Andrew Prifer
@prif.bsky.social
UI, UX, 3D things. Graphics, UI & Interaction Design @ Luma AI. Prev: Theatre.js, Spline
https://bento.me/andrewprifer
https://bento.me/andrewprifer
What is the reason that the macOS genie animation looks like a trapezoid instead of a smooth bezier near the bottom?
November 24, 2025 at 5:03 PM
What is the reason that the macOS genie animation looks like a trapezoid instead of a smooth bezier near the bottom?
I find it very suspicious that the single benchmark where Gemini 3 Pro didn’t show any improvement is the one that most punishes overfitting to the training data by nondeterministically sourcing from a large set of novel real world problems.
November 18, 2025 at 4:56 PM
I find it very suspicious that the single benchmark where Gemini 3 Pro didn’t show any improvement is the one that most punishes overfitting to the training data by nondeterministically sourcing from a large set of novel real world problems.
Investors please form an orderly line
November 17, 2025 at 5:02 PM
Investors please form an orderly line
Haven’t seen anything like this before
November 16, 2025 at 6:48 PM
Haven’t seen anything like this before
GPT-5 now supports constrained decoding with context free grammars. This is similar to structured outputs, but instead of forcing a schema, we can create custom languages and force the LLM to adhere to the syntax.
Huge.
Huge.
November 11, 2025 at 4:58 PM
GPT-5 now supports constrained decoding with context free grammars. This is similar to structured outputs, but instead of forcing a schema, we can create custom languages and force the LLM to adhere to the syntax.
Huge.
Huge.
I hear a lot of talk about code mode tool calling with @AnthropicAI and @threepointone publishing some really interesting findings. Fun fact, the below demo is very real, and one of the reasons it was able to be so accurate (3 years ago!!!) is because of code gen tool use.
x.com/theatre_js/...
x.com/theatre_js/...
November 10, 2025 at 5:04 PM
I hear a lot of talk about code mode tool calling with @AnthropicAI and @threepointone publishing some really interesting findings. Fun fact, the below demo is very real, and one of the reasons it was able to be so accurate (3 years ago!!!) is because of code gen tool use.
x.com/theatre_js/...
x.com/theatre_js/...
Crazy that the encryption export laws of the Cold War still aren’t completely gone yet.
October 23, 2025 at 11:37 PM
Crazy that the encryption export laws of the Cold War still aren’t completely gone yet.
The quickest hashing library for react native by @pioner_dev
The thing I'm building makes heavy use of caching at every step and this has been invaluable for calculating cache keys from stuff that’d be too large to key by directly.
github.com/pioner92/re...
The thing I'm building makes heavy use of caching at every step and this has been invaluable for calculating cache keys from stuff that’d be too large to key by directly.
github.com/pioner92/re...
GitHub - pioner92/react-native-xxhash: 💪 A high-performance React Native library for generating xxHash hashes using C++ and JSI
💪 A high-performance React Native library for generating xxHash hashes using C++ and JSI - pioner92/react-native-xxhash
github.com
October 22, 2025 at 5:02 PM
The quickest hashing library for react native by @pioner_dev
The thing I'm building makes heavy use of caching at every step and this has been invaluable for calculating cache keys from stuff that’d be too large to key by directly.
github.com/pioner92/re...
The thing I'm building makes heavy use of caching at every step and this has been invaluable for calculating cache keys from stuff that’d be too large to key by directly.
github.com/pioner92/re...
If Lovable can one-shot your project, it probably isn't worth doing.
October 22, 2025 at 3:58 PM
If Lovable can one-shot your project, it probably isn't worth doing.
Lord, I see what You're doing for others; grant me the fortune of a half-billion exit right before OpenAI kills my product, amen.
October 22, 2025 at 12:04 PM
Lord, I see what You're doing for others; grant me the fortune of a half-billion exit right before OpenAI kills my product, amen.
As far as real world utility goes, coding models are a really smart typist with a CS degree. It's hard for me to agree with the crowd declaring the death of coding in <insert number of months>. Even today if I am any less specific in my prompts than this, the llm veers off in undesirable directions.
October 21, 2025 at 5:03 PM
As far as real world utility goes, coding models are a really smart typist with a CS degree. It's hard for me to agree with the crowd declaring the death of coding in <insert number of months>. Even today if I am any less specific in my prompts than this, the llm veers off in undesirable directions.
OpenRouter charging 5% is crazy, because it prices manually wiring up your routing at 1/20 of the value AI models provide. Absurd for something that’s at most 400 loc, and then you'd even get nice things like sorting by historical latency as measured from your backend.
October 21, 2025 at 3:56 PM
OpenRouter charging 5% is crazy, because it prices manually wiring up your routing at 1/20 of the value AI models provide. Absurd for something that’s at most 400 loc, and then you'd even get nice things like sorting by historical latency as measured from your backend.
When I said I wished we moved away from chat boxes as a means to interact with AI, I didn't mean I want no interaction at all.
x.com/AndrewPrife...
x.com/AndrewPrife...
October 20, 2025 at 5:00 PM
When I said I wished we moved away from chat boxes as a means to interact with AI, I didn't mean I want no interaction at all.
x.com/AndrewPrife...
x.com/AndrewPrife...
If only someone built the AI infrastructure for the agentic era. Is there anyone already doing this? Anyone at all?
October 20, 2025 at 4:04 PM
If only someone built the AI infrastructure for the agentic era. Is there anyone already doing this? Anyone at all?
Learning about Lisp and I keep seeing the word sexpression everywhere. Man the 60's were wild.
October 17, 2025 at 5:21 AM
Learning about Lisp and I keep seeing the word sexpression everywhere. Man the 60's were wild.
it's not looking good bros
October 16, 2025 at 5:57 PM
it's not looking good bros
This extreme subsidizing of users and selling things at a loss is the wrong axis to compete on imo and I can't wait for the bubble to burst. I know game theory won't let us have it any other way, but man is it annoying.
October 16, 2025 at 5:00 PM
This extreme subsidizing of users and selling things at a loss is the wrong axis to compete on imo and I can't wait for the bubble to burst. I know game theory won't let us have it any other way, but man is it annoying.
my primary source of anxiety
October 16, 2025 at 4:01 PM
my primary source of anxiety
This is still one of the most audacious things I’ve ever put in a real-world app, but somehow what I’m building now feels a 1000x cooler. The future is bright.
x.com/andrewprife...
x.com/andrewprife...
October 14, 2025 at 3:57 PM
This is still one of the most audacious things I’ve ever put in a real-world app, but somehow what I’m building now feels a 1000x cooler. The future is bright.
x.com/andrewprife...
x.com/andrewprife...
The hard thing about microVM-based services like @flydotio is that you need to know Docker.
The easy thing about microVM-based services is that you only need to know Docker.
x.com/AndrewPrife...
The easy thing about microVM-based services is that you only need to know Docker.
x.com/AndrewPrife...
October 13, 2025 at 5:01 PM
The hard thing about microVM-based services like @flydotio is that you need to know Docker.
The easy thing about microVM-based services is that you only need to know Docker.
x.com/AndrewPrife...
The easy thing about microVM-based services is that you only need to know Docker.
x.com/AndrewPrife...
My progressive blur library has weirdly picked up steam lately. If we can reach 200 stars I'll make a React Native version.
October 10, 2025 at 4:02 PM
My progressive blur library has weirdly picked up steam lately. If we can reach 200 stars I'll make a React Native version.
Async generators are becoming the standard way to create streaming outputs (AI SDK tool calls, TanStack Start). Here's a fun challenge you WILL need to solve at some point, and which surprisingly has no builtin equivalent.
October 9, 2025 at 3:56 PM
Async generators are becoming the standard way to create streaming outputs (AI SDK tool calls, TanStack Start). Here's a fun challenge you WILL need to solve at some point, and which surprisingly has no builtin equivalent.
Learned the hard way: When using tool calls or structured outputs with LLMs, tool and schema descriptions barely have any sway.
To actually steer LLMs, always put this directly in your system prompt.
To actually steer LLMs, always put this directly in your system prompt.
October 1, 2025 at 4:03 PM
Learned the hard way: When using tool calls or structured outputs with LLMs, tool and schema descriptions barely have any sway.
To actually steer LLMs, always put this directly in your system prompt.
To actually steer LLMs, always put this directly in your system prompt.
The absolute GOAT of coding music is the OST of AC Unity by a long shot.
Especially vol 2 by @SarahSchachner is all bangers.
Especially vol 2 by @SarahSchachner is all bangers.
September 29, 2025 at 6:04 PM
The absolute GOAT of coding music is the OST of AC Unity by a long shot.
Especially vol 2 by @SarahSchachner is all bangers.
Especially vol 2 by @SarahSchachner is all bangers.