The method
Dilo is a competency-based loop. It drills a small chunk set until you can say it without thinking, forces it into an unscripted conversation in the same session, corrects you specifically, and, once the oral check ships, refuses to move you up a level until you can do the level's task out loud. That is Stevick's FSI "microwave cycle" plus the Peace Corps oral gate, delivered by an AI partner instead of a classroom.
There is no secret. There are four levers.
FSI and the Peace Corps take adults selected for the job, in class four or more hours a day in small groups, from zero to a working conversation in 12 to 30 weeks, and about two-thirds of FSI's full-time students reach their goal. Their own documents describe no trick. They describe volume, structure and a test you can fail.
| Lever | Who has it today | Gap |
|---|---|---|
| Dosage, four or more hours a day | Nobody in consumer apps | Partly replicable with a deadline and a daily structure |
| Drill welded to immediate use in the same session | Drill apps have the drill half. AI conversation apps have the talking half. | Nobody welds them |
| Accurate corrective feedback in interaction | AI apps, but shallow and often wrong | The hard part, and the whole bet |
| An oral test that gates progress | Nobody | The clearest structural gap in the market |
FSI's finding, in their words: focused practice of some kind, including drills, appears necessary for almost all language learners to develop confidence and automatic language use. Peace Corps volunteers get 10 to 12 weeks, about four hours a day in small groups, then a live oral interview, typically at Intermediate-Mid, before swear-in. One audited cohort, learning Tagalog: 58 of 63 reached Intermediate-Mid.
One cycle, 30 to 45 minutes
- Introduce. Six to ten high-frequency chunks for one real task, with a grammar note of under two minutes. Logic, not tables.
- Drill. About ten minutes, spoken only. Prompt, pause, produce, confirm. Then substitution and transformation: swap the noun, change the person, make it a question. Speech recognition checks what came out of your mouth.
- Use. Ten to fifteen minutes. An unscripted roleplay of that exact task with an AI partner who stays in Spanish, knows your real name, city and plans, and pushes for output.
- Feedback. About five minutes. What you said, what we would actually say, why. Chunks you fumbled go back into the drill queue.
- Review. Spaced repetition of earlier chunks, spoken. Roughly seventy percent review to thirty percent new across a session.
The step from mimicry to communication is short on purpose. Drilling for weeks before you are allowed to talk is how most courses lose people.
The gate
Between levels there will be a three to five minute unscripted oral check scored against explicit can-do criteria, such as order food and handle a problem with the order without switching to English. No streak, no XP path around it. Miss it and you drill what fumbled, then try again straight away.
The beta ships without the check. Levels 1 and 2 are open; the gate arrives after we have measured the loop. We say this on every page because a screen that implies a working feature is a lie.
The ladder
- Introduce yourself, ask and answer basic personal questions
- Order food and drink, handle a problem with the order
- Get around: directions, transport, tickets
- Shop and negotiate: market, pharmacy, phone plan
- Make and change plans with a friend
- Talk about your past week and your plans
- Describe a problem and ask for help: landlord, doctor, lost item
- Hold a ten-minute open conversation on a familiar topic
Vocabulary across the ladder is 1,500 to 2,000 words, verbs and function words first, learned inside phrases you will actually say. The top 1,000 Spanish words cover about 88 percent of speech; the top 2,000 about 93 percent.
Honest hours
B1 takes roughly 350 to 400 hours. At one hour a day that is about a year; at ninety minutes a day, about eight months. Intensive beats extensive: the same hours in five weeks outperform them spread over seven months. Dilo will not tell you it is faster than that, and it will not sell you a streak to hide the number.
What we have not proved yet
Feedback accuracy. Expert testers have caught AI apps correcting things that were not wrong and missing real errors. Before launch we blind-rate a sample of Dilo's corrections against a human tutor and publish the agreement rate, whatever it is. Marketing claims wait for that number.
Get a seat
You are on the list.
We keep your name, email and phone number for the beta invitation and the notes, nothing else. The invitation comes by email when a seat opens.