Albert2088

Building the most performant video editor

[Albert2088 @water2088]

Prediction/Cold take: designing and developing good UI will be more important than ever in the agent age.

The popular rumor is that
1. AI can just look at APIs or CLIs to operate the software
2. The future of software is for agents
3. Therefore, building textual APIs for agents is the way.

My take:
1. Many behavior intentions can not be precisely and concisely communicated in texts. An image is worth a thousand words.
2. Image understanding will become so fast and accurate and cheap.
3. Therefore, having agents directly working on GUIs will be more useful.
AI works on latent space of audiovisual data, not on text is the future.

[Albert2088 @water2088]

I had a hunch years a while back that ROC can be a great sandbox language for the agent age. We are entering the ad-how code age where compile and cold start time matters a lot. For small code that will only be run 2 times forever, LLM can generate new version before Rust compiles the old. Roc has the sweet spot of fast compilation and fast running, and sandboxed.

[Albert2088 @water2088]

@zeddotdev Super excited to try it

[Albert2088 @water2088]

@poteto @davis7 @bot @theo That’s brutal

[Albert2088 @water2088]

Opus5.5 optimized my video decoder 8X faster.

The previous slow version was written by AI as well. AI can't produce a fast version from the beginning.

But after tweeking the code many days, AI was able to see a clearer picture of the code intention and thus do a much better job

[Albert2088 @water2088]

@rtfeldman Have been speculating Roc for 2 years and I feel 2026/2027 will finally be the year to do it. AI just makes this kind of exotic experiment 100X cheaper for me. Started a thread in zulip and let's see if any developers are interested. I will do a version myself.

[Albert2088 @water2088]

@rtfeldman I wonder if it’s practical to use Roc as a library, meaning statically linking, so that I can call it as a function. But my main program is written in Rust, not Zig. The use case is a generative plugin system where the user generates plugins in Roc for their own use.

[Albert2088 @water2088]

@ChShersh So you changed a job once a year before 2022?

[Albert2088 @water2088]

@poteto Chat is just an append-only document. The agent should of course have some way of recording the interaction history and display that history. But the history does not have to be the main UI

[Albert2088 @water2088]

@_chenglou Absolutely beyond my comprehension

[Albert2088 @water2088]

@SteveOnSpeed somewhere in the middle. I have 9 years of software engineering experience, but not yet an expert in any domains. A normal generalist. It took me 2 months to get an interview from 10 applications. It’s already faster than when I graduated.

[gingerBill @TheGingerBill]

One of the big problems with everyone vibe-coding everything rather than actually just checking and reviewing the output, is that LLMs are going to train themselves on their own output.

LLMs are going to have problems related to inbreeding. Other AIs are beginning to have this.

This model collapse or Habsburg AI is a serious problem in all of the kinds of models (LLM, generative, etc), and Goddamn do I worry it'll be an issue sooner rather than later.

They could stop say "only train on data before this date", but that'll limit them too for when new things happen.

I believe many companies adding the fingerprinting/watermarks is probably a big reason now of doing it. They are marketing to allow people to see if something is AI made, when in reality they are adding it to prevent their own models from consuming it, producing inbreed cannibalism.

[Albert2088 @water2088]

@neogoose_btw No video gen models used?

[Albert2088 @water2088]

@TheGingerBill @Hal_Wayland I view this as an anti-fragile's nonlinear problem.

10KG lift for 10 times != 100KG lift 1 time.

AI can do 10KG lift 1000 times very quick and that's like 99% of all jobs. But to build strength, a man has to lift heavier, to solve the 1% problem.

[Albert2088 @water2088]

@huacnlee Then it opens a posibility of implementing generative UI for complex apps.

I am making a video editor in GPUI. I envision an editor with basic UIs and users can generate new UIs for their own work style. It's like vibe coding but for customizing a productivity app.

[Albert2088 @water2088]

Haven’t been interviewed for years. Just did the 1st interview in Tokyo in years. Interviewers were very nice.

I also felt that my knowledge was lagging a bit for enterprise, as a side effect of working on exotic personal projects for too long.

[Albert2088 @water2088]

@FishAudio Drama

[Albert2088 @water2088]

@poteto The reset guyger is not that bad

[Albert2088 @water2088]

It’s great to write some code in Tokyo on a Tuesday.