What services does Young Tsai offer?

A 1-week prototype sprint, a 4-week MVP, and 2–4 week AI strategy / team training.

How long does an AI MVP take?

Four weeks: from requirements to a deployed, tracked product handed off to your team.

Do I need an engineering background to work with Young?

No. Young teaches you to let AI build for you — shipping a working system with one tool, Claude Code.

How do I book a consultation?

Book a free 30-minute call on Calendly, or email young.tsai.9765@gmail.com.

Three Years of AI in Re-Education — From Khanmigo to Claude Learning Mode

In three years AI users went from 0 to 18M, but even the creator said "the revolution hasn't happened"

On April 9, 2026, Sal Khan said something in a Chalkbeat interview

He said Khanmigo was basically a "non-event" for students — they just were not using it much

That made me pause

Because at the same moment, Khanmigo's numbers were:

18M students globally
350 US school districts using it daily
Across 34 languages
Integrated with both Microsoft and Google

Scale exploded, but PMF still did not click

I have spent almost three years doing AI projects around education in Taiwan, and this contradiction kept spinning in my head

This post is my notes on the last three years of edtech AI evolution, and also prep work for a short talk I recently gave

What happened in these three years

2023 was the first GPT-4 wrapper wave

In the same month GPT-4 was released, Khanmigo and Duolingo Max both launched as AI tutors built by "wrapping GPT-4"

Back then everyone asked the same question: will AI replace teachers

2024 was when the platform war started

Microsoft integrated Khanmigo and offered it free to educators worldwide

Khanmigo K-12 users jumped from 40,000 to 700,000 in one school year

Duolingo Max added roleplay + explain-my-answer, targeting advanced learners at $30/month

Edtech players in many countries started testing the waters, including Junyi in Taiwan

2025 brought a qualitative shift — LLM companies entered directly instead of waiting for edtech companies to call their APIs

Anthropic launched Claude for Education in April

Learning Mode expanded to all users in August, not just education plans

Google started pushing Gemini for Education

The game shifted from "edtech companies using big-tech APIs" to "big-tech companies shipping education products themselves"

2026 shifted the battlefield again

Anthropic + Teach For All — 100,000 educators / 63 countries / 1.5M students
Google + ISTE — free Gemini training for 6M US K-12 teachers
Microsoft AI classrooms — broad rollout
McKinsey data: 78% K-12 + 92% higher ed using AI (vs 23% / 41% in 2023)

The battlefield quietly moved from "student-facing tutors" to "teacher AI enablement" + "classroom infrastructure"

Three player types, very different realities

Type 1: Student-facing AI teachers (Tutor products)

Representative players: Khanmigo, Duolingo Max, Quizlet AI

Scale is large, but PMF is uncertain

Problems:

Khanmigo's own founder admitted "the revolution hasn't happened"
Students do not open it proactively, friction is high
Direct answers create brain-rot risk (the more you use it, the less you think — this is not exaggeration, this is a design issue)
Teacher feedback is mixed: sometimes helpful, sometimes students regress

Type 2: Tool platforms (Tool + AI)

Representative player: Canva for Education + Magic Studio

Massive scale: 260M users / 190 countries / AI tools used 10B times / 15M paid + 150M users (including free K-12)

PMF exists

Why does this work while Khanmigo struggles

Canva students do not need to "want to study" first before opening it. They open it to finish assignments, slides, posters — and AI helps right there

AI becomes a productivity multiplier with immediate visible output, so motivation appears naturally

Learning is a byproduct, not the main character, and that distinction is critical

Type 3: LLM giants entering directly

Anthropic (Claude for Education led by Drew Bent) / Google (Gemini + NotebookLM) / OpenAI (ChatGPT Edu)

This is infrastructure-level scale with three legs: models, compute, and enterprise sales

Anthropic ARR was close to $19B in 2026, OpenAI over $25B

For content companies, competing head-on with that stack is very hard

There is also a fourth type: those who got eaten (the Chegg story)

I intentionally skipped one before talking about the three types

Chegg

The largest US homework-help / textbook-rental / tutoring platform, with a peak market cap of $14.7B in February 2021

On May 2, 2023, its CEO said one sentence on the earnings call: "ChatGPT is impacting new customer growth"

The stock fell 50% that day

By early 2026, the stock dropped to about $1, with only $156M market cap left

99% of market value evaporated in 39 months

Chegg did not ignore AI. In April 2023 it partnered with OpenAI and launched CheggMate (its own AI product)

But wrapping ChatGPT to compete with ChatGPT is a dead end

In 2024, Google AI Overviews took another bite out of Chegg's organic search traffic

In May 2025 they cut 22% of staff (248 people), and in October the same year another 45% (388 people)

Subscribers fell from a 7.8M peak to 3.2M

The most brutal survey number: students planning to use Chegg dropped from 38% to 30%, while those planning to use ChatGPT rose from 43% to 62%

The lesson is not "AI killed edtech"

The lesson is "old edtech dies fastest when 'we'll just add AI' is the strategy"

If your core value is paid answer-copying / homework-solving, ChatGPT free tier drives that value toward zero

This is different from Khanmigo. Khanmigo is scaling without clear PMF, while Chegg had its core business model hollowed out

What market data is saying

The AI-in-education market was $7.05B in 2025, $9.58B in 2026, and projected at $32B-$41B by 2030 (different firms estimate roughly 4-5x)

Overall e-learning market reaches $365B in 2026

Numbers look beautiful, but two anchors matter

First anchor: student use of ChatGPT is seasonal

OpenRouter data: during finals season on May 27, 2025, ChatGPT peaked at 97.4B tokens/day. In June (summer break), it dropped to 36.7B (down 54%)

Rutgers researchers analyzed 10,000 prompts and found strong school-calendar correlation; spring break + summer were dead zones

Meaning: a large part of ChatGPT traffic is homework copying

OpenAI burned $9B and made $4B in 2024 — already a cash-burn cycle. Layer in seasonality and the model gets fragile

Second anchor: growth is concentrated in tech giants + upgraded legacy edtech

The PMF window for new local entrants is already very narrow

Khanmigo has 18M users but still a "non-event", Chegg lost 99% of market value, Canva has 260M users but is not positioned as pure teaching

Differentiation is the only path — content IP / local verticals / tool-first entry points

Three insights from the contrast

Insight 1: The "answer machine" problem is not weak AI

Sal Khan's "non-event" comment reminded me of people I coached on Claude Code

When many people first try AI-assisted development, their instinct is to use it like Google: ask one question, get one answer

People who learn fast do not do that

They ask AI: "Don't give me the answer yet, guide my thinking"

The difference is not in AI capability, but in usage design

Khanmigo's root problem is design philosophy:

Drew Bent's (Anthropic Education Lead) corrective design:

This is Socratic interaction, also anti-brain-rot design

But this mode is opt-in and OFF by default, and most people do not know it exists. I used Claude Code for half a year before I noticed it myself

Insight 2: Canva's lesson matters more than Khanmigo's

Canva has 14x Khanmigo's users (260M vs 18M), yet almost nobody seriously analyzes Canva in AI-education conversations

Why

The takeaway for parents and educators: students learn better when they encounter AI while doing real tasks, not when they are told to sit down and "go learn AI"

This is not saying we should abandon active learning. It says AI intervention-point design matters

Insight 3: After big-tech entry, local players have fewer choices

LLM giants arrive with three legs: models, compute, and FDE sales

For local content players, only three paths remain:

Chinese language / local markets / family education are long-tail for global giants, but home turf for local players

I think Taiwan's education community still has not discussed this seriously enough

The person I most want to highlight: Drew Bent

I have followed many people in AI education, and Drew Bent is the most underestimated by far

His background is unusual:

MIT double major in Physics + CS
Stanford master's in education
Knight-Hennessy Scholar
Former Khan Academy engineer
Co-founded Schoolhouse.world with Sal Khan (170K students / 180 countries, free peer-to-peer tutoring)
Now Anthropic Education Lead

A few lines from his public interviews stuck with me:

"human to human interaction is at the core of education"

"personalized done right should be personal"

"bullish on tutoring as a way to have a human in each person's life"

He is not saying AI replaces human teachers. He is saying AI completes the stack so everyone can have a human-like companion

Claude Learning Mode's design rules come from that philosophy:

Rule	Behavior
No direct answers	Guide students through questions
Socratic interaction	"Do you think the core is X or Y?"
Anti brain rot	Prevent AI dependency
Code learning scenario	Use `TODO(human)` to force student-written code

How to enable it:

# Inside Claude Code
/output-style learning

# Inside chat.claude.ai
Select the "Learning" style in the chat UI

It is off by default, so you have to choose it yourself

I personally ran it once with "teach me what continual learning is"

The feel is very different from a normal answer machine. It is not AI dumping knowledge; it anchors from what I already know and guides me to articulate new concepts myself

After the session I noticed something: the parts I thought I understood became reinforced, and the parts I could not explain were forced into clarity

That feels like a different world from Khanmigo's "non-event"

Taiwan perspective: Jian-Lifeng sees the same thing

While writing this, I kept asking why Drew Bent's design philosophy is rarely discussed seriously in Taiwan

After reviewing Jian-Lifeng's talks from the last two years, I realized he and Drew Bent are pointing in the same direction from different angles

A few of Jian-Lifeng's lines are best quoted directly

On learning and questioning:

"One Q&A is plagiarism, ten Q&As is learning, a hundred Q&As is creation"

"The answer comes from smart ask"

"Good answers come from smart questions. If you don't understand fundamentals, you can't guide AI to solve problems"

These align directly with Learning Mode — AI is not an answer machine, it is a partner that helps you ask better

On thinking and the brain warning:

"Don't outsource your brain to AI"

"Always think for 10 minutes yourself before asking AI"

This is the Taiwan version of the brain-rot warning, just more direct

On talent structure, he uses the "pi-shaped talent" idea:

"Talent needs two legs like the Greek letter pi: one leg is domain depth, one leg is AI collaboration ability, and the top bar is breadth of knowledge"

"Breadth matters more than depth" (breadth determines how well you can ask)

"The passing line has moved from 60 to 80" (AI can already handle 60-level work, so the human baseline moved up)

His two hardest lines about education:

"We are all AI immigrants. The true natives are not born yet"

"Schools teach knowledge AI already knows, while companies need what AI still cannot do but you must understand (tacit knowledge)"

The first line admits everyone is still figuring it out

The second is a fundamental challenge to schools — if schools only teach open knowledge (the part AI learns fastest), students graduate as a "lost generation" (his phrase)

His societal warning:

"AI may create tool-using superhumans where 1% capture 99% of wealth and opportunity"

It sounds extreme, but look at Khanmigo's 18M users still being a non-event, Chegg losing 99% market value, and ChatGPT summer traffic dropping by half. The 1% vs 99% split is not a prediction, it is already emerging

So what does this mean for you

If you are an educator or parent:

Do not use AI as an answer machine — ask, answer, forget, ask again is exactly the brain-rot loop. It is real, not panic talk

Try Claude Learning Mode — turn it on when learning, turn it off when executing

Let students encounter AI while doing real work. Do not make AI a separate place they must deliberately "go study"

Teachers' roles are shifting too: from "teaching knowledge" to "designing game rules" so students + AI can produce meaningful outcomes

If you build education products or courses in Taiwan:

Content IP is the moat, not just the content platform — big-tech cannot replicate your 20 years of local nuance

Do not build the Nth tutor product. If even Sal Khan says it struggled, you are unlikely to beat that

Learn from Canva, not Khanmigo — tool-first, learning as a byproduct

Partner with big tech strategically instead of competing with them — Anthropic Education and Google Education both need vertical localization partners. Chinese markets are your home turf

Closing

Three years ago everyone asked whether AI would replace teachers

Now the question is whether teachers who use AI will replace teachers who do not

Khanmigo proves scale does not equal PMF

Canva proves tool + AI can land earlier than tutor + AI

The Learning Mode pushed by Drew Bent at Anthropic is still the closest design I have seen to "AI making people better thinkers" — but it is opt-in, and you have to turn it on yourself

That opt-in design is interesting in itself

You know it exists, but you still have to choose it

Like many things

One last time, Drew Bent's line:

"personalized done right should be personal"

In the AI era, personalization is not throwing you to an algorithm. It is preserving your own learning path and thinking style

Sources / Further Reading

Written on 2026-05-09 as prep work for a short talk. AI in education changes fast, so the scale numbers in this post will keep moving. The point is not the snapshot numbers, but the design philosophy differences across these player types. It will be interesting to revisit Khanmigo / Drew Bent / Canva PMF evolution in six months

Related posts