When the AI Giants Start Talking About Pacing, Education Needs Pacing Even More
Illustration · not a service record“When the AI Giants Start Talking About Pacing, AI Adoption in Taiwan’s Education Needs Pacing Even More”
Dario Amodei, Sam Altman and Elon Musk, leaders who compete fiercely with one another in AI, have recently arrived at a rare consensus: the capabilities of frontier AI are growing fast, and safety measures and society’s readiness need time to catch up.
Dario put forward “Pace the Frontier,” arguing that the speed at which frontier capabilities advance should be restrained, so that safety research, independent evaluation and governance have more time. Sam and Elon expressed support one after the other, but they still disagree about how to slow down and who should oversee it.
[Dario’s original post](https://darioamodei.com/post/we-must-pace-the-frontier)
When the people at the very front start talking about speed, it is something everyone should take seriously.
Yet the tech leaders’ appeal soon ran into political reality.
On September 14, on stage at the All-In Summit in Los Angeles, Jensen Huang took a phone call from Trump and put it on speaker for the whole room to hear. Trump rebutted the AI doomsday narrative and questioned whether the calls to slow down might carry political motives, and might even benefit China. Jensen Huang, for his part, stressed that rapid innovation and safety can go hand in hand, and that companies should take responsibility.
From China, too, came voices opposing this case for slowing down. Chinese researchers and state media questioned whether it might entrench the technological advantage the United States already holds and make it harder for latecomers to catch up.
These responses show us how many sides a single problem has: even when they agree that safety matters, all parties still worry that they will slow down first while others keep accelerating.
Who sets the rules? Who does the verifying? After a slowdown, who keeps the lead, and who loses the chance to catch up?
Once a technological risk enters politics, it becomes at the same time a question of trust, power and competition between nations. We can hardly expect the world to stop in an orderly way and wait until everyone is ready.
But the economy has a different speed.
A recent LinkedIn article offered a reminder: AI keeps getting stronger, yet the value that companies create does not necessarily increase in step, because the bottleneck may already have shifted to the “absorptive capacity” of people and organizations.
[Original article](https://www.linkedin.com/....../ais-alignment-problem......)
A new model can go live overnight. For a company to change its workflows, connect its data, redivide the work and build trust often takes longer.
The “productivity J-curve” proposed by Erik Brynjolfsson describes exactly this kind of phenomenon: a new technology needs complementary investments in talent, processes and organization before its benefits have a chance to be realized step by step.
This time lag makes me think of education.
While the economy is still digesting AI’s advances, education has an opportunity it must seize: to prepare actively and turn the capabilities of technology into the capabilities of people.
This opportunity does not rest on the assumption that “AI will definitely slow down.” The tools we already have deserve to be put to serious use, and whether models move fast or slow, education needs to build its own absorptive capacity.
What is more, education itself is raising the workers, citizens and creators of the future. Whether education can digest AI will affect whether society as a whole can make good use of AI.
I believe we should advance two things at the same time: bring education into AI, and bring AI into education.
▍First, bring education to AI and develop Pedagogical AI.
Being able to answer questions and knowing how to help people learn are two different abilities.
Teachers know that students sometimes need a demonstration, sometimes need a hint, and sometimes need to be asked a follow-up question: “How were you thinking about that just now?” And sometimes they need to quietly try again on their own.
These judgments should go into the design of AI: how to understand a student’s level, identify misconceptions, provide the right amount of help, and confirm whether the student has really learned.
A number of studies with strong evidence have already found that AI can help students do better on their assignments, yet does not necessarily improve their performance once the tool is taken away. Other research shows that AI tutors combined with instructional design can improve learning in specific settings. Together, these results remind us that what needs to be examined is the specific instructional design and the learning outcomes.
For education AI to mature, therefore, teachers, learning science researchers and engineers need to work together.
Teachers say where children get stuck; researchers design ways to verify learning; engineers turn methods that work into products that can be used reliably; and then feedback from the classroom keeps correcting the next version.
This is the deeper meaning of “bringing education to AI”: letting educational expertise take part in defining what good AI really is.
Junyi Academy is continuing to develop several education AIs. On the one hand it makes sure of basic safety and learning effectiveness, but to truly solve the problems in schools, it will need more feedback from students, teachers and parents.
▍Second, bring AI to education, so that teachers and schools are able to take it on, and so that children grow the abilities to face the future.
I hope every teacher will have the chance to first get familiar with one major general-purpose AI. ChatGPT, Claude, Gemini: just pick one to start with. Of course, it would be even better if the whole school had one together.
Start with one real piece of work: organizing teaching materials, adjusting reading difficulty, discussing students’ mistakes, or helping to prepare administrative documents. Gradually come to understand how to ask, how to verify, what can be handed over to it, and which judgments you still have to be responsible for yourself.
For this to happen, schools need to provide time, tools and peer support. Only when teachers feel helped in the course of using it does change have a chance to last.
Next comes the digitization of administrative and collaborative lesson-planning systems.
When curriculum plans, collaborative lesson-planning records, teaching materials and administrative rules are still scattered across paper, personal computers and chat groups, it is hard for AI to grasp the school’s context.
We need to organize the data so that it can be found and updated, with clear versions and clear permissions, and at the same time review duplicate reporting and unnecessary procedures.
Only then is there a better chance of letting AI, through a suitable Connector / MCP and within the scope it is authorized for, connect to the school’s existing data and systems.
When a teacher prepares a unit, AI then has the chance to refer to the course schedule, the jointly prepared materials and the teaching objectives, and to offer help that fits the classroom more closely. Data quality, personal data protection and system compatibility also have to be done properly along the way.
What really matters is not fancy technology, but whether these changes in the end truly reduce teachers’ stress and save them effort, especially at a time when people’s willingness to become teachers is falling.
If AI produces ten more reports and teachers then have to read ten more reports, we still need to keep revising. Once the time spent verifying, operating and reading is all counted in, whether teachers have more room to understand their children is the criterion that matters.
As teachers gradually become familiar with it, schools gradually get ready, and Pedagogical AI matures through feedback, we can step by step expand the applications that are suitable for students.
Of course, we can already begin now with supported, small-scale trials.
Abroad there is Khanmigo; in Taiwan there are the Junyi AI Notebook, e-Du (e 度) on the Taiwan Adaptive Learning Platform (因材網), and CoolE Bot on Cool English (酷英網).
Junyi also continues to explore Jutor, an English AI tutor; ActiveAI, an AI literacy tutor; PlayGo, a multi-faceted AI learning companion; and Lingoleap, a Chinese language arts AI tutor that is still a very early-stage experiment.
I have already started letting my own daughter try Lingoleap. The photo shows her using it.
As a father, and as someone involved in education products, what I care about is whether she is more willing to read, better able to understand, and more capable of expressing her own ideas. Having my own child start using it makes me care about these questions even more; the effects still need continued observation and verification.
This is a sweeping change to the system. On the one hand it needs to be pushed forward actively, but on the other it needs, even more, respect for the pace of adoption, because education really is hard to hand over to AI alone. Even tech giants such as Bill Gates, who has cared about K-12 education over a relatively long period, have set the tone this way.
How teachers prepare lessons, how administrative work is coordinated, how students ask for help, how assignments are designed and how learning is assessed may all change along with it.
A healthy Adoption Curve requires that those willing to try first, from Pioneers to Early Adopters, get support, that their experience is tested and revised, and that more people then come to see the value that relates to them. Schools with fewer resources need help all the earlier, so that they are not left behind in the transition.
Having only deadlines, usage rates and outcome requirements, without time, support and adjustments to processes, easily leads to a Hard Landing: the tools add pressure, and a single failure may also use up people’s trust in change.
But if, out of fear, we simply ban it and leave it at that, use may move to places educators cannot see, and we lose the chance to guide students and improve the tools.
What we must work to create is a Soft Landing: letting AI land steadily in schools and gradually become a support that teachers can rely on and students can benefit from.
Preparation should be active, trials should be solid, and scaling up should be well-founded.
And I believe Taiwan has a very good chance of becoming one of the countries in the world best suited to developing and validating education AI.
We have several conditions that are worth cherishing, and worth bringing together well.
The first is an educational foundation strong enough for the world to take seriously. In the latest PISA 2025, Taiwan ranked 3rd in reading and 4th in both mathematics and science. This can be an important starting point for Taiwan to build an international education brand.
Second, I believe Taiwan’s education is a rare meeting point: we value fundamentals, practice and discipline, and in a democratic and open society we also keep developing self-directed learning, inquiry and diverse choices. Of course, this does not deny that morale in our schools is currently rather low, and I hope the authorities in charge can find solutions at a more fundamental level.
In short, Taiwan’s environment, shaped by both Eastern and Western educational influences, gives us the chance to think about how AI can support solid learning and free exploration at the same time. The pressure of entrance exams, learning gaps and teachers’ workloads are still there, and they are precisely the real problems we need to solve in earnest.
The third is the engineering capability built up by Taiwan’s hard-tech industries.
Semiconductors, information and communications technology, and manufacturing have given Taiwan a group of people who are good at breaking complex problems down, testing, revising, and turning them into reliable products. I hope to combine this engineering culture with teachers’ understanding of learning and of people. Of course, the shortage of mathematics and science teachers is also still a reality we have to face.
To return to the combination of the spirit of education and the spirit of engineering: education provides the problems worth solving and the professional judgment; engineering provides the ability to keep verifying, improving and scaling.
If schools, research institutions, technology companies and education NPOs can work together, first producing results in Taiwanese classrooms under different conditions, and then adapting language, national curriculum guidelines and cultural context together with international educators, we will have the chance to take Pedagogical AI from home to the world.
At the same time, AI’s entry into education should also help us nurture the next generation of talent: people who can research and develop AI, people who can make good use of AI in every profession and industry, and people who can judge, govern and take responsibility for the results of its use.
Ask, Use, Manage, Build (問用管造) can become abilities that more people have, and can give more children the chance to take part in shaping the future.
This makes me think that TAIWAN can carry two hopes that echo each other, which are also the calls we have made at our annual conferences over the past few years:
Bring education to AI, and develop education AI that makes people more capable:
TAIWAN — The AI We All Need - Pedagogical AI.
The AI we all need: education AI.
Bring AI to education, and together nurture talent who can make good use of, create and govern AI:
TAIWAN — Talents of AI We All Nurture.
The AI talent we all nurture.
One direction makes Taiwan’s education and technology into a capability the world can benefit from.
The other direction gives every child the chance to grow their own abilities in this era.
We may find it hard to lead the pacing of development at the frontier model layer, but I hope Taiwan will have the chance to use a mature civil society and its technology industry base to demo the pacing of education entering AI and AI entering education.
- Original: Ray Lu’s personal Facebook post (published 2026-09-16)
Course titles follow the source. Check each course page for its teaching language.
Generative AI Leader Certification Learning Path
Five preparation courses are free and address business adoption and transformation. Course completion is not certification; the separate exam costs US$99 plus applicable tax. Its 90-minute duration is exam time, not course length.
AI capabilities and limitations
Understand next-token prediction, knowledge, working memory, and steerability to diagnose hallucinations and context limits. No technical prerequisites.