Memos

One of the most common things I hear when I talk to people about what they are doing to adapt to the new world, whether they’re noticing the change, is something along the lines of ‘Yeah, of course everything is changing, but I’ve always been adaptable, so I’m sure everything will be fine, I’ll be able to adapt’, which might sound like a decent argument at first glance.

But I believe they miss a crucial point. They were able to adapt due to either their intelligence or unique skills. In a world where a machine has higher intelligence and possesses all your skills, you have no leverage to adapt.

We don’t have decent humanoid robots yet, and the current AI systems mostly lack the infrastructure and scaffolding to connect to the real world. The intelligence is clearly improving at a very rapid pace, however, the infrastructure and scaffolding will be the bottleneck for the near future, which may give a false impression that this technology is not as transformational as it is.

I’m deeply worried that my skills, once a differentiating factor, are becoming commoditized. Building something functional and polished in hours used to be a competitive advantage. It no longer is. No one now has any leverage that used to provide stability in terms of both career and individuality.

The most frightening part is that people won’t notice until the infrastructure catches up. Then suddenly they’ll realize they have no leverage, nothing that differentiates them from anyone else.

Today I found NeuralOS, a really cool project where a diffusion model was trained to simulate Linux running Xfce. While playing with it, I had a thought. What if instead of training a model to predict frame sequences, you trained it directly through a differentiable engine? The model takes scene state as input, outputs an image, and you backprop through the renderer itself. The gradients encode how rendering actually works: light transport, material interactions, occlusion. The model learns compressed proxies for these functions instead of memorized outputs.

This means you could have arbitrary rendering complexity at constant inference cost. Ray-tracing, 16K textures, millions of objects, whatever. The model runs in fixed time regardless of scene complexity. And because you’re training through the renderer rather than on frames, the model learns the compositional structure of rendering. It generalizes to all viewports and configurations because it learns the function proxies instead of learning to interpolate between examples.

Of course, today this is largely impractical given current GPU performance. But in the near future it could simplify hardware requirements significantly. Imagine games specifying a ‘Tier 1’ or ‘Tier 3’ requirement, corresponding to the VRAM and compute needed to run the learned engine.

I’ve tried to make this website timeless. I purposefully decided to avoid skeuomorphism (or liquid glass, neumorphism, etc.) and modern design trends that may quickly become outdated. I am mostly inspired by classic Swiss design of the 20th century, Metro by Microsoft (which is largely inspired by airport navigation), Vitsœ, and the New York City Subway.

I figured what works best is creating a sense of order and clarity with grids and typography, then adding some details that are not perfectly aligned to make everything feel more human and less sterile. This completes the overall aesthetic I’m aiming for.

Even after 15 years since the iPad’s release, most note-taking apps are still skeuomorphic. I mean it in the fundamental sense: they use the paper metaphor — pages, ink, eraser, and highlighter. This worked to make people believe that it could replace their notebook. But with more time, it becomes more and more pointless. In fact, this hurts the experience significantly and limits the possibilities.

It’s like if cars were designed to look exactly like horses and carriages instead of being built for speed and efficiency.

For instance, why do we have issues like ‘I started writing a word but realized there’s not enough space’? This problem is inherited from paper and has no reason to exist. Formatting writing between pages also causes problems. Even though highlighting definitions doesn’t correlate with memory retention and is proven to be purely for structuring, we still do it as if writing on paper.

What if you escaped from the paper legacy in note-taking and tried to build the experience from scratch? Just like a typewriter supercharged the handwriting experience, or a computer supercharged the typewriter.

Speaking of new perspectives, most users of note-taking apps are students. They use the app as a utility to preserve, decompose, and structure information for the purpose of learning. For some reason, the apps that are marketed as educational tools are still mostly designed to be the best recording tool, not a general-purpose learning tool, which is how they are actually used.

Some complementary factors are:

• Since 2022, students use AI as one of the core tools for learning
• They provide context to the AI — essentially, the same context they write down in their notes
• AI models become more and more capable and efficient every year and can be run on a tablet
• Modern tablets have insane compute power, which is mostly idle during note-taking

What if you built a note-taking app that is not just a recording tool, but a learning tool? Then the primary focus becomes the retention of information, the engagement, and the evaluation of progress. I don’t mean simply adding a chat sidebar, but rather building the app around the new technology and the new use case.

Went to a Raycast meetup at the amo office in Paris. Well, ‘office’ is an overstatement.