AI's new bottleneck isn't in the model: it's in what we're able to imagine for it

Few tech reviews are worth reading for their argument more than for their product, especially when that product can no longer be used. Such is the case with the piece Nate devotes to Fable 5, the model he describes as the most capable he has ever tested and which, as he recounts, the US Government pulled from production a few days after its launch, leading Anthropic to shut it down worldwide. The author recorded his assessment before the outcome, briefly withdrew it, and finally decided to publish it, with a reasoning worth keeping in mind: staying silent about what the model can do would be, in his view, the wrong response to having lost access.

The anecdote that frames the article is eloquent. Nate handed Fable 5 a deliberately poisoned database —phantom records, corrupt files, planted traps— and, instead of supervising as he usually does with every new model, he went off to do something else. When he came back, the work was not merely answered, it was done: the database clean, the junk in quarantine rather than 'fixed' behind his back, and a review queue built on its own initiative with the doubtful cases, as if the model anticipated it would be verified. That behavior —completing a real task, deciding and flagging its own uncertainties without being asked— is what the author identifies as the qualitative leap.

From that he draws a provocative claim: Fable 5 would be the first model bigger than our habits. For years, he argues, the model set the limit and the user learned to ask below that line; here, for the first time, it was he who ran out of things to delegate before the tool exhausted its capacity. The ceiling shifted from the model to the imagination of whoever uses it. He names that skill 'detailed task imagination': the ability to conceive complete jobs, not prompts, and he describes it as concrete and learnable, even though almost no one teaches it because three years of AI culture have been devoted to optimizing instructions instead of assignments.

These claims should be read with the caution an individual testimony deserves: they are one user's experience of an already inaccessible model, not a reproducible measurement. Nate himself qualifies that anyone using it for summaries, rewrites or code snippets will not notice the difference, and that the reviews calling it 'excessive' describe with precision the smallness of the task posed. He also clarifies that this capability cannot be reconstructed from a system prompt nor by stacking smaller models, as he assures after having tried.

Beyond the specific case, the article points to an underlying debate increasingly present in the agentic AI community: as models autonomously carry out multi-step tasks, the competitive advantage stops being about knowing how to program the machine and shifts to knowing how to rigorously define what work it is handed. It is a shift that is both optimistic and demanding. Optimistic because it suggests that human value migrates toward conception, judgment and verification; demanding because it requires developing a new literacy. The sentence with which Nate closes his diagnosis sums up the mood of the moment well: people are not tired of AI, they are tired of being told it is amazing while their real experience remains small.