An AI tutor that deliberately does not solve the problem for the student

The most interesting part of Knowable isn’t the promise to “understand any school subject.” It’s an attempt to make AI less convenient: Milo is designed to help students arrive at the solution on their own, rather than instantly generating an answer for their homework.

The user places their notebook in front of a Mac, turns on Desk View, and asks a question by voice or via chat. The system analyzes the image of the page—including formulas, diagrams, printed text, and handwritten notes—and then responds with a hint. In a demonstration involving an integral, Milo does not continue the calculation but asks what will happen if the upper limit is substituted.

This is a significant shift in the interface. A typical AI tutor sees only what the student has managed to type. Here, the entire page becomes the context: the problem statement, intermediate steps, and the mistake the user hasn’t yet articulated in words. For math and physics, this is potentially more useful than a long text-based query.

But “seeing the notebook” doesn’t mean automatically understanding the thought process. Milo receives visual context and responds only after the student asks a question. This means the product doesn’t monitor the session like an independent tutor and doesn’t intervene without being prompted—it’s more like a camera above the desk plus a conversational layer on top of it.

There’s a notable limitation: currently, you need a Mac running macOS Ventura or later. The built-in camera in Desk View mode is used, and support for iPhone via Continuity Camera is also announced. For families where the main computer is a Windows PC or Chromebook, this isn’t a universal tutor, but rather a fairly specific use case for the Apple ecosystem.

There’s also a less obvious trade-off between convenience and privacy. Knowable states that camera footage is stored only in memory during the session and isn’t saved to disk or a database. However, chat messages and session metadata are stored in AWS DynamoDB, and video frames, voice transcriptions, and chat are transmitted via TLS to AWS Bedrock in the us-east-1 region. This doesn’t look like an “all-local” setup, and parents should read this specific section—not just the promise that video recordings aren’t saved.

The pricing model also ties usage to the frequency of classes: the free tier provides 1,000 credits per month, while the Plus plan offers 10,000 credits for $24.99 per month or $249.99 per year. The company does not explain in the provided material how many typical questions or minutes of conversation correspond to one credit, so it’s difficult to estimate the actual cost of regular lessons in advance.

The main criterion for a product like this isn’t how effectively it recognizes handwritten text. What’s more important is whether it adheres to its own rule of “hints, not answers.” If a student can consistently coax a ready-made solution out of it by asking a few leading questions, the limitation will remain nothing more than a marketing slogan. If, however, Milo truly maintains the right level of assistance, the visual context transforms the AI from a solution generator into a tool for real-time feedback.

For now, Knowable doesn’t look like a replacement for a teacher, but rather a carefully defined learning environment: a Mac, a camera, a question from the student, and limited intervention from the model. Its value will be determined not by the breadth of its subject list, but by how well the system can stop one step short of providing an answer.

Source: Hacker News - Newest: ""AI" "LLM""