AI development remains on an exponential trajectory. The author, who publishes periodically on Substack, notes that events are accelerating so quickly that publishing every few weeks can feel too slow: in recent weeks an AI reportedly cracked a famous math problem (with controversy) and public debate about AI risks and responses intensified (also controversial). Many of these worries stem from the same issue the author sees in his posting cadence: human systems and processes are much slower to adapt than AI.
The present is already transformative
Worries about future models are important, but current systems are already remarkably capable. The author points to GPT-6 Astra and Fable 5.1 as examples of models that today can produce transformative effects across large parts of the economy; with correct guidance and tooling they can reliably perform weeks of human work.
Concrete examples from the author’s experiments:
- Using GPT-6 Astra, he converted Zork, a 1977 text-adventure game, into a full 3D action-adventure game. Zork has no graphics and each location is a paragraph of prose, so the model had to invent visual details (for example, what the white house or a grue look like) and turn commands like “fight the troll” into an action sequence.
- With Fable 5.1 he attempted to reconstruct Umberto Eco’s library in Milan in 3D. Eco kept tens of thousands of books; lacking a floor plan, the model worked from roughly a dozen videos, foundation photographs of each bookcase, and two catalogs. It read spines frame-by-frame, inferred rooms, and placed about 5,000 identifiable books among roughly 27,000 shelf slots, tagging each book as certain, guessed, or unknown and drawing the unseen bookcases in fog.
These tasks, and similar real-world work the author has had AI perform recently, would have taken human teams of researchers, coders, and designers many weeks to complete, yet the models did them in short order.
Capability overhang: a gap and an opportunity
The author argues that most attention focuses on future model risks, while the capabilities of existing models are underused and often poorly understood. He did not know GPT-6 Astra could operate Blender until the model demonstrated it.
He gave the model a draft of his upcoming book, Co-Existence, and asked it to create a trailer from an AI’s perspective. Without detailed technical instructions, the model used Blender to build a full animated 3D scene, wrote a script with jokes and reveals (the author rejected the first joke but kept the second), generated voices, music, and sound effects, and produced a film about 45 minutes later. When asked to make an action-movie style trailer no longer than 30 seconds, the model again wrote a script, built a 3D prototype in Blender, then used that animation as a storyboard to operate a browser-based video generator and edit the generated shots into a final trailer.
These outputs contain flaws, but they also demonstrate forms of judgment and creativity that not long ago were seen as uniquely human. Importantly, these projects required the author’s choices: he selected the tasks, he knew the source material well enough to spot errors, and he requested revisions when necessary. That gap between what models can do and how most people use them—the capability overhang—is large, and it represents an opportunity because few users bring the background knowledge needed to fully exploit AI.
Four human advantages to cultivate
Rather than compete with AI on producing outputs, the author recommends using human advantages to work with AI to achieve things neither could do alone. He identifies four personal traits that matter:
- Deep knowledge
- Deep, domain-specific expertise builds intuition that enables quick, accurate decisions. Examples include an accountant spotting errors at a glance or a golf pro diagnosing a swing. Deep knowledge helps experts map the “jagged frontier” of where AI succeeds and fails in their field, and makes it easier to shift from doing the work to managing it. The author also cites work from Anthropic suggesting that expertise improves both the quality and quantity of what AI returns: experts not only get better outputs, they get more out of models.
- Wide knowledge
- Large language models are trained on a broad slice of humanity’s output and have learned many conceptual patterns — design principles, Bayesian reasoning, production systems, therapeutic techniques, literary theory — but they usually surface those patterns only if prompted correctly. Wide reading and study across fields helps you know what to ask for (for example, knowing to tell a model not to add extra small subheadings when it generates webpage copy, or recognizing that a Blender animation can serve as a sensible storyboard). Broad knowledge amplifies your ability to direct and correct AI.
- Taste
- Before generative AI, production was the bottleneck; now creation is fast and cheap, and the scarce resource is the ability to select among many similar outputs. Taste allows you to pick the best outputs, discard the rest, and use AI-generated material as raw material for something the model wouldn’t have produced on its own. The author’s trailer revisions—rejecting one joke, keeping another, asking for a more cinematic edit—are examples of taste-driven decisions.
- Agency
- Agency is harder to define but here refers to the willingness to test boundaries and explore what’s possible when the frontier is uncertain. It’s the difference between waiting for others to report an AI capability and discovering it yourself by experimenting. The author’s many odd experiments (for instance, trying to get AI to play games) are ways to learn what AI can and cannot do.
Book release and pre-order bonus
The author expands on these four advantages in Co-Existence, which is published October 20. If you pre-order and notify the author at co-existence.ai (pre-ordering helps authors), you receive a link to a free AI-conducted voice interview. The interview asks about what you know, what you like, and what you have tried, then gives a report on your deep knowledge, wide knowledge, taste, and agency, along with suggested use cases and prompts built around those traits.
Where this leaves us
Much current anxiety centers on future models and whether we can control them; policymakers and labs are right to debate management and pacing to mitigate risks. But slowing development does not remove the capabilities that already exist. Even if all labs stopped training new models tomorrow, GPT-6 Astra and Fable 5.1 would still be able to change parts of the economy. The capability overhang between what these models can do today and how people typically use them is massive.
Change is coming regardless of how we pace the frontier. It will be uneven and incremental, not instantaneous, but inevitable. That does not mean the form of change is fixed: as a society we should develop and share models of human–AI work that enhance human labor rather than merely replace it. Individually, we should use AI to augment our efforts, not only substitute for them. The four human advantages—deep knowledge, wide knowledge, taste, and agency—offer a practical starting point for doing that today.
