AI Ingredient Labels
I was looking over article/document covering “what we think the big AI companies are doing” but shouldn’t we not be guessing about these things? Shouldn’t we compel all these companies to release details of what they are trying to use to control the world? “Secrets are leverage” after all.
It’s a scenario reminiscent of when Apple releases a new i14 or iWatch then a dozen companies rush out to take it apart and xray the components to figure out what’s inside — we know what’s inside because it was already developed by people, so this is just a huge waste of time? (similarly, why don’t we compel companies to release chip designs for all public hardware too1?)
anyway, what if we had AI Nutrition Labels?
What goes into an AI Nutrition Label2?
MAGI via EE
A label for Matt’s AGI running on a patented Eschatology Engine:
AI Model Facts: MAGI via EE
Total Training: 66 days @ 1 EFLOP (512 x H100)
Training Energy: 500 Megawatt-hours
Inference: 33 ms @ 300 TFLOP
Inference Energy per Sample: 20 Watt-hours
| Amount Per Serving | ||
|---|---|---|
| Model Parameters 22 T | Model Memory 40 TiB | |
| % Training Data* | ||
| Text Language Inputs 50 TiB | 60% | |
| American English 40 TiB | 80% | |
| Japanese 10 TiB | 20% | |
| Text Code Inputs 170 GiB | 20% | |
| C 50 GiB | 30% | |
| Python 80 GiB | 47% | |
| ECMAScript 20 GiB | 12% | |
| Turbo Pascal 20 GiB | 11% | |
| Static Image Inputs 170 GiB | 20% | |
| Stock Images 70 GiB | 41% | |
| OnlyFans 100 GiB | 59% | |
| [AI] Protein 3g | ||
| epistemological risk 44% | alignment risk 33% | |
| capability risk 99% | goo risk 33% | |
* Percent Training Data is based on statistical estimates. Final byte-by-byte results may vary ± 7% from stated values.
Input data obtained by sampling subsets of the following datasets:
| Dataset Size: | 550 TiB | ||
|---|---|---|---|
| Common Crawl 2022-12 | Up to | 450 TiB | |
| Stack Overflow 2022-12 | Up to | 33 GiB | |
| LAION-5B | Up to | 15 GiB | |
| Wikipedia | Up to | 100 TiB | |
| Custom Image Crawler (data available by request) | Up to | 200 GiB | |
We legally certify our deployed model does not deviate from this label definition. This label definition includes all datasources our model trained against and any future finetuning will be amended with a supplementary label update.
Further work
You could imagine this at a glance label format also being used to record implementation-level details like architecture, width, depth, epochs, embedding criteria, position criteria, hyperparameters, any non-standard usage of standard components, etc. Labels could also be arbitrarily recursive if data sources also want to break out their own contents or if we continue breaking down models referencing sub-models ending up with 87 turtles all the way down.
At-a-glance labels could be adapted for subjective conditions too. Let’s provide more insight into per-model alignment work since so much alignment is now 87 levels deep into insiders just creating more and more abstract fantasy ideas from other insider scifi writers forever and nothing makes sense anymore. or worse, all the “corporate alignment” nerds who only say “Sure, let’s do alignment — as long as it protects our egos, reputations, personal brands, and grows our capital as fast as possible, but not alignment for what human live people actually want.”
Maybe instead of caring about fantasy scifi ideas by a discount zac oyama we can create systems for widespread use guarded by personal responsibility?
CONCLUSION GOES HERE
so yeah, stop letting people with aspirations of becoming thousand million trillion dollar companies release unsecured world-ending interactive models into the world without global oversight into both their motives and their methods.