I no longer switch on the day a new model comes out. Leaning toward the cheaper one, moving up to the stronger one, adding more workers: I have stopped all of them. What I look at instead is only what each mechanism costs. There is 1 unit for it, and what comes out costs more than what goes in — how much you have it write weighs more than how much you have it read. This series takes the mechanisms built across the previous 5 series and measures them again from a single point, the allocation of compute.
First, a word on how to read it. This series can be followed by reading alone. Each article starts from what happened in front of me and goes on through what I put down and what went away. Near the end of each article there is a section called "How to verify." You do not have to read it first. Come back to it when you want to check whether my numbers come out in your own environment.
Learning the remedy from the symptom
Each article starts with a symptom, something that is not working. The remedy comes with the relevant behavior and specifications of the AI, and an explanation of why that remedy fits.
You do not need to learn why it fits before you use it. Put one mechanism down and the symptom is handled.
Even when you know the behavior and the specifications, the AI can still get it wrong depending on how they combine. Ordinary design and development is unlikely to catch these pitfalls, where several behaviors and specifications are tangled together. You learn them from the symptom, one at a time.
Start by checking what you recognize
Start by recalling just one thing.
The last time you switched a model or a setup.
What happened, that made you decide on that switch?
After the switch, did you count what had got better than before?
If you have never switched a model or a setup yet, it is natural that nothing comes back. This series is a record of counting again what the mechanisms built across 5 series are costing. You can read it as it is.
What 5 series and 55 articles built
This is the 6th series.
- Series 1, "The Art of Not Reading": it put the AI's output into a shape you do not have to read
- Series 2, "The Art of Not Listening to the AI's Opinions": it put the AI's proposals and judgments into a shape you do not have to weigh up twice
- Series 3, "The Art of Not Telling": it put the human's input into a shape you do not have to say
- Series 4, "The Art of Not Checking": it put checking itself into a shape you do not have to decide on the spot
- Series 5, "The Art of Not Running": it replaced running things in order to find out with problems whose answer is known in advance
What the 6th series deals with is the step before all of that — which model, how many workers, and in what setup you run it.
The mechanisms of the previous 5 series had not reached this part yet. That is because a mechanism decides how to hand work over to the same counterpart, while on the judgment of choosing that counterpart I had put down none at all.
The 2 things this series does not chase
The title is broad, so let me narrow it first. The 2 below are what I do not chase, and everything else I do chase.
| # | What I do not chase | What that means |
|---|---|---|
| ⭐⭐ 1 | The latest information | Not weighing up whether to switch on the day a new model or a new feature comes out |
| ⭐⭐ 2 | Digging deeper once the answer is in | Going on investigating even though enough material to decide is already there |
Start reading with only 1 in mind and half of it will not reach you. 2 is handled in its own article.
There is 1 principle that works across every article
The bill is attached to both what goes in and what comes out. And what comes out costs more.
I do not write the multiplier (that is the latest information, so it goes stale in half a year). Remember the direction alone: cut the same amount, and cutting it on the output side works better.
This 1 line works across most of the articles in this series. Where and how it works, I go through 1 at a time in each article.
What this series hands over
Let me separate up front what can be handed over from what cannot.
| ⭐⭐ Can be handed over | A procedure for counting, on your own machine, how much is going where (it sits at the end of each article) |
| ⭐⭐ Can be handed over | The numbers I actually measured, and their denominators (⚠️ the predictions that came out wrong are written down as they are) |
| ⚠️ Cannot be handed over | Which model you should use (⚠️⚠️ it goes stale in half a year, so not 1 of them is written) |
| ⚠️ Cannot be handed over | The multiplier in your environment (⭐ run the procedure and your own number comes out) |
Model names, prices, performance figures: not 1 of them appears in this series. What gets compared I call only "the expensive side" and "the cheap side." In exchange, every multiplier and every denominator is there.
Next time, the art of not digging deeper. I stopped going on investigating once the answer was in. And even so, the material I need in order to decide has not shrunk.