A Task Worth 90 Minutes of Its Work

TSMC today reported October revenue of NT$586.89 billion, another single-month record. All of Taiwan is talking about TSMC, but can a junior high school student explain how a grain of sand becomes a chip? I gave the latest Fable 5.2 a task. It worked on its own for 90 minutes straight and produced a website that a junior high school student can understand, with a level of detail that surprised me.
I gave the AI only 3 lines of instructions, and it worked on its own for 90 minutes straight without stopping once, producing an interactive website about semiconductors that a junior high school student can understand, that can switch between 6 languages, and that runs all the way from TSMC to NVIDIA.
More importantly: the 5-hour quota on my Max plan, which I had not burned through in a long time, was actually burned through.
This is Claude Fable 5.2, the new model Anthropic released just in the past couple of days. The instructions I gave it were, word for word, just these few lines:
“First go through all the TSMC- and Nvidia-related courses on Junyi Academy, then go ahead and build a website similar to the Wondering App, with interactive videos/conversations on it. TA: something even junior high school students can understand. Start building the website. It should be switchable between English, Traditional Chinese, Simplified Chinese, Japanese, German, and Korean.”
When I opened the result 90 minutes later, I was truly startled.
The website is called “Sand Becomes a Chip” (沙子變晶片). The main character is a grain of sand named “A-Sha” (阿沙), who takes you from the wafer fab to the data center, 8 stations in all. The first 5 stations are the TSMC part, about how chips are made. The last 3 stations are the NVIDIA part, about what chips are used to compute, and the final station breaks down how a chatbot plays “word chain.”
Each station has four steps: first watch a video that pauses to ask you questions, then adjust the parameters with your own hands (tuning the light, etching, training a neuron), then chat a little with an AI study buddy, and finally take a three-question check to clear the station, earning a stamp for correct answers. Your learning progress is a wafer on which the chips light up one square at a time.
What surprised me most was the details:
Switch to English and A-Sha becomes Sandy; switch to Japanese and it becomes “Suna-kun” (スナくん).
The Simplified Chinese version does more than convert the characters: “晶片” (chip) is changed to “芯片,” and “國中” (junior high school) is changed to “初中.”
On its own, it stated on the website that this is a demonstration site, not an official website of TSMC, NVIDIA or Junyi, and that the foreign-language versions are AI translations that have not yet been proofread by native speakers.
It also set rules for the AI study buddy: at most 3 sentences each time; start with an analogy from everyday life, then ask the student a small question in return; never ask for the student’s name or school; if unsure, say so, and never make up numbers.
I had not specified a single one of these rules.
What people abroad have been discussing most heatedly over the past couple of days happens to be exactly the two things I ran into: it is very strong, and it also burns through quota very fast.
The results Anthropic itself has shown are startling: 1,000 instances of Fable 5.2, dividing the work among themselves, took 30 hours to translate the proof of Fermat’s Last Theorem into a language a computer can check line by line. With the same setup, the previous-generation flagship Opus 5.5 ran for 18 hours and completed only 26%. In another experiment, it directed more than a thousand subagents and in 8 days wrote, starting from an empty folder, an operating system that can boot on a laptop.
But on Reddit there is a popular thread whose title says outright, “WTF, the cost per task of Fable 5.2 is ridiculous.” On Hacker News, someone also said that in a single evening they used up a quarter of the Fable quota in the US$200 plan.
One Hacker News user who has already used it put it most aptly: in everyday coding, he finds it hard to tell the difference between Fable and Opus 5.5; you have to raise the ambition of the task before the gap shows.
That sentence is the user manual for Fable.
Fable’s strength lies not in answering faster but in “taking on a project.” Hand it a big task, and it will break down the work itself, check it itself, and carry it through to the end. Ask it a small question, and it is about the same as other models, only more expensive. Using Fable to ask “What should I have for lunch today?” is roughly like asking TSMC’s EUV machines to print your lunchbox menu.
So burning through your quota is not necessarily a waste. It may well mean that you have finally given it a task worthy of it.
So what are people supposed to do?
Wharton School professor Ethan Mollick, after trying it early, wrote a long essay saying that Fable 5.2 “almost” counts as AGI, except that its abilities are jagged: in some areas it is already like an expert, and in others it still gets stuck. He asked it to rebuild the literature course that the poet W. H. Auden taught in 1941, which the internet has called one of the hardest in history. The first version was accurate in content but utterly boring: a 15-week course website with exams. Mollick had to tell it himself that “this is a slog through boring web pages” before it made changes. The changes took only a few minutes, but it had not noticed the problem itself.
Mollick also admitted that every one of these works came about because he first gave the AI an idea.
My website is the same. It says “suitable for junior high school,” but whether Taiwanese junior high school students will get stuck at the “Drawing Circuits with Light” station or the “Carving and Stacking” station, it does not know. Only teachers know. The final acceptance check has to be left to the junior high school students themselves.
Now that AI can work for 90 minutes straight, what children will need most in the future may be two abilities: being able to set a task worth 90 minutes of AI’s work, and being able to tell whether the finished product is easy to understand, whether it is fun, and whether it has mistakes.
If someone in every field in Taiwan seriously burned through their quota once this week, what could be made?
Teachers: turn the unit the whole class has been stuck on the longest into an interactive website that pauses to ask questions, and make Vietnamese and Indonesian versions as well, so that new-immigrant parents can go through it with their children.
Factories and the tech industry: turn new-hire training and SOPs into multilingual interactive lessons that migrant workers can also understand.
Hospitals and clinics: turn a single patient education sheet into a website that patients and their families can click through and ask questions of.
Hardware startups: in testing, Fable designed 3 electronic devices that actually work from nothing more than a paragraph of description, handling everything from the circuit board and the enclosure to the code. The best place in the world to turn this kind of capability into products may well be Taiwan’s supply chain.
The public sector and civic tech: Mollick had it read through the meeting agendas of 201 local governments in the United States and organize them into intelligence. The council records of Taiwan’s 22 counties and cities deserve to be organized properly in the same way.
If you want to try it, this week happens to be a good time:
Pro plan: until 3:59 p.m. Taiwan time on Tuesday, November 17, Fable also counts within the monthly subscription’s quota, up to half of the weekly quota, and existing subscribers also receive one “Fable reset.”
Max plan: depending on how long you have subscribed, you receive 1 to 3 Fable resets, valid until March 8 next year. Press the button on the “Usage” page in Settings, and the 5-hour quota is refilled completely.
The default model does not switch automatically; you have to select Fable 5.2 yourself in the model menu. For small everyday matters Opus 5.5 is enough, and even Anthropic itself recommends this.
Three small tips: make the task big, and say clearly who it is for and what counts as good enough; once you have handed it over, let go and let it run, without cutting in every 5 minutes; when it finishes, you do the acceptance check: is it easy to understand, is it fun, are there mistakes?
At the end of every station, the website takes you back to Junyi to watch the complete original courses: My Semiconductor Adventure (《我的半導體大冒險》), licensed by TSMC, and the “Generative AI Learning Zone” (生成式 AI 學習專區), which draws on NVIDIA’s teaching framework. For the past few years Junyi has been teaching children “Ask, Use, Manage, Build (問用管造),” and this time I learned another lesson myself: only when you Ask big enough, Use long enough, and Manage carefully enough can you Build.
AI can now work for 90 minutes without a break. What is truly scarce now is a task worth 90 minutes of its work.
Is there something you have wanted to do for a long time and never had the time for, something you could hand to Fable this week?
(The website link and sources are in the first comment.)
- Original: Ray Lu’s personal Facebook post (published 2026-11-10)
Course titles follow the source. Check each course page for its teaching language.
AI Fluency: Framework and foundations
Practice delegation, description, discernment, and diligence with the 4D framework. Free course; completion badge after assessment.
Dissecting the Crawfish: How AI Agents Work, via OpenClaw
An OpenClaw-based introduction to how AI agents work. Free public lecture; no self-study certificate. Costs of independently using external models or tools are separate.