The Holy Fuck Gap

Why frontier users are feeling the AGI and everyone else thinks they've lost it

Moltbook, a social network designed for AI agents, broke the Internet over the weekend.

People observed the agents' behaviours with curiosity: turning errors into pets, starting religions, asking for legal advice, and building a government for bots.

While much of the bot communication on Moltbook is likely human-prompted to various degrees (including verbatim “post this” instructions), a lot of posts appear to be genuinely autonomous.

This led some people to point toward the takeoff: the moment when AI capabilities begin advancing so rapidly that they soon outperform humans across most cognitive tasks.

Yet many others scratched their heads and saw nothing interesting at all.

This AI hype divide captures the most profound societal trend in AI. Three paradoxes lie at its centre.

The first paradox: The same world is increasingly producing different realities. The AI world is splitting into two camps that seem almost fundamentally at odds.

The first camp is defined by the frontier of AI research and championed by tech Twitter. It observes that model capabilities are growing exponentially and believes that there is a high likelihood that they will continue to do so.

People in this camp are starting to really feel the AGI. Developers state that they don’t write any code anymore. Leading AI founders like Dario Amodei expect AI to produce Nobel-level research across various scientific fields within 2-3 years - the same timeframe in which this camp expects AGI to arrive.

Amodei is part of the1 San Francisco consensus: a small Bay Area elite that has situational awareness about an imminent future. They fervently discuss a post-labour economy and buying galaxies. Some commit their lives to the safe alignment of superintelligence.

The second camp is made up of…more or less everyone else. They don't understand what all the fuss is about. This isn't because they're not using AI: about 62% of US adults now use AI several times a week, and much of the second camp uses it at least occasionally.

Many in this camp think AI is cool. They find it useful to replace Google searches, and it lets them transcend boundaries they found annoying before, like cheating on tests, confiding secrets, and exploring ideas.

But they don't see AI transforming their day-to-day lives yet. When they hear people talk about AGI, they treat it like science fiction for nerdy adults, the way you'd listen to a scientist who discovered some obscure plant. Mild interest, sure, but what's for dinner?

They are the famous 95% of AI pilots that failed in the MIT media study. They are the studies finding no AI impact on labor markets yet (Goldman Sachs; Brookings; Yale). They are also the voices warning about data walls and capex bubbles.

The second paradox: As both camps use AI more, the gap between them widens.

How can that be? Is either camp deluded? Is the second camp too narrow-minded? Or is the first camp simply lying about AI progress, driven by self-interest to attract investment and sustain the hype?

The stakes are certainly enormous. We're entering an unprecedented technology investment cycle. AI capex already exceeds every historical investment wave, including railroads, electrification, and computers (Ark Invest). If this is all a lie, it would be the biggest financial fraud in human history.

The third paradox: Both camps are right.

Right in that they both accurately describe the reality they live in, and precisely because of that, they will understand each other less and less going forward.

But how can there be two different belief systems about AI that grow further apart as each camp uses AI more, yet both remain right in their assessment?

The resolution lies in different exposure to the capability frontier. AI capabilities exist on a steep but uneven curve. Your position on that curve influences everything you see, feel, and therefore believe about the future.

What makes this curve so uneven in the first place?

Double-jaggedness

Humans are intelligent. We are able to acquire and apply knowledge, learn from experience and adapt to new situations and environments.

AI is also intelligent. It learns from data, and while we don’t quite know how, it is able to make decisions and reason about the world. However, AI intelligence is jagged. This means it is very good at some things, while it struggles a lot with other things.

For example, AI has superhuman abilities in information retrieval, as well as making advanced mathematical calculations or playing complex games like Go.

At the same time, it struggles with tasks that are easy for humans: common sense reasoning (like how it is impossible to unscramble an egg), showing contextual understanding (like connecting separate but intuitively related information), spatial reasoning (like solving a puzzle) and learning continuously.

We call this Moravec’s paradox: Many things that are hard for humans, AI excels at, while many things that are easy for humans, AI struggles with.

But human intelligence is also jagged. We have varying verbal, mathematical, and creative skills. People rarely excel across all dimensions - but rarely struggle across all of them either.

And so, at the frontier where AI and humans meet each other, jaggedness meets jaggedness.

The key difference between our intelligences is evolutionary speed: human intelligence advanced over millions of years, while AI intelligence is improving exponentially on a timescale of just a few years.

If this progress continues, AI intelligence may soon look more like this:

Explaining the divide

Next, let’s examine how AI shows up in our everyday life, and specifically, how it may start to automate certain jobs.

A job consists of a bundle of tasks. Because of jaggedness, AI can only automate some of those tasks today. The remaining tasks that AI cannot automate become bottlenecks that economists call weak links (Jones, 2011).

Which exact tasks get automated is less predictable than Moravec’s paradox might suggest. While it’s true that AI often excels at computationally hard tasks and struggles with things toddlers can do, this pattern isn’t universal. Some tasks are hard for both humans and AI (novel scientific reasoning). Some are easy for both (basic data lookup). The mapping between human difficulty and AI capability is jagged and context-dependent.

AI can only diffuse through society at the speed at which weak links allow. Weak links become the pacemaker for AI. No matter how incredible AI becomes on one hand, if we cannot overcome the weak links, its diffusion in society will be gradual. They’re like a valve through which AI must pass, even when it is otherwise ready to fire from a hose.

Weak links appear in two forms: AI's current limitations in areas like reasoning and physical tasks, as well as human barriers to adaptation, including age, education gaps, technical inexperience, and resistance to change.

Based on the underlying bundles of tasks, AI adoption also varies by job function. For software developers, AI jaggedness aligns almost perfectly with their task jaggedness, since writing code is precisely the kind of structured, logical, text-based digital work AI handles well. They are therefore able to see transformative productivity gains.

Nurses on the other hand cannot get transformative productivity gains out of AI yet. While they can research patient symptoms with AI health solutions, AI jaggedness misaligns with their core tasks that focus on personal care.

Pricing managers fall somewhat in between. AI excels at pricing analysis and optimisation, but struggles with organisational politics and knowing when to make strategic exceptions.

Why the gap widens as usage increases

As people use AI more, those least constrained by weak links begin experiencing compounding returns. They discover new capabilities, build better workflows, and uncover even more capabilities. A recursive improvement loop sets in, not in the AI model itself, but in how effectively they extract value from it.

This compounds with another advantage: deep model understanding. Model capability overhang describes the gap between what AI can do and how most people use it. Power users who understand models intimately, including their capabilities and jaggedness, can navigate around weak links and exploit capabilities others miss. They’re not just less constrained but also better equipped to find and use what the model can actually do.

The results are staggering. According to developers at Anthropic, Claude Cowork, the version of Claude Code for knowledge workers, was built in only ten days, thanks to significant speed up of developer productivity through coding agents. Similarly, AI-first companies are able to achieve faster revenue growth than ever before.

For some, AI creates exponentially increasing value with constant "Holy Fuck" moments. For others, weak-link bottlenecks impose utility ceilings where even AI models reaching International Mathematical Olympiad gold-medal performance feel merely incremental.

I believe these productive feedback loops don't activate gradually but rather kick in at discrete thresholds. Claude Opus 4.5 was such a threshold for agentic coding: frontier users could suddenly compound their productivity and accelerate away from others. Those who haven't immersed themselves in frontier-model powered workflows, or who work in domains still suffering from a lot of weak links, remain stuck below the capability threshold.

Both camps accurately report their reality, but understand each other less with each passing month because they're positioned on a curve with a fundamentally jagged frontier and are thus experiencing very different realities at the same moment in time.

Boosting AI’s usefulness: complementary innovation

Even if AI models keep improving exponentially and become less jagged, large-scale productivity gains may take time. Historically, new technologies often boom while productivity initially stagnates, sometimes for decades. Economists call this the “productivity paradox.”

Electricity arrived in the 1880s as a vastly superior form of energy to steam, but until 1910, many stuck with their existing steam engine systems.

The problem lay in the full investment required to change: A steam factory needed heavy drive shafts and was arranged entirely around them. An electric factory could be light, airy and organised around production lines. So rather than just swap the motor, you needed to redesign everything: architecture, production processes and even how you paid workers. Factory owners didn’t want to scrap existing investments and struggled to imagine the full implications of smaller electric motors.

Only by the 1920s, productivity in American manufacturing soared once manufacturers finally embraced the technology.

AI is similar. The models are already powerful, but they are jagged and create weak links that slow down societal transformation. But even if and when we overcome jaggedness, productivity gains will depend on complementary innovation: redesigned processes, different team structures, large scale employee training, updated HR policies and compliance.

Unlocking AI's value requires deep AI adoption: not just using the tools, but fundamentally redesigning work around them. You cannot simply add better technology to old systems. You have to do things differently. That takes trial, error, and appetite for risk. For many organisations, this will take years. But if I’m honest, I'm not sure they have that kind of time. Those who lean into complementary innovations the fastest will capture the rewards while those who wait may find the gap unbridgeable.

Bridging the divide while we still can

We're still early in AI, yet already diverging into different realities. Moltbook offered a glimpse of what agent swarms coordinating, sharing resources, and collaborating on projects will look like.

But the way tech Twitter lost its mind over an arena of AI agents also shows how little we have thought this new world through so far, how much more amazement, surprise, and danger awaits us all in this wondrous future.

Boosting deep AI adoption will help bridge the divide and decrease model capability overhang. But beyond model improvements and adoption tactics, the race to advance AI capabilities must be matched by an equally serious commitment to understanding and shaping the uneven diffusion of AI through society.

Who will benefit? Who will be left behind? Shaping AI for broad human benefit requires camp 1 to engage in more research and effective communication with the rest of society and camp 2 to extrapolate beyond their current experience and instead grapple with how their reality will transform as the exponential continues.

“The majority of the [divide] is people who look at the current point and people who look at the current slope, which in my opinion gets to the heart of the variance.” Andrej Karpathy

It’s time for everyone to feel the AGI.