There’s a lot going on in AI. Sending you a handful of curated news articles helps, but right now I can only tell you so much through them. Every time I click ‘Send,’ I feel like I’m missing out on so many things I’d like to tell you.

So, today, I’m going to try and summarize how I feel about the AI industry today.

We’ll cover:

Behind the paywall, we’ll cover my honest thoughts about ‘AInomics,’ and the key stuff to know about the real state of the model frontier, with one model in particular that you should definitely know about.

There’s lots to cover and learn today, so let’s dive in.

The reality is that, although I’m very skeptical about many things in this space, one of them is not agents, at least not when used well. However, that’s not the feeling most people have outside the SF/Twitter AI bubble.

My concerns about whether AI has real demand, which are very real and are addressed here, are real, but that’s certainly not because I don’t use the technology.

The problem is that, for some reason, most of society doesnt’ feel the same way as I do; I find using them for almost anything these days, when the vast majority of people, including well-educated persons in my surroundings, really struggle to find reasons to even pay $20/month, as they don’t see the value in something I would happily pay $400/month for.

What most people in San Francisco fail to realize is that this perception gap is real, and neither they nor I represent the opinion of the vast majority.

And while I’ve argued multiple times that adopting AI models is a ‘muscle’ that has to be trained, meaning the more you use them, the more reasons you find to use them, the harsh reality is that studies show that the average person gives almost zero value (median of around $11) to these products. We’ll circle back to this topic later in more detail because it’s particularly important to tackle the recent ruse the industry is throwing at us.

Either way, I’m a hardcore AI user who feels AIs create a disproportionate advantage (especially when most society fails to see the value). Thus, my ‘job’ here is to convince you by showing how AI has improved my life lately in and out of work.

In my work use cases, most are obvious: my job is to deeply understand hard topics and distill that knowledge for readers and clients alike, so LLMs are an incredibly natural fit for me. They help me test ideas, clarify doubts, build Excel models, crunch numbers to see if my theories make sense, and so on.

But, as regular readers know, I’ve been notoriously incapable of adopting agents as automation software. For example, I don’t use OpenClaw-type agents, nor do I have “my entire life automated” using Hermes, Zapier, or n8n.

In my experience, most of those workflows are just toy examples or, at the very least, aren’t needed for the type of work I do.

In other words, agents are consistently thrown at me as ‘automation solutions,’ and no matter how much I try, I struggle to see their value that way for now, mostly because the real tasks they can meaningfully automate with enough robustness are surprisingly limited.

Instead, I’ve become much more bullish on a particular type of use case: just-in-time implementations.

Eight months ago I published a viral article on Medium about the idea of ‘good enough software’, the notion that one of the benefits of AI-generated code is that you can pop up entire products in minutes that you need for that specfic time; that product can’t be put into a production environment for others, can’t really be sold to anyone, but it’s “good enough” for you at that point in time.

Before AI, code was not something you would do unless you were committed to monetizing it, because code required knowledge, expertise, and most importantly, time. This created a moat around software, and software companies could command 90% gross margins, selling their products at an absurd premium relative to the production costs.

In that article, I argued that many software companies would struggle to compete, because not only would they have real competitors, but their clients would also build rival solutions for internal use. This reality is no longer the case, and you can create valuable software (for you, at least) in minutes.

Incidentally, this could also create a software explosion, as software could be used in areas where it was inconceivable beforehand. If software “becomes easy” and “free,” there’s now plenty of new use cases for it that wouldn’t have existed otherwise.

Especially now, you can build software for things that have nothing to do with work, but simply make life easier by solving a particular problem you may be facing right now.

In my case, just in the last three days, I have used agents for the following:

As you can see, agents are incredibly versatile, crafty, and generally get things done. They make mistakes and need guidance, which is why they don’t work well as automation tools; but as a coworker, they are an incredible unlocker.

To be clear, you don’t have to be impressed by any of my “achievements”; that’s not the point. What I’m screaming at you is that “you can just do things” and your “things” might be more complex or not, but the lesson is that these tools can really meaningfully help you in ways you probably haven’t realized yet.

Hidden in the last paragraphs is what I believe will be a huge trend over the next months and years: marketing for agents.

Across all the examples above, the agent does some pretty technical stuff I don’t see, approve, or even know about. The truth is, despite being very technical myself, for most agent builds I give the agent “creative freedom.” In other words, the AI chooses the technology stack we'll use.

Now, picture a world where a vast majority of created software is human-unopinionated. Yes, right now enterprise software is absolutely opinionated; humans make the stack decisions.

However:

But even if you might disagree with my timings, the truth is that, as a technological provider (or any vendor, for that matter), it’s going to be vital that you’re well positioned amongst agents.

Agents won't only build software; they'll also do groceries, book flights, and much more in the future. Agents will have considerable freedom to act on user commands with tools of their choosing. And if so, you need to make sure agents consider your product or service as an option.

For example, I’ve been a Supabase customer for months now, and funnily enough, it was never my choice. The agent chose that fate for me as, for some reason, they decided all my software projects needed a Supabase database.

Interestingly, as this study by AmplifyingAI shows, Claude obsessively defaults to Supabase and prefers PostgreSQL over SQLite. In the case of Supabase, it mentions alternatives like MongoDB, but as researchers explain, “Models know it; they just don’t default to it.” This is terrible news for Mongo, and it’s a free customer-acquisition flywheel for Supabase.

Importantly to our topic above about models building stuff on the fly, models are also much more inclined to build stuff than to recommend tools. This can be easily explained by biases labs introduce to choose the option that requires more tokens they can charge for, but it’s nonetheless something to consider.

Nonetheless, I recently went “lawyer shopping” and had ChatGPT do the entire search. Even if I had the last word, ChatGPT had an enormous influence on who I eventually chose. If you’re a lawyer and ChatGPT isn’t considering you, you’ll have a long-term problem.

And what form could marketing for agents take?

Of course, the search API OpenAI uses introduces search bias, but the fact that the agent can tailor search to your highly specific needs introduces a factor that will be absolutely vital for marketing in the world of agents: ad hoc writing. Lots of it.

Having a presence isn't just about writing a handful of blog posts. It will require very specific copy that targets specific users, because unlike humans, agents will actually read your stuff.

Agents will choose based on what you offer, not solely on whether you're a high-ranked option for certain keywords; if you look at their reasoning traces to see the types of Google searches they do, they are incredibly targeted.

Two years ago, Sequoia estimated that the AI industry had already amassed a “revenue debt” of $600 billion; the amount of money it would have to make to give money back.

Today, that number is in the multiple trillions, but it's gained traction again for a different reason: it’s the value Anthropic’s own strategists project the company will reach in ARR sometime… next year.

Viable, or just another baking layer of the “AI hypers can’t get any more delusional” cake?

First of all, where does that number come from? It represents an extrapolation over Anthropic’s own revenue growth over the last three years, which has been an outstanding 10x, or an average revenue growth of 21.2% per month. Truly outstanding.

This also comes at a time of intense chatter about Anthropic’s upcoming IPO at, potentially, $2 trillion, plus their alleged $11.5 billion in second-quarter revenue.

What can I say, Silicon Valley loves linear extrapolations. So if the company has grown at that rate for three years… why not a fourth?

Let’s find out.

In a recent research study by Ramp, they show that the median employee “pays” $12/month for AI.

This number is surprisingly close to Stanford’s WTA value (Willingness to Accept) in the study we discussed earlier on, of how much money you would have to pay someone to stop using AI (a proxy for the value they perceive from these products), at a little over $11.

Crucially, if we are trying to understand what the masses are doing, the median is a much more telling value than the average, which can be skewed (we’ll go back to this in a moment), so we’ll use that for our estimates.

To measure how big the AI market can get, a defensible global estimate today for “knowledge workers” is roughly 1 billion people, with something like 800 million–1.2 billion depending on how broadly you define it.

But what is a knowledge worker really?

The cleaner way to approximate it is occupationally. The ILO classifies managers, professionals, and technicians/associate professionals as “high-skilled” occupations.

If you add clerical support workers, which are also predominantly information-processing jobs, you get reasonably close to the intuitive definition of knowledge work: people whose main economic output involves manipulating information rather than performing manual labor.

As not all knowledge workers will need AI, we can apply some discount and use a range between 600 and 900 million workers.

All things considered, assuming AI commoditization doesn’t force prices down, this market could be between $86.4 and $129.6 billion per year at current median spending.

Can Anthropic alone reach $600 billion per year under these assumptions? Nah.

But if I stopped my analysis here, I would be being extremely misleading to you. The reason is that the median doesn’t seem like the best metric for measuring AI revenue projections either. Maybe AI was never about “overall society adoption” and the distribution ends up being extremely skewed.

In fact, I wouldn’t be surprised if, in the future, most people don’t spend a single dime on AI and most of the ARPU (Average Revenue per User) derived from these freemium cohorts comes exclusively from the one market that doesn’t need you to pay to use products: advertising.

The real telling figures in Ramp’s study (which match Stanford’s study) are the top percentiles of users.

As you can see below, the median employee spends almost nothing (kind of proving my point that AI adoption requires putting the adoption muscle to work), but the top 10% spends around $660/month, and the top 1% an astonishing $7,500/month, or $90,000/year, higher than most human salaries are.

Thus, could these smaller cohorts be enough for Anthropic to meet its goal?

What does this tell you? Anthropic’s projections could make sense, but it’s a massive bet on over-the-top spending from a “small” cohort of people that needs to assume that, somehow, the AI industry retains enough pricing power to sustain those revenues per user.

But is this a given? Hardly, because as we’re about to see, frontier labs are under enormous pressure to drop prices.

Lastly, an often underestimated problem is a clear ‘US-bias’ in these numbers. I can see a software developer getting paid $600,000/year with an added $100,000 in AI spending, but that same developer in Poland or Spain is getting paid between $40-$70k.

Can that SWE justify spending $100k/year on AI? No.

So, in reality, the cohort that has to pay the bills likely concentrates in a few high-income places, not the best base from which to carry out ambitious growth projections.

But just like revenue projections require a lot of faith, profits, ironically, have a much clearer call.

Just like any teenager dreams of being the prom’s queen and king, all AI startups these days aspire to be the same thing.

And what seems like a harmless desire perfectly illustrates the calamity that my continent’s AI strategy, Europe, has been, and also pretty much reveals the industry’s biggest problem.