Daniel Karuna
@karuna.dev
always enthusiastic but sometimes not really
There is no way it is a healthy relationship (even for a parasocial), if every time time you message ChatGPT you lowkey think something along the lines of “ok let’s see what kind of bullshit you’ll spit out this time”
August 6, 2025 at 12:07 AM
There is no way it is a healthy relationship (even for a parasocial), if every time time you message ChatGPT you lowkey think something along the lines of “ok let’s see what kind of bullshit you’ll spit out this time”
After years of chatting with LLMs, I think I subconsciously assigned an identity to this abstract concept of “AI”. The problem is, I see it as an unreliable individual.
I mean, we all know LLMs are unreliable, but this emotional impression aspect makes working with them infinitely more exhausting
I mean, we all know LLMs are unreliable, but this emotional impression aspect makes working with them infinitely more exhausting
August 5, 2025 at 11:51 PM
After years of chatting with LLMs, I think I subconsciously assigned an identity to this abstract concept of “AI”. The problem is, I see it as an unreliable individual.
I mean, we all know LLMs are unreliable, but this emotional impression aspect makes working with them infinitely more exhausting
I mean, we all know LLMs are unreliable, but this emotional impression aspect makes working with them infinitely more exhausting
Spent a couple of hours writing a unified log collector to feed into Claude Code: github.com/iamkd/llmtail. Still a lot of work to be done, especially around surfacing the actually useful logs, filtering by error type etc, but anecdotally giving access to logs noticeably improves the iterations
GitHub - iamkd/llmtail
Contribute to iamkd/llmtail development by creating an account on GitHub.
github.com
July 13, 2025 at 12:23 AM
Spent a couple of hours writing a unified log collector to feed into Claude Code: github.com/iamkd/llmtail. Still a lot of work to be done, especially around surfacing the actually useful logs, filtering by error type etc, but anecdotally giving access to logs noticeably improves the iterations
Vibe coding is exactly the way I built my first apps and small games when I was like 12 years old. My context window was small and I did not have a big enough dataset to generate the accurate embeddings
March 11, 2025 at 12:18 AM
Vibe coding is exactly the way I built my first apps and small games when I was like 12 years old. My context window was small and I did not have a big enough dataset to generate the accurate embeddings
Somebody really thought it was a cool idea
March 4, 2025 at 12:25 AM
Somebody really thought it was a cool idea
Just finished writing a (longer than expected) post about some problematic loading behavior in Next.js, hope it'll be useful to someone: karuna.dev/nextjs-prelo...
Next.js Images, Preloading, and React Suspense
A story about the unclear behavior of lazy image loading in Next.js, especially when used with React Suspense. Quite useful to know if you are trying to optimize for performance or SEO (load times/LCP...
karuna.dev
March 3, 2025 at 1:19 AM
Just finished writing a (longer than expected) post about some problematic loading behavior in Next.js, hope it'll be useful to someone: karuna.dev/nextjs-prelo...
Incredible feat of engineering: youtu.be/0mCsluv5FXA
TypeScript types can run DOOM
YouTube video by Michigan TypeScript
youtu.be
February 28, 2025 at 11:59 AM
Incredible feat of engineering: youtu.be/0mCsluv5FXA
Crazy stuff happening in Instagram Reels right now, feeling sorry for the folks at Meta who’ll have to clean this up
February 27, 2025 at 2:12 AM
Crazy stuff happening in Instagram Reels right now, feeling sorry for the folks at Meta who’ll have to clean this up
It just dawned on me that the most precious part of Cursor as a company is the behavior telemetry.
I always thought of these tools as “gather open source -> LLM -> copilot”.
But Cursor collects the behavior(!!!). Crazy things could be done with it once we stop thinking about AI in terms of LLMs
I always thought of these tools as “gather open source -> LLM -> copilot”.
But Cursor collects the behavior(!!!). Crazy things could be done with it once we stop thinking about AI in terms of LLMs
February 24, 2025 at 6:48 PM
It just dawned on me that the most precious part of Cursor as a company is the behavior telemetry.
I always thought of these tools as “gather open source -> LLM -> copilot”.
But Cursor collects the behavior(!!!). Crazy things could be done with it once we stop thinking about AI in terms of LLMs
I always thought of these tools as “gather open source -> LLM -> copilot”.
But Cursor collects the behavior(!!!). Crazy things could be done with it once we stop thinking about AI in terms of LLMs
> Claude 3.7 Sonnet shows particularly strong improvements in coding and front-end web development.
This makes me think of how invaluable the telemetry from Cursor is for basically any AI company. They have privacy mode, but I’m sure not everyone uses that
This makes me think of how invaluable the telemetry from Cursor is for basically any AI company. They have privacy mode, but I’m sure not everyone uses that
February 24, 2025 at 6:41 PM
> Claude 3.7 Sonnet shows particularly strong improvements in coding and front-end web development.
This makes me think of how invaluable the telemetry from Cursor is for basically any AI company. They have privacy mode, but I’m sure not everyone uses that
This makes me think of how invaluable the telemetry from Cursor is for basically any AI company. They have privacy mode, but I’m sure not everyone uses that
Had a random (and likely unoriginal) thought: if a gazillion-parameter AI (LLM or not) were trained on everything humanity had, then each of our personalities — or even identities — could be represented as a relatively small vector of weights. Possibly in the ballpark of a few kilobytes per person.
February 21, 2025 at 8:13 PM
Had a random (and likely unoriginal) thought: if a gazillion-parameter AI (LLM or not) were trained on everything humanity had, then each of our personalities — or even identities — could be represented as a relatively small vector of weights. Possibly in the ballpark of a few kilobytes per person.
For people who do web with Next.js: `ReactDOM.preload` and, therefore, Next.js's `<Image priority />` will not always work if it is behind a Suspense boundary.
<link rel="preload"> will be streamed to the end of the body and your precious LCP and image loading times will drop. In-depth post coming
<link rel="preload"> will be streamed to the end of the body and your precious LCP and image loading times will drop. In-depth post coming
February 20, 2025 at 12:14 AM
For people who do web with Next.js: `ReactDOM.preload` and, therefore, Next.js's `<Image priority />` will not always work if it is behind a Suspense boundary.
<link rel="preload"> will be streamed to the end of the body and your precious LCP and image loading times will drop. In-depth post coming
<link rel="preload"> will be streamed to the end of the body and your precious LCP and image loading times will drop. In-depth post coming
Wrote a small post about the fact that #ff0000 is not the reddest red. Inspired by a recent hackernews thread/post, links inside: karuna.dev/colormaxxing/
Colormaxxing
A short post about the new color space in CSS that allows users to render colors beyond the default sRGB. It compares the classic sRGB colors with similar ones outside the sRGB color space.
karuna.dev
February 17, 2025 at 4:00 PM
Wrote a small post about the fact that #ff0000 is not the reddest red. Inspired by a recent hackernews thread/post, links inside: karuna.dev/colormaxxing/
Nine patch versions in ten days, exciting
February 3, 2025 at 1:41 PM
Nine patch versions in ten days, exciting
Did they really artificially darken The Weeknd’s filming footage to create a fake impression of iPhone’s low light performance?
The white balance diff between scenes was quite apparent. I’m no pro, but to get that image quality on an iPhone the set def should’ve been filled with high power lights
The white balance diff between scenes was quite apparent. I’m no pro, but to get that image quality on an iPhone the set def should’ve been filled with high power lights
September 9, 2024 at 7:29 PM
Did they really artificially darken The Weeknd’s filming footage to create a fake impression of iPhone’s low light performance?
The white balance diff between scenes was quite apparent. I’m no pro, but to get that image quality on an iPhone the set def should’ve been filled with high power lights
The white balance diff between scenes was quite apparent. I’m no pro, but to get that image quality on an iPhone the set def should’ve been filled with high power lights
Also their TikTok-ish music discovery feed is really cool. Spotify’s “Discover Weekly” really sucks, don’t get why people love it
September 6, 2023 at 2:40 PM
Also their TikTok-ish music discovery feed is really cool. Spotify’s “Discover Weekly” really sucks, don’t get why people love it
I keep paying for both Spotify and Apple Music and all their recommendations are always the same songs with half of them being already in my library.
Enter YouTube Music, random shit every time I refresh the page. MUCH better
Enter YouTube Music, random shit every time I refresh the page. MUCH better
September 6, 2023 at 2:39 PM
I keep paying for both Spotify and Apple Music and all their recommendations are always the same songs with half of them being already in my library.
Enter YouTube Music, random shit every time I refresh the page. MUCH better
Enter YouTube Music, random shit every time I refresh the page. MUCH better