Picture a flight instructor sitting in the right seat while you fly a practice pattern. She is not grading you at the end. She is watching your hand on the throttle right now, listening to your radio calls as you make them, and correcting your descent rate before you overshoot the numbers. That continuous, present-tense attention is what makes a good instructor different from a good textbook.
When we started building Jade Labs, that was the problem we were actually trying to solve — not "how do we deliver content online" but "how do we put that right-seat instructor inside a browser tab." The answer turned out to require 53 distinct AI teachers, each with a different voice, a different domain, and a different model of what a mistake looks like in their subject.
This is the story of how we built them and why each design choice went the way it did.
The first prototype used one general-purpose AI that could answer questions about anything. It worked fine as a question-answering tool. It failed as an instructor.
A chemistry student titrating an acid-base solution does not need the same kind of feedback as a pilot managing an engine failure or a learner trying to hold a conversation in Mandarin. The mistake in titration is adding a half-drop too many and watching the endpoint colour overshoot. The mistake in an engine failure drill is reaching for the wrong checklist item first. The mistake in Mandarin is a tonal error that changes the word entirely. These are not variations on a single feedback template. They require fundamentally different models of what correct performance looks like.
A single general AI kept producing feedback that was technically accurate but instructionally useless — the equivalent of a tennis coach saying "your swing was suboptimal" without specifying whether your grip, your footwork, or your follow-through was the actual problem.
So we stopped trying to build one teacher and started building many.
When we say each AI teacher has its own voice, we do not mean a different text-to-speech avatar with a different name. We mean something more structural.
Each teacher was built with:
The Language Lab teacher, for example, knows that a false cognate error in Spanish is usually a vocabulary gap, but a consistent subject-verb agreement error usually signals a gap in grammatical intuition that vocabulary drilling will not fix. It treats them differently. The Aeronautics Lab teacher knows that a student who consistently floats long on landing may be carrying too much power, flaring too early, or misjudging ground effect — three different root causes with three different corrections.
Building that specificity for 53 subjects took considerably longer than building the browser interface did.
The pillar that surprises people most is that the AI instructors listen in real time. Not "submit your answer and wait for a grade." Genuinely listen, the way a person in the room with you listens.
In the Language Lab, you are holding a spoken conversation. The AI instructor is processing your speech as you produce it — not after you hit a stop button. It hears the tonal error on the third syllable of a Mandarin word while you are still finishing the sentence. By the time you have completed the sentence, it has already decided whether to interrupt immediately (because the error will confuse the conversation partner) or to let the exchange finish and then loop back (because correcting mid-sentence would break your fluency more than the error itself did).
That interrupt-or-wait decision is not random. It is governed by rules we built specifically for spoken language acquisition, based on what the research literature actually says about corrective feedback timing in second-language learning.
In the Aeronautics Lab, the listening is not to speech but to control inputs. The AI instructor watches your throttle position, your pitch attitude, your airspeed trend, and your heading deviation simultaneously, the same way a CFI watches the instrument panel and the outside world at once. If your airspeed is decaying toward stall speed on a base-to-final turn, the instructor does not wait until you have stalled. It speaks before the event, not after it.
In the Typing Tutor, the feedback is nearly instantaneous — a keypress error is flagged before your fingers have left the key. The instructor tracks not just accuracy but rhythm, because uneven rhythm at low error rates often predicts accuracy collapse at higher speeds.
The technical infrastructure that makes this possible is not the interesting part of this story. The interesting part is that each instructor needed its own definition of what "real time" means. For typing, real time is measured in milliseconds. For a language conversation, real time is measured in conversational turns. For a chemistry titration in the Science Lab, real time is measured in the colour change you are watching in the flask.
There is a specific reason we built the feedback to be immediate rather than deferred, and it is not because immediate feedback always feels better. Sometimes it does not.
Deferred feedback — the kind you get from submitting an assignment and reading comments the next day — has one serious structural problem: the learner has usually moved on. The mental state in which the mistake was made is gone. The correction arrives as abstract information about a past event rather than as a signal that modifies ongoing behaviour.
Immediate correction lands in the context where the error occurred. You are still holding the pipette. You are still in the turn. You are still mid-sentence. The correction can actually change what you do next.
The design challenge is that immediate correction done badly destroys confidence and disrupts flow. A system that interrupts every minor deviation produces learners who freeze before every action, waiting to be corrected rather than developing judgement. So the 53 instructors each have calibrated thresholds:
In the Quantum Core lab, a learner building their first Bell state circuit will make gate-ordering errors that, in a real quantum computer, would produce a maximally mixed state instead of entanglement. The instructor flags that immediately because leaving it uncorrected would mean the learner runs the circuit, gets garbage output, and has no idea why. But if the same learner uses a slightly non-standard notation when labelling their qubits, the instructor notes it and mentions it at a natural pause — it is not worth breaking concentration over.
Every lab in Jade is built around a performance, not a lecture. This is a deliberate constraint, not a feature we added.
The Robotics Lab does not have a video explaining how to wire a motor driver. It puts you in front of a wiring diagram and asks you to connect the components. The Medical Lab does not show you a labelled diagram of the cardiovascular system. It asks you to name and locate the structures as you encounter them in a working simulation. The Cosmos Explorer does not narrate a tour of the observable universe. It gives you a real star catalogue and asks you to navigate it.
This is not because lecture video is bad. Lecture video is excellent for some things. But lecture video cannot teach you to do something. It can only teach you that you watched someone else do something, which is a different cognitive event entirely.
The reason this matters for the AI instructor design is that performance-based learning creates observable behaviour. The instructor has something to watch. A learner reading a paragraph gives the AI nothing to work with. A learner pipetting a solution, flying a heading, conjugating a verb, or soldering a connection gives the instructor a continuous stream of data about what the learner actually knows and can do.
The certificate at the end of a Jade lab is not a certificate for completing the lab. It is a certificate for passing a performance standard. In the Language Lab, that means holding a real conversation and passing a spoken exam. In the Aeronautics Lab, that means flying a full pattern, handling a simulated engine failure, and completing a checkride brief. The certificate means something because the standard it certifies is explicit, and the AI instructor was watching the whole time.
The number is not arbitrary, and we are not done.
Each of the 53 AI instructors maps to a specific domain, and within each domain, a specific performance outcome. The Aeronautics instructor is not the same as the instrument flying instructor, even though they overlap in knowledge. The Spanish conversation instructor is not the same as the Mandarin instructor, because the error taxonomies are genuinely different — tonal errors do not exist in Spanish; grammatical gender errors barely exist in Mandarin.
The current labs — Language, Typing, Aeronautics, Quantum Core, Robotics, Science, Cosmos Explorer, Medical — each contain multiple instructors. The Science Lab alone has instructors specialising in titration, reaction kinetics, and molecular biology, because the performance model for reading a titration endpoint is not the same as the performance model for interpreting a gel electrophoresis result.
Building a new instructor requires three things that take real time:
1. Domain mapping — working with people who actually have expertise in the subject to build the skill tree and error taxonomy, not just scraping curriculum documents
2. Performance calibration — defining the pass standard clearly enough that the AI can evaluate it consistently, which is harder than it sounds in subjects with any ambiguity
3. Feedback testing — running learners through the instructor and identifying where the feedback is too harsh, too lenient, too late, or missing the actual root cause
The 53 we have now represent subjects where we were able to do all three rigorously. The ones we have not released yet are in various stages of the calibration and testing process.
It is worth being honest about what this approach does not do well, because the genuine strengths of other approaches matter.
Coursera and edX have deep relationships with universities and a breadth of accredited content that we do not. If you need a credential that a hiring manager will recognise because it comes from MIT or Stanford, those platforms are the right choice.
Duolingo has extraordinary distribution and a gamification system that is genuinely good at maintaining daily habit. If the problem you are solving is "how do I stay consistent over a long period," Duolingo has thought harder about that problem than we have.
Khan Academy's free access and curriculum depth, particularly in mathematics, is something we have real respect for. The explanatory content is excellent.
What those platforms have not solved — and what Jade is specifically built to solve — is the right-seat instructor problem. The continuous, present-tense, domain-specific correction that modifies your behaviour while you are performing, not after you have submitted something for review.
If the right-seat instructor problem is one you have been trying to solve for yourself — you want to actually be able to do something, not just know about it — the most direct next step is to pick the domain where you have the clearest performance goal and start there.
The Aeronautics Lab is a good test case if you want to see what real-time corrective feedback feels like at its most demanding, because the performance standard is the most unambiguous: fly the pattern correctly, handle the emergency correctly, pass the checkride brief. There is no partial credit for almost flying the approach.
But the same logic applies in the Language Lab, the Science Lab, or any of the others. The AI instructor is already there, already calibrated, and already listening. What it needs is for you to start doing the thing.