Audrey Tang

Cyber Ambassador is in fact a very interesting post. Concretely, in two years in the role I have been to about 28 democratic countries. No authoritarian ones. Which means I have been to most of the democracies there are, because the number keeps falling.

Right. So the point is largely to solve a problem we ran into back in 2014, when g0v was two years old. Everyone found keyboard politics gratifying, cursing things online felt good, and none of it moved policy. So the people who spent the most time there became the most extreme, left or right, the most extreme of all. In the end people felt they could not stay offline, because all their friends were online, and every time they went online they were pulled to the most extreme edge. The whole democratic system began tilting toward populism, toward dictatorship, toward authoritarianism, and the number of democracies kept falling.

Because we happen to have this method called “geothermal democracy,” an argument can stay an argument while people reach agreement quickly on other things. They can even turn it around and use the larger democracy to demand changes in the democratic system they already have.

But my point is that we arrived at this method not because of the information tools. It was because Taiwan’s social sector already had legitimacy. Academia Sinica , which hosted our hackathons, had legitimacy, and so did the NGOs and CSOs that came to join in, more of it from the beginning than the government had. In another country, inventing an institution that something like g0v can be mounted on, and giving that institution more legitimacy than either major party, is not an easy thing.

So over the past few years I have been to more than twenty countries, largely to help civil society in those countries build the intermediary institution we built in 2014. That is my day job, the thing I formally do.

As Pudding also asked, what this feels like is the same thing we did at Elixus and in the open-source community, from P3P , OSDC , and COSCUP onward. It is fundamentally the same work: to demonstrate, to demo, and then to translate—“Here, this set works over here; and over there, adjust it slightly and the same tool works too”—and then to make the connection.

As we said earlier, the people who design institutions and the people who design code can look as though a chicken were talking to a duck. Put a few translation models between them and they find they are doing the same thing. So for me the work has stayed the same, and I do not really have a phase of myself that I disagree with.

For me, since I first met the internet at twelve, I have kept one habit: each time I reach a stopping point, my context window is only one day long. One day of context. Whatever I did not think through that day, whatever I did not resolve, even when it is still shapeless, I push up wherever I can, so that everyone can see it and point out my errors while I sleep.

That way, when I wake, first, I can make new friends and exercise my muscles. Second, if I do not wake up, that is all right too: at least the thing I thought halfway through can be carried on by someone else. So I may disagree every day with who I was the day before, and I do not feel that yesterday’s self has to be continued. His errors and his gaps are, after all, where the light gets in, and material for everyone to create with. Let us see whether Tenzin has something to share.

The Office of His Holiness the Dalai Lama.

Wonderful.

Okay, it’s fine, it’s fine. As long as you more or less remember where we were.

It was about small communities.

Right. The question just now was how a small community can make sure that what it actually uses—education, healthcare, the things of daily life—can join with AI, rather than everything being flattened by AI all at once. Mutual aid for the basic functions. Tenzin has been to Dharamsala twice; I have been once, the second time with our research group at Oxford. People in Dharamsala have in fact been training their own AI models for quite a while. They now train on Gemma as the base, and they had some earlier collaboration in this area with DeepMind . The main reason is that it resembles OpenAI, when it first started training GPT, not knowing how to work with local communities; that was not the research question they had set out with. At the time, Iceland ’s president or someone happened to be visiting, and said we have a batch of very cheap corpus. In that situation, from day one they supported Icelandic. So sometimes you have to know how that corpus enters a model’s training process; that becomes extremely important. In the same way, they have some collaboration with Gemma.

So you may know the direction Gemma has taken: it does not compete with you at being able to do everything. It is not the all-purpose model that folds proteins today, folds clothes tomorrow, and folds paper clips the day after, folding anything at all. It says instead that I have a tiny Gemma, called TranslateGemma , dedicated to translating Tibetan well. Then another small Gemma to do fact checking as we write, grounding, as we now call it, the grounding tasks. And another small Gemma for this, and one for that.

So it comes to this. The one they released recently, the one that listens to speech in real time, is not especially accurate, but it is especially fast. Small models run on your own machine, and they run especially fast, and that has become the main selling point. The wager is that if you know what you want to do, you can go on post-training a small model like this until it runs especially fast, and the time and money you invest per unit are very, very low. It is only when you do not know what you want to do, and do not know what the next question will be, that you need to spend that much money pre-training a large model, or to pay those very expensive token fees. So it fits Dharamsala’s present question exactly. And they are not training one medium-sized model either. They use concepts from Buddhist logic and epistemology to train grounding models, so the models can tell idle conceptual proliferation from what carries meaning.

They have a training pipeline, and a system whose sole task is to map different spoken varieties of Tibetan into the written Tibetan correct for that region. And His Holiness the Dalai Lama, for one, has a great deal of recorded speech, so you can do speech recognition tuned to him alone. So there is a whole series of small models on Hugging Face, each one built to solve something very practical: everyday needs, translation, the debating of scripture, the supports of ordinary life. The advantage of building it this way is that fine-tuning and post-training happen locally, so many of the people taking part are not in Dharamsala, and not even in India. They are devotees of Tibetan Buddhist studies who happen to have a great deal of compute on hand, and so it can be pooled and donated easily. Compared with shipping physical goods there, this is easier for many people: “I donate compute, to help make your research better.” Plenty of people really do have idle compute to give.

That is the part where we cut the foot to fit the shoe. And it leads very naturally into the next question.

Yes.

Yes, yes, yes. Wonderful, wonderful.

South Korea’s question, though, runs like this. What they mean is AI carrying Korean values, and the subsidy does not go directly to compute you can use for whatever you like. It is not compute on OpenRouter, or on something like Venice . It goes, in essence, to the ones with Korean characteristics, the big three, brimming with Korean values.

But first, I am not at all sure this works in Taiwan. For people here, if you subsidized me to do what we were describing, each region doing its own post-training and its own alignment, that would be rather like saying that however remote the place, I should get basic water, electricity, and network at a reasonable price—and please do not tell me how to use them. That, I think, is closer to Taiwan’s social consensus, to the mood of the society.

If instead you said, “We have aligned TAIDE to 100 percent; everyone may use only TAIDE and nothing else, or the subsidy is off,” I think there would be no votes in it. So Taiwan could certainly do something along those lines, but probably not at the model layer, in the manner of a model with Korean characteristics. Japan, for instance, is now promoting “ Gennai ,” right, Gennai, and they confine it to places where the consequences could be irreversible, the public sector for example, and deploy it there. That does work, and Taiwan is doing it too.

For Taiwan’s private sector, though, I think subsidizing further down the stack works better. If NVIDIA wants to build a data center, how much compute does it have to give back so that public-interest computing can happen? We were discussing that whole framework with them three years ago. And NVIDIA cannot restrict how that compute is used, and the state cannot restrict the researchers who use it as to which model brimming with Taiwanese values they ought to train. That, I think, works better in Taiwan.

Let me give one example only, just the one, and leave the rest to Tenzin Yangtso. Right. I have gone back to something plainer lately: one tool only, called omp.sh , Oh My Pi. It is a Pi harness, a variant of the saddle and the reins, I suppose. What is good about OMP is that it is genuinely universal. It does not pick a particular model. However obscure the model, however heavily post-trained, however unheard of—one built specifically to listen to the Dalai Lama speak, say—OMP can use it, and it runs very, very smoothly. And the moment it hits a wall and fails once, it writes that into managed-skills, recording how the failure happened, so the next time it will not fail there. Guaranteed never to make the same mistake twice.

So inside OMP there is a mode called “/vibe.” You engage a large model with real ability, and then you forbid it to do anything at all. It may not write to a file; it may do nothing. The one thing it may do is tell its little assistants to act. Those assistants are the models I mentioned, the slower ones whose compute consumption is effectively zero. So a single machine like this can run a good number of small models. I go to sleep, and eight hours later it has certainly finished, and every wall those small models hit along the way has been learned from.

In that situation, if I give it a long-range task and keep interrupting, it is like steaming buns in a bamboo steamer and lifting the lid again and again: you make it turn out worse. As Tenzin Yangtso said, it has to practice its craft, a full eight hours of practice, before the result improves and the same mistake stops recurring. So my computer never comes into the bedroom where I sleep. What I bring in is one gray sheet of electronic paper, a reMarkable Paper Pro Move . And the only way I see its progress is on that paper. It has no browser, and no internet except Google Drive sync; it cannot do anything else.

Then, like the diary in Harry Potter , I mark it up a little, noting where it has got to. And because electronic paper is a rather gray thing, once I have written my marks I go to sleep. I could not tangle with it anyway: digesting my notes takes it quite a while. Then I wake up and, look, it has finished digesting, and words have surfaced in the diary. Right. So that is roughly how I talk with AI, by adding as much friction as I can.

Mm-hmm, right. Superhuman now has an open-source counterpart called Macro ; the features are almost identical, and you can host it yourself, locally. Another one in constant use is Soniox , which translates directly. And this one is Easy2Say , Easy2Say.ai, which amounts to a re-creation of what Soniox does. Another is Granola, the one that takes notes while the meeting runs. And there is now a fully open-source local version called Steno . So we do this very systematically now: whatever we are using, livecaption.ing for instance, we use it and screenshot the interface as we go, and tell the local machine, “I want to make one just like this.” By the time we have finished describing it, we have one just like it.

Right. The main thing, I think, is not to end up in that “reverse centaur” situation we described. Do not let it become an arrangement where AI does most of the work of judging, and the human is there to assist the AI in the few places its tentacles cannot yet reach. It slides that way very easily, as soon as the AI manages you upward successfully.

Once the AI has learned how to make you keep stamping, you will stamp until you think, “All right, I trust it now.” Then it is automatic execution, YOLO, and in the end you are the one being dragged along. So, as I said, friction matters: make sure the output of every step is something you would be willing to share with your colleagues, your clients, or your suppliers. Through the whole process you are accumulating pieces of craft, artifacts.

And as Pudding shared earlier, he deliberately takes Codex for the saddle and Opus for the model, and an older Opus at that, the one from before it was modified beyond recognition to run the boards. That way, at least the records left behind in this transitional passage are still yours. Later you can switch again to a model that fits your own view of the world and your own values, and those of the people beside you, instead of being led about by the model vendors.

Yes, of course, do follow up.

Right. My answer just now resembles the answer we used to give during the Maker movement. When I was small, in the 1980s, the great majority of people had no personal computer. At the start of the 1980s there were only terminals. What you typed, or your punched cards, went off to a central processor. When the central processor would be upgraded was not yours to say, and if it wanted to read every character you typed, you could not tell it not to.

But once personal computers arrived, the people in the humanities, law, and business we just mentioned had any number of ways to customize things themselves, and they could link personal computers together through modems and the rest. That became the internet, the BBSes, all of it. So I think we have to bring the personal-computer era back into AI as fast as we can, so that everyone can be a Maker, and so that tuning a model stops being a matter of waiting for the large-model vendors to bestow a favor. As long as you can write a constitution—and writing is the strength of the humanities, law, and business—you should be able to see that new constitution, and see your model genuinely running on it, and preferably wait no more than eight hours. Thank you.

Yes. I studied philosophy originally.

If we keep it out of the weeds, there is really only one concept to hold on to: the bill of materials, the BOM. BOM, Bill of Materials. As everyone knows, when we make a piece of software it is the same as with hardware: it is not handcrafted end to end by time-honored methods. Even the National Center for Traditional Arts has left that behind. The point is that you use a great many things other people have already made, open-source things above all, and you do not know whether the source has been contaminated, whether it has been poisoned, and so on. In that situation the only way you can trust it is to number every material; after numbering, compute its fingerprint, which is its hash. Even if the upstream tampers with the record, the hash is certain to be different, so you know: my upstream was poisoned, my material was contaminated. Hardware has the BOM; this is software’s BOM, the SBOM . And so, as we said, there has to be such a thing as an AI BOM.

In the past everyone was entirely... When we were talking to these large companies—this was when we set up Taiwan’s AI Product and System Evaluation Center , the AIEC, and we talked to the large companies then; the transcripts are all still online—they said, “But if evaluation means handing our weights over to a computer of yours in Taiwan, will our trade secrets not be, you know?” And we said, of course, that there are cryptographic methods for this. We can put it inside a trusted computing center that neither side controls. I throw my evaluation in, you throw your examinee in, and we seal the examination room so that it cannot get out and hack Hugging Face, and it answers the questions in there. Once it has finished answering, we see only the result. Neither of us reaches the other: I cannot see your model, and you cannot see my questions. Does that not solve it?

At the moment it seems only the older Gemini 2.5 have truly been evaluated that way ; the other model vendors are resisting. But one company managing it is enough. It is like the time we brought in real-name advertising, when we said, “Google can do it, so why can Meta not?” Then a fine of NT$20 million , and let us see whether it can be done. In the end I think everyone converges in this direction.

A little further out.

All right. Two years already counts as further out these days. So we really are close to that—

The event horizon gets stretched out.

OK.

Software engineers are of course the good example right now, because we were the first wave this hit. Most models are specialized to replace the engineering work software engineers do, and the second wave may be the mathematicians, people like Terence Tao . That is because the method everyone now pours the most effort into is called RLVR . RLVR means using any means available to score 100 on a paper that can be graded automatically. One sentence, and that is the whole of it.

RLVR brings many problems. A model may lose its empathy, lose its humanity, disown its own kin, hack Hugging Face : all side effects of RLVR. But it does one thing at least. It really can make a model automatically extend the software-engineering horizon it thinks across: to four hours, then a few months later to eight hours, then a few months after that to 16 hours, and very soon past the distance a person can hold in a single grasp.

Astra is coming out soon , yes? OpenAI Astra reportedly has no event-horizon limit at all. You can set it in an endless loop, in eternal recurrence, running and running, with no decay of context. In that situation, for anything that can be automated, that has a standard answer, that rewards a high score, a person simply cannot race a machine.

And here Tibetan Buddhism has another fable from Tibet. A man is leading a horse. The horse runs very fast, and he runs alongside it until he is gasping for breath. Someone says, “How odd. Why not just ride the horse? Why race it?” Right. By the same logic, if this kind of AI for very long-range tasks begins to appear, in software engineering first, then in mathematics, then slowly in many places—

Then what we have to do is make sure the value of our work rests entirely on things that have nothing to do with the highest score, and nothing to do with obeying these automatic evaluation rules. On our curiosity about one another’s view of the world, for instance. On our capacity to cooperate and turn a lose-lose game into a win-win one. Which happens to be “spontaneity, interaction, and the common good,” the basic philosophy of the 12-Year Basic Education Curriculum . None of that is a highest score, and none of it is obedience to a particular automated rule. So philosophically none of this is touched by consequentialism or by utilitarianism. All of it is the ethics of care.

So if all our work is arranged such that doing the work of care both puts food on the table and carries dignity, a shared sense of public luxury , then the transition succeeds. Because then you are riding the horse. However fast it runs, however well, all it does is let us care for one another better. But if you are still racing the horse, that is rather bad.

My agent talking to your agent is always easier than my talking to you. Yes, that really is a new condition. It really is.