AI Models Made the List

In early June, many of us had a three-day fever dream.

Claude Fable 5, the model released by Anthropic, gave the people a sneak peek at what the next generation of AI might look like.

Until now, using AI in the workplace meant explaining things step by step. Fable was different. It behaved more like a senior researcher: Instruct the mission, and it would break down the problem, organise the processes and keep going for hours without coming back to ask for guidance. Even as it managed the work, it would ask: “Is this really what you want?” It tried to think with a longer-term perspective.

Across many forms of knowledge work, Fable was already clearly ahead of other models. In software engineering and a few other fields, some considered it beyond many human experts.

Then the fever dream broke, at least for now. The reason was national security.

The U.S. government was worried that a model with advanced vulnerability-discovery capabilities could be used to attack critical infrastructure. In cyber offence and defence, an attacker needs to find only one hole; the defender has to cover them all. If an AI agent like Fable fell under the spell of hostile actors, the consequences could impact critical infrastructure around the world.

So, the U.S. has begun to treat advanced AI agents as it treats high-level chips — an item on a controlled list requiring export and user authorisation.

Under the strictest interpretation, what matters is not only where the model is hosted, but who is allowed to use it. Even if Fable’s servers are inside the U.S., the moment a foreign national gains access to the model, that access can count as an “export”.

Once Anthropic received the control order, it could hardly prove complete compliance. This is why it chose to shut the service down entirely.

The most advanced AI capabilities are now being treated as dual-use strategic assets on par with chips. That means most people may in future get only civilian versions resembling open-weight models. The truly capable systems, such as Mythos 5 sans many guardrails, may remain in the hands of a small number of governments and tech giants.

To reduce dependence on any single model, Sakana AI has introduced Fugu Ultra, which lets many models work as a team, then synthesises their deep research; the results are comparable with Fable. At the same time, this development has pushed the issue of “sovereign AI”, already under discussion in many countries, back to the forefront.

For Taiwan, I do not believe the answer is to pour vast sums into building another Fable. Our real advantage comes from the density of knowledge built up over decades in specialist fields such as semiconductors.

One example is SiliconMind-V1, a post-trained model for chip design recently published by Team Taiwan researchers. Its goal is not to become the best AI agent all-rounder. Its goal is to be the best in a specific field. If it can be deployed at lower cost, tested against clear standards, and keep sensitive data inside the organisation, it may create more practical value than a general-purpose frontier model.

The more important question we must now address in Taiwan is not how to chase the next Fable. It is this: In which fields does our can-do country hold knowledge capabilities that the rest of the world cannot easily replicate?

(Interview and Compilation by Yu-Tang You. License: CC BY 4.0)