What does a three decade-long path of exploring the jagged frontier of AI and asking what separates human cognition from a machine’s look like? The answer says as much about the machines as it does about us.

By George Anadiotis

At eleven years old, Carlos E. Perez was already choosing between two possible futures. One was first contact with an alien civilization. The prospect thrilled him, given an early obsession with astronomy, but was entirely out of his hands. The other was something he could actually work toward: building AI. He picked the second path. Inner space over outer space.

The route there wasn’t direct. Perez studied physics as an undergraduate, decided he wasn’t sharp enough for it, and switched to something he found easier: computer science. A stint at IBM Research working on internet and B2B protocols followed, then a string of startups tinkering with the AI of the era: subsumption architectures, constraint engines – the classical, symbolic kind.

When deep learning broke through around 2012, Perez assumed his physics background would give him an edge. Gradient descent, after all, is just continuous mathematics. By his own admission, he was wrong, but the way he was wrong turned out to be the more interesting finding.

Perez co-founded The Intuition Machine in 2015, and he has published a number of books and essays on AI. He also works with Fannie Mae as a consultant, providing AI strategy and architecture guidance for the Enterprise Architecture organisation.

We explore his path of discovery and tinkering: from AI’s uneven capabilities to what’s actually changing between model generations, enterprise architecture, AI harnesses, loop engineering and graph engineering for AI agents.

In “the Jagged Frontier,” Perez names something most people who’ve used a frontier model have already felt: islands of superhuman competence sitting next to gaps a first-year student wouldn’t fall into. A legal AI can shred a contract in seconds and then miss context a paralegal would catch on the first read.

That unevenness isn’t noise, Perez argues: it follows the shape of the training data. “Whatever curriculum was used was the low-hanging fruit available for training,” he says, and that curriculum, running from Reddit threads to books to code, is a poor substitute for a body.

Humans build intuition through decades of embodied experience: what a fall from a height feels like, what a cliff edge means. LLMs build outward from text, with no body at all, which Perez says is why their failures often look nonsensical to us: the developmental path runs in reverse.

The gap compounds because we’re jagged too, or as Perez puts it, our own evaluation of AI is unevenly capable. We judge models against a working memory limit of roughly seven items that they simply don’t share. When a model instantly renders a ranked top-ten list that would take a person twenty minutes to assemble, it doesn’t just save time. It reshapes how we think.

Perez reaches for Marshall McLuhan’s line to point out that media amplify certain faculties while letting others atrophy: GPS users gradually lose the capacity to navigate without it. The same trade is underway with AI, quietly deciding which of our own cognitive muscles we keep exercising.

Software Engineering, Abundance, and Meaning

Nowhere does that jagged frontier of AI play out faster than in Perez’s own craft. Because generative AI can produce code prolifically, the scarce resource has moved downstream, from writing to checking.

The symptom is already visible in open-source repositories, where pull requests generated faster than any human reviewer can reasonably absorb are starting to pile up. This is a strain other engineering leaders are describing in similar terms, and one that echoes broader concern about what AI-driven productivity is doing to junior hiring.

Perez’s explanation for why senior engineers benefit disproportionately from AI tools is architectural, not experiential in the vague sense: junior training focuses on optimizing an algorithm, not shaping a system. Architecture is a taste acquired over years of shipping and fixing things, and it’s the one skill that lets an experienced engineer keep an overarching shape in mind while steering an AI through the details.

What he’s certain of is that engineering has to move “toward expressing our humanity” as the machine-like parts of the job get absorbed. The things people find meaningful, he notes, tend to be the hard and inconvenient ones: climbing Mount Everest would be immensely more meaningful than being helicoptered, even if that was possible.

Perez makes an unusual claim that cuts against the usual scaling-law story on what’s actually improving between AI model generations: raw ability barely moves. What moves is the division of labor – how much of the job gets handed over.

He compares it to onboarding a new hire, watching an LLM progress from producing text, to following instructions, to holding a conversation, to reasoning, to retrieving information, to acting, to persisting, to eventually setting its own goals.

Pushed on whether that hand-off isn’t itself gated by real capability gains, Perez concedes ground carefully. GPT-3, which he considered close to AGI in 2020, stayed recognizably the same model through several years of refinement.

Instruction-following was added, then code training, which is where Perez thinks causal reasoning first showed up, then reasoning layers on top. “It was essentially the same underlying model needing more experience over time,” he says, “like a child developing from birth to age five.” Humans, by contrast, arrive with years of evolved priors and learn from a 3D world, not a vocabulary.

That gap is the engine behind Perez’s library metaphor: today’s models resemble someone who has read every book in a library and lived none of it: fluent about the world, ungrounded in it. He offers two ways that resolves.

Either models eventually grow a felt sense of their own, one that won’t quite match ours and will be hard for anyone to revise once formed, or they keep manipulating the language of meaning indefinitely with nothing underneath it. Bodies, in his view, are the only route to the first outcome, and today’s curated, cleaned-up training sets are a poor substitute for the failure modes a body actually encounters in the world.

Perez is careful to caveat one thing he’s noticed testing the newest wave of models: a shift he describes as “a different feel,” an apparent sense of how complex objects (buildings, machinery) assemble into a working whole, rather than just a fluent description of the parts. It’s an early, subjective read, offered as exactly that.

Perez’s enterprise architecture background is where the conversation turns most concrete. His claim: in an AI-native organization, EA and the AI harness are the same artifact. Traditional EA documents a mess after the fact and goes stale the moment it’s finished. AI-native EA instead keeps the enterprise’s structure – things such as decision traces, boundaries, accountability lines – explicit inside a model’s context window, queryable rather than archived.

In Beyond the Decision Trace, we’ve argued that Context Graphs need Knowledge Architecture and showed how ArchiMate EA models can become live, reasoning graphs rather than static diagrams. Perez arrives at the same convergence from his enterprise architecture experience: an organization’s context, once it’s legible to a model, stops being documentation and starts being infrastructure.

That matters because roles and departments have always functioned as modules of human responsibility, and AI erodes the reason those boundaries need to be rigid. Perez reaches for British naval history to make the point: ships kept formation in battle mainly so admirals could verify everyone had followed orders, not because it was good tactics.

Nelson broke that pattern by giving individual commanders the freedom to engage as the situation demanded. Native AI organizations face the same trade between legibility and fluidity. As Perez sees it, specialization is still the safer default even when a single model could technically do everything.

Individual AI agents improve through a feedback loop borrowed from control theory: pick a metric, set a target, measure the gap, act, repeat. But single loops break in predictable ways: Goodhart’s law, blind spots, conflicting objectives, decaying signals.

Perez echoes OpenClaw’s Pete Steinberger about fleets of agents outgrowing “loop engineering” in arguing that mature AI systems need graphs of loops instead: networks where different loops observe, constrain, and override each other when one breaks down.

Which is why graph engineering is gaining traction in building AI agents too. But that, too, has its own failure modes. It’s a fitting place to leave things, because it’s the same argument in miniature: whether it’s a model, a codebase, or an org chart, isolated components eventually need a structure around them to stay honest.

What that structure preserves, in Perez’s view, is exactly what’s hardest to automate: the deliberate, effortful part of the work that makes it worth doing in the first place. The single loop was how systems learned to get better. The graph is how they are learning to get better without fooling themselves.

Staying honest about what “better” means is a different lesson than either. Perez will present his graph engineering thesis at Connected Data London 2026.