What comes next — multimodality, reasoning models and embodied AI

Short answer: the useful way to read future trends is not to predict what will happen but to separate what you can buy today from what may mature next year and what is still a research topic. This lesson sorts the three and gives you criteria that will not go out of date.

It starts with a vendor deck

A principal came back from an education technology exhibition with a slide deck containing three words: multimodal, reasoning, embodied AI. He asked the IT coordinator: "Should we be doing these three this year?"

The right answer is not yes or no. It is that the maturity gap between those three words is far too wide for them to belong on the same slide. One is already running in your school, one should be applied selectively by task, and one is still in the laboratory.

Here is each in turn.

Already in use: multimodality

Multimodality simply means the model handles more than text: it can look at images, listen to speech and read handwriting. This is not a future prospect; it already runs in classrooms.

Examples that work in Hong Kong schools today:

• Speech. Students read English or Putonghua passages aloud and the system transcribes them and gives pronunciation feedback; teachers dictate lesson plans. Cantonese recognition now holds up with primary pupils reading aloud, unclear articulation and background noise.
• Images. A teacher photographs a worksheet page and asks for variant questions of the same type, or photographs whiteboard notes and gets a handout outline.
• Handwriting. Scanned handwritten answers are converted to text and passed into the marking flow. Accuracy still depends on the handwriting, so teacher review is needed.

The limits deserve stating plainly: image understanding still makes mistakes on complex charts, mathematical notation and messy handwriting, and it sounds just as confident when wrong. Wherever marks are affected, teacher review cannot be dropped.

Maturing now: reasoning models

What distinguishes a reasoning model is that before answering it produces an internal chain of working and then states a conclusion. Picture a student who has stopped shouting out the first answer and now works it through on rough paper first.

That brings three concrete changes:

• Clear gains on multi-step problems. Mathematics word problems, logical reasoning and debugging all show lower error rates.
• Slower and dearer. Thinking longer means using more computation. For "tidy up this English sentence" it is pure waste.
• Adjustable thinking time. Current models generally let you set an effort level, from a quick answer to deep deliberation.

One common misunderstanding is worth clearing up: the reasoning a model shows you is not necessarily what it internally computed. It is a helpful piece of text, usually consistent with the conclusion, but it should not be treated as proof. A teacher who reads a rigorous-looking chain of reasoning still has to check the conclusion itself.

The practical implication for schools is simple: choose by task. Composition feedback, circular drafts and vocabulary exercises go to something fast and cheap; step-by-step mathematics and complex administrative calculations go to a reasoning model. For the cost side, see context caching and cost control.

Still research: embodied AI

Embodied AI means putting AI inside something with a body — a robot, a robotic arm, an automated device — so it perceives its surroundings, plans actions and carries them out.

The state of play, stated accurately: this is one of the most active research and demonstration areas, and several laboratories have shown robots carrying out everyday manipulation from natural language instructions. But the distance from a laboratory demonstration to something stable, safe and affordable in a school environment is large, and nobody can honestly tell you how many years it will take.

Its meaning for schools today comes down to two things:

1. As teaching material. STEM and technology education can use it to teach perception, planning and control, which is solid curriculum content.
2. As training in judgement. Having students compare a demonstration video with a purchasable product is an excellent information literacy exercise.

As for procurement, writing it into a three-year plan now is not defensible.

An example from the classroom

The IT team at one secondary school was asked to respond to that vendor deck. Rather than writing a trend analysis, they produced a three-column table that answered the principal's question directly:

The valuable part of the table is the third column. Most school AI projects fail not because the technology was wrong but because the problem it solved was one the school never had.

Three criteria that will not go out of date

The technology changes every year; these three questions keep working:

1. Can I buy it today and afford it? Not "has anyone demonstrated it" but "are schools using it, what does it cost, and who maintains it".
2. What happens when it fails? An oral practice error makes a student laugh; a marking error affects university applications. The higher the stakes, the greater the share of human review.
3. Does my school actually have this problem? If not, even excellent technology is a burden. Start from the pain point, not from the technology.

And one more for procurement: the contract must let you take all your data and leave whenever you want. Betting on any vendor's roadmap carries risk; a contract you can walk away from does not.

What this means for your classroom

1. Multimodality can be trialled this term. Oral practice and image-based question setting are the two easiest entry points with the most visible results.
2. Do not buy the strongest model for every task. Matching model to task is how you cut cost and raise speed, and it is something the platform should be doing for you.
3. Treat future trends as information literacy material. Teaching students to tell a demonstration video from a product you can buy is one of the most practical skills of this era.
4. Do not pay for a roadmap. Pay only for capability you can use today, and make sure you can leave with your data at any point.

How this works inside Edor.ai

We describe only what is live and make no commitments about unreleased features.

On multimodality, the platform already uses speech recognition and text-to-speech for Cantonese, Putonghua and English oral practice, and supports image input for understanding teaching materials and student work. On model selection, the platform allocates by task difficulty: day-to-day teacher tools and student conversations run on cost-efficient tiers and only escalate to a reasoning model for long reasoning tasks, which is why a school pays a fixed annual fee rather than a fluctuating inference bill. On embodied AI, we have no product and will not write one into a school's plan.

One further point that has nothing to do with trends and everything to do with risk: the data belongs to the school, with a free one-click full export (JSON, CSV or original files) and no lock-in. If something better suits your school in three years, you should not be staying because the data will not move.

The course ends here

From lesson 1 to this page, you have covered how a model works inside, the three stages of training, the ecosystem and how models are obtained, the core concepts of building applications, local deployment and efficient fine-tuning, and safety, evaluation and ethics.

You should now be able to explain to a colleague why AI gets things wrong, judge whether a vendor's safety claims are specific, read the limitations of a leaderboard, and decide what data must never be pasted in.

Two suggested next steps.

To revise or fill gaps, go back to the LLM Classroom syllabus or the Learning Centre.

To turn this understanding into classroom output, continue with the Prompt Engineering Classroom — eighteen lessons set entirely in Hong Kong classrooms, beginning with what makes a good prompt and covering planning, tiered worksheets, question setting, marking feedback, SEN adaptation and panel template libraries, each with copy-ready prompts in both languages. The judgement taught here plus the technique taught there is the complete set.