The Last Interface

Consider how far the input has traveled. We flipped switches and fed punch cards. We wrote assembly, mnemonic by mnemonic, and carried programs on floppy disks that held less than a single photo does today. Then high-level languages let us describe logic instead of registers. The GUI let us point instead of command. Search let us ask instead of navigate. Autocomplete finished our sentences. Then, quite suddenly, we stopped instructing computers at all and started talking to them — and in early 2025, "vibe coding" entered the vocabulary: describe what you want in plain language, and code appears. Within months, even that felt slow. Agent harnesses now take a two-line instruction and work for hours — planning, writing, testing, correcting — while you do something else.

And the next step is already shipping. Meta's Neural Band — a wristband reading the electrical signals your muscles fire — launched alongside its Ray-Ban Display glasses, and at CES 2026 Meta announced neural handwriting: write on any surface, or barely move at all, and text appears. This is not science-fiction brain reading — it's peripheral neuromotor signal, trained intent. But the direction is unmistakable: the cost of telling a machine what you want is heading to zero.

Here is the question that trajectory forces, and it isn't a technology question. When articulation is free and execution is delegated, what exactly is left for the human to contribute?

The answer: everything that was always the hard part. What you choose to want. How clearly you conceive it. What you predict will come back. What you accept. What you're willing to sign your name to. AI will do what you tell it — which means the quality of what you tell it is the quality of your thinking, and the quality of what you accept back is the quality of your judgment. Thinking is the last interface. It always was; the other layers just hid it.

The uncomfortable data

You would hope that as AI got better, human thinking would sharpen alongside it. The early evidence points the other way.

MIT Media Lab wired essay writers with EEG and found what they called "cognitive debt" — measurably reduced neural engagement, and less ownership of AI-assisted work. A Microsoft and Carnegie Mellon study of knowledge workers found that the more people trusted AI, the less critical-thinking effort they invested — confidence in the machine substituting for evaluation of it (summarized by Duke's teaching center). EDUCAUSE called the pattern "better results, worse thinking". And KPMG's 48,000-person global study found 66% of employees don't evaluate AI outputs before using them — while 57% hide their AI use from their employers. Researchers are now publishing frameworks for protecting human cognition in the GenAI era — the atrophy risk has become a field.

Read those together and the picture is stark: organizations are buying millions of AI seats while the skill that determines whether those seats create value or ship confident nonsense — human judgment — is quietly deteriorating from disuse.

Thinking fast, thinking slow — at the gates

Daniel Kahneman gave us the vocabulary: System 1, the fast, automatic, pattern-matching mind; System 2, the slow, deliberate, effortful one. Fast thinking isn't the villain — expertise is trained intuition. The problem is where the fast mind runs unsupervised. Agentic work has three such moments, and they repeat all day:

Delegate — before the agent runs. The System-1 failure: vague intent, no success criteria, unbounded scope. The trained move: frame it, bound its blast radius, and make one falsifiable prediction — what do I expect back, and how sure am I?

Accept — when the output arrives. The System-1 failure is the most dangerous one: polished output feels correct. Fluency masquerades as truth. The trained move: verify before adopting, and reconcile against the prediction you made — surprise is information.

Deploy — before the deed. Merge, send, spend, publish. The System-1 failure: "the agent did it." The trained move: a premortem on anything consequential, a reversibility check, and the oldest rule of professional life — you sign your work.

None of this is trainable by watching lectures about bias. The brain-training industry learned that lesson expensively — the FTC fined Lumosity $2 million for claiming puzzle gains transfer to life. They don't. What does work is deliberate practice on your real work with feedback — and one skill above all has robust evidence behind it: calibration. Philip Tetlock's forecasting research showed that people who practice making probability predictions and confronting the outcomes genuinely, durably improve their judgment — the science behind Good Judgment's superforecasting programs. Until now, that training happened in workshops, on geopolitical questions. It has never been available on the work you actually did today.

What we're building

That's the gap Miatz Tuukul closes — tuukul is Yucatec Maya for thought, and it's the thinking strand inside miatz's AI-Fluency program. A daily check-in where you speak one intent and one prediction. A coach that lives inside the tools you already use — Claude Code, Cursor, your IDE — asking why at the three gates, never doing the thinking for you. A personal calibration curve computed from your real AI work — when you say "90% sure," how often are you actually right? — something almost no professional on earth has ever seen about themselves. Drills on your own backlog, never puzzles. A journal that writes its own first draft from your sessions. And a credential — Itz'at AI-Fluency — whose higher levels require observed practice and a spoken exam, because you cannot memorize your way into a calibration curve. All of it private by architecture: your prompts stay on your machine; only skill signals sync; your coach reports to you, not to your boss.

The interfaces will keep dissolving — keyboard into voice, voice into intent, intent into agents that act. Every dissolution raises the stakes on the one layer that remains. When a two-second gesture can launch a four-hour agent run, the ratio of consequence to deliberation becomes the highest in the history of tools. That's not a reason to fear the band on your wrist. It's a reason to train the mind behind it.

When the interface disappears, thinking is all that's left. Train it. → [Try the free Calibration Lab](/play/calibration-lab) · [Join the Tuukul early-access list](/tuukul)