Does Reality Have a Surprising _Lack_ of Detail?

The week's AI news is a litmus test of your sensemaking chops

This essay is part of the Mediocre Computing series.

Last week’s newsletter (Massed Muddler Intelligence) was hard to write and delayed because for a change I was trying to think like an engineer first, and a writer second. Which means sweating first principles over dreaming up banger memes and metaphors. And since I’m a better writer than engineer, it was hard. But gratifyingly, as I’d hoped, I got a lot of good responses from actual practitioners trying to build with AI technology. I think I’m going to continue with this bias for a while (though this essay should still mostly be comprehensible to non-technical readers), but might attempt a more general-audience synthesis in a few months. Right now developments are happening too fast to translate to general audience mental models.

This week’s two major bits of AI news — the shock-and-awe launch of Sora, OpenAIs video-generation engine, and the rather sorry launch of Gemini 1.5, Google’s new 10 million-token context LLM (look them up before proceeding if you haven’t caught up) — constitute a real test of your sense-making of developments.

Which do you think is the more significant development? Vote even if you have really low confidence in your take and would rather not judge. Let’s use the convenient poll tool Substack just introduced. Vote before you read on.

Second question, why do you think what you do?

Before I share my own answers and analysis, let me offer a provocation that I think sheds light on the correct answer. Here is a video (via this tweet) of a stormy sea not generated by either AI or the real world.

It is generated by this fragment of clever fractal-based procedural generation code which my friends tell me is WebGL:

for(float e,i,a,w,x,g,h;i++<90.;o+=.01-.02/exp(max(s,e)*3e3)/h){vec3 p=vec3((FC.xy-.5*r)/r.y*g+2.,g);p.zy*=rotate2D(.5);e=p.y;h=e+p.x*.3;p.z+=t;for(a=.6;a>.001;a*=.7)p.xz*=rotate2D(5.),x=(p.x+p.z)/a+t+t,e-=w=exp(sin(x)-3.)*a,h+=abs(dot(sin(p.xz/a*.3)*a,r/r));g+=e=min(e,h*.5-1.);}

I have no idea how it works, but fortunately someone asked ChatGPT to explain, so you can read that tweet if you’re interested.

Now ask yourself, how much information is actually in that video? There’s a decent argument to be made that there is no more than 280 characters worth (though well, it calls a bunch of library code…). That fragment of code fits in a tweet.

I often cite the idea that “reality has a surprising amount of detail” and the great John Salvatier essay, which makes that argument, and this bit of code raises an uncomfortable antithesis question. Contra Salvatier: Does reality have a surprising lack of detail? A richly textured view of a turbulent ocean by a rocky coastline…reduced to 280 characters?

Now consider the Sora vs. Gemini question again in light of the implications of this non-AI video. Our three sensemaking questions are:

1. Which is more significant, Sora or Gemini 1.5?

2. Why?

3. Contra Salvatier, Does reality have a surprising lack of detail?

My answers, in order, are:

1. About the same

2. Because boring inference advances are more valuable than flashy training advances

3. No, but it’s kinda complicated.

Let’s dive in.