Print gallery transform: https://github.com/abetusk/CursedLogic/tree/main/sandbox/slo...
This is a web application that you can upload a picture to and do an Escher like print gallery transform to. Two examples from using the application:
https://github.com/abetusk/CursedLogic/blob/main/sandbox/slo...
https://github.com/abetusk/CursedLogic/blob/main/sandbox/slo...
Dragon rider (aka "Dragon ring"): https://github.com/abetusk/scratch/tree/release/experiments/...
Hooked up an AVR atmega32u to my exercise bike to poll the hall effect sensor to sense when the crank has turned and send that to a web application as a MIDI message to indicate when the dragon should 'beat' it's wings.
The web application is a relatively simple game where a dragon flies through rings for 15, 30, 45 or 60 mins. The idea being, a simple, relatively boring game is much more interesting when you tie it to another relatively boring task of cranking on the exercise bike.
Print gallery was a few rounds of Claude (Opus 5.5). I would guess that would've taken me 20-40 hours. So maybe an hour of wall clock time to save me 20-40 hours? 20x-40x? That's conservative, because a lot of that was iterating on the application, using it, etc.
Dragon rider I have a better sense for because I've made a game of similar complexity, and that took me 40-60 hours. Dragon rider took a few iterations, maybe 15 rounds or so with Claude, but I would guess that's a couple hours. Considering all the effects I experimented with (and mostly abandoned) the refactors, extensions and feature, I would easily guess that an equivalent of my time is in the 100+ hour range. So again 20x-100x with the same caveats as before.
When 10x is so effortless, it seems like 100x is easily within reach. Wait a few model generations and the factor easily compounds.
Yes, those are 20x-100x tasks indeed. They are into the intersection of "small tasks" and "niche personal projects used by the owner". Which is totally fair use of AI.
The article doesn't claim to be impossible to achieve those results on particular instances/cases. It proposes that this multiplier does not become the new baseline for all new tasks. And that those are multipliers are crazy high in the Corporate context. Unless your corporate job is solely creating image transformers, I guess :)
Each of those examples is a 10x increase in productivity, before the advent of LLMs.
For anyone who's used these modern agents to do programming, it's obvious there's a 100x increase for certain programming tasks. Anecdotally I've seen the 100x increase for projects myself.
Tao talked about a 10x slowdown when trying to convert a proof to Lean (from how long it took to do a 'paper proof'). Modern agents can do a conversion to Lean, for a variety of topics, 10x faster than a human, at least.
Humans do not have a good intuitive understanding of exponential growth. The reason the OP is so skeptical is because they're comparing it to everyday experience (lifting thing, driving time, etc.). It's easy to discuss these points in context (lifting things can be done with cranes which do provide 10x human lift power, commercial airlines allow us to traverse the globe 10x faster than walking, sailing or driving) but I think this is missing the point. Computation has limits but we're nowhere near them.
> In 3 days, at 100x and you achieve the work of 300 working days. There are 260 working days in a year.
This has literally happened to me. A project that would have taken me a year was done in 3 days, and this was for a domain that I was well versed in. A project that would have taken me 10 years (because I wasn't well versed in it), took me 7 days.
> That sounds unrealistic.
It may sound that way but it absolutely isn't.