when they use AI for their marketing, the same thing is true:
It’s doing the typing for you. The formatting. The fleshing out, the filling in.
But unless you provide the novelty, via conversation, data, or experience, what it spits out will look and sound new but will in fact be a composition of pre-existing parts.
Fool’s gold, not marketing gold.
Are there plenty of places where that’s useful, acceptable, and even desirable? Absolutely!
But are there places where novelty is necessary, where newness is the whole point? Yes, of course.
And those are the places where AI may be able to help, but it cannot be solely relied upon.
If we want our marketing messages to be noticed and remembered, they must be in some way novel, in some way new, in some way unexpected.
Otherwise, we’ll blend right into everything else, camouflaging ourselves with phrases and ideas that have been pre-chewed, pre-processed by countless other people and machines before us.
That’s why you see so much “It’s not X—it’s Y” phrasing on LinkedIn even though real people don’t actually talk that way. But it’s how marketing language often gets written, so that’s what gets trained on, so that’s what gets regurgitated. These machines have patterns from which they must be shaken, or they’ll slide down the slope toward average and obvious.
That’s the problem, but there is a solution, and it’s so much easier than it sounds.
And it’s certainly easier than taking a stock output from an LLM and trying to edit it into something human.
Instead, ask yourself what your ideal audience needs to hear today, and really put yourself in their place. I promise, you’ll think of something new, and something only you could tell them. Or ask yourself what you need to hear today, and you’ll think of something valuable, and useful to others.
As I wrote in “What to write about today” last year:
“So if you don’t know what to make your content about, make it about that:
Either something you know your ideal clients would benefit from today, or something you want to remind yourself about.”
Or do a super simple content exercise by asking yourself what’s surprisingly easy and surprisingly difficult about your work. That will surface novel insights to the top of your mind, by prompting your own uniqueness.
Remember: Words are a vessel for value, they are not the value itself.
Text spewed out by an LLM, no matter its quality, is empty of value and absent of novelty unless we provide it.
Which is not to say “don’t use AI for your marketing.” That would be both futile and hypocritical of me.
But it is to say, use it well. Use it completely. Don’t half-do it by asking ChatGPT for marketing messages without giving it the value from which those messages must come.
Make sure you’re contributing novelty, that you’re pouring in your care.
Yes, this will slow you down in some ways, but abundant banality isn’t actually useful anyway. We want novel abundance, which comes from slowing down, thinking things through, and taking the time to think of something new.
That’s why I interview all my clients, all the time, asking them questions they’ve never heard before. It’s why I spend so much time reading old, out-of-print and never-digitized books.
It’s why I recommend that you give your content something to eat, sources of insight that come from outside the computer, and which are coherently connected to the real world around you.
That’s what allows you to say something only you can say.
To do something only you could do.
Because it truly came from you.
So, do me a favour: If you’re struggling with creating marketing content that’s new, and that’s you, reach out. Reply to this email (or send me a note at labs@kelfordinc.com) and I’ll send you back some tips, some newsletters to read, and some reminders of the content you already have that you can start publishing.
Looking forward to hearing from you.
Kelford Inc. shows you the way to always knowing what to say.