iknowharsha

I wanted to watch The Odyssey in IMAX. I'm lucky, there's an actual IMAX...

Fri, 24 Jul 2026 00:00:00 GMT

I wanted to watch The Odyssey in IMAX. I'm lucky, there's an actual IMAX theater near me here in Indianapolis, at the Indiana State Museum. It's one of 41 theaters in the entire world that can project it in true IMAX 70mm, and one of only 25 in the US.

But every show was sold out till August 20th, and the museum hadn't announced anything yet about new dates after that. I didn't want to check the site every single day myself, I wanted to know the moment new shows dropped.

While looking into it, I found Distill. It's a browser extension that watches a webpage for you and tells you when something on it changes. You just point it at a page, tell it how often to check, and how you want to be notified, email or a popup.

I set it up on the museum's ticket page. Now it checks for me, and I'll get notified the second new dates go up, instead of finding out later that they already sold out again.

Worth a look if you deal with this kind of thing too, waiting on tickets, waiting on restocks, whatever it is.

Saw this tweet about TogetherLink and running GLM 5.2 with it. My first ...

Thu, 16 Jul 2026 00:00:00 GMT

Saw this tweet about TogetherLink and running GLM 5.2 with it. My first thought was: these models are open source, so why are we paying companies to use them? Asked Claude to break the whole thing down. Leaving the notes here for future me.

TogetherLink
A free, open source tool. It lets you temporarily run coding tools like Claude Code or Codex on a different AI model, without changing your normal settings. Install it once, run a short command like tclaude to switch models just for that session, close it and everything goes back to normal. The tool itself costs nothing.

What does "open source" actually mean?
The model's code and weights are publicly available, anyone can download and inspect them. Think of it like a free recipe. The recipe is open, but you still need a kitchen, ingredients, and someone to cook it.

Why open source models still cost money
Running these models needs powerful GPUs, heavy electricity, and technical setup. Most people don't have that hardware. Companies like Together AI buy the GPUs and host the model on their servers. When you use it, you're paying for computing power and hosting, not for the model itself.

What is a token?
A token is roughly a chunk of a word, about 1000 tokens equals 750 words.
- Input tokens: what you send to the AI
- Output tokens: what the AI writes back
- Cached tokens: repeated text the AI already saw before, cheaper to process

What tokens cost in real money
GLM 5.2 on Together AI:
- Input: $1.40 per 1 million tokens
- Output: $4.40 per 1 million tokens
- Cached input: $0.26 per 1 million tokens
A normal chat of a few thousand tokens costs a tiny fraction of a cent. The bill only adds up with heavy, repeated use, like an app calling the model thousands of times a day.

Why GLM 5.2 is cheap
Together AI hosts it efficiently and competes on price. Platforms like OpenRouter sell it even cheaper by routing requests to whichever host is cheapest, and using caching to cut repeated costs by 60 to 80 percent.

Who makes these models
- GLM: Z.ai (also called Zhipu AI), a Chinese company known for strong coding and long, multi-step tasks
- Kimi: Moonshot AI, a Chinese startup known for very long context windows, holding whole books or codebases in memory at once
- DeepSeek: another major Chinese lab, competing closely with the closed models from OpenAI, Google, and Anthropic

Inside the iconic Pentagram Design studio in New York

Wed, 15 Jul 2026 00:00:00 GMT

Lunch break watch: Really loved how this designer thinks about design.

Inside the iconic Pentagram Design studio in New York by Piercarla Tozzi

Been experimenting with Flora to design my footer section and it's so fu...

Thu, 09 Jul 2026 00:00:00 GMT

Been experimenting with Flora to design my footer section and it's so fun.

The first image was actually created with ChatGPT, and I feel like it got the style I was expecting pretty well. Then I used Claude to structure the prompt as detailed JSON, pasted that into Flora, and iterated from there. The right prompt keeps the image structure and style intact across iterations.

Honestly just having a lot of fun with this.

Footer experiments in Flora

Been using Compresto lately to compress images and videos before uploadi...

Fri, 03 Jul 2026 00:00:00 GMT

Been using Compresto lately to compress images and videos before uploading them to this site. Does the job, and it was a one-time purchase instead of another subscription, which is rare these days.

Every time I designed interfaces by vibecoding straight into code withou...

Sun, 28 Jun 2026 00:00:00 GMT

Every time I designed interfaces by vibecoding straight into code without touching Figma, things would end up looking out of place. The layout was technically right, but something always felt visually off.

Today I went through the Means & Methods module in Josh's Interface Craft. It answered why things weren't looking optically right.

So well explained, especially with the designed artifacts for each case. Amazing stuff.

Tweet by storyofharsha

Fri, 26 Jun 2026 00:00:00 GMT

Figma's shaders are sooo good.
Small experiment I tried. Video was made with Figma motion too.

Even though it's simple, crazy that prompting can create a video inside Figma. That's nuts.

Tweet by storyofharsha

Tweet by figma

Wed, 24 Jun 2026 00:00:00 GMT

Figma woke up listening to Everything At Once.

Tweet by figma

I was always amazed by how people figured out that you could print full-...

Fri, 12 Jun 2026 00:00:00 GMT

I was always amazed by how people figured out that you could print full-color images with just three inks, Cyan, Magenta & Yellow, and tiny dots.

Where each color gets its own plate, each plate has its own screen angle, and you have to line all of them perfectly or the whole image comes out blurry or fringed.

That curiosity led me to build CMYK Studio over a few days.

Tweet - 2026-06-09

Tue, 09 Jun 2026 00:00:00 GMT

Claude Fable dropped today, leaves the Pro plan on June 22nd, and burns through a session so fast I get maybe 3 uses total. But you're not nerdy enough to understand that.

The Thinker by Rodin: Great Art Explained

Sun, 07 Jun 2026 00:00:00 GMT

There's something about this sculpture that made me just stare at it. I still don't know why I like it so much.

My best guess is that it just looks like me, always thinking about something random that sometimes doesn't even make sense. Went down the rabbit hole to find out what it actually was.

It's The Thinker by Rodin. This is the story behind it.

The Thinker by Rodin: Great Art Explained by Great Art Explained

how to enter side doors

Sat, 06 Jun 2026 00:00:00 GMT

As someone early in their design career, constantly applying for roles, this is the kind of hard truth I wish I'd understood sooner. Worth a read if you're in the same boat.

how to enter side doors

Always thought it was hard to write case studies in Notion and pull them...

Fri, 05 Jun 2026 00:00:00 GMT

Always thought it was hard to write case studies in Notion and pull them into your site with your own styling. All my life I wrote them in Framer.

Today I finally did it using Claude. Built custom components for my needs, connected the site to Notion through their API. Works exactly how I wanted.

Now working on small animations to highlight information. Let's see how this goes.

Animation Vocabulary

Mon, 01 Jun 2026 00:00:00 GMT

Emil shared this. Having the right vocabulary for animations makes it way easier to describe what you want to an AI.

Animation Vocabulary

YKW, I was completely wrong about Opus 4.8. It simply just comsumes w...

Sat, 30 May 2026 00:00:00 GMT

Me when the 5-hour session ends after 5 prompts.

YKW, I was completely wrong about Opus 4.8.

It simply just comsumes way too much tokens per session, not even 3 days, I havve already reached the weekly limit 😡. And I already feel that it's been nerfed

Right now, GPT-5.5 feels better than Opus 4.7 and 4.8.

Building a CMYK editor. This feels oddly satisfying to work on. I can'...

Fri, 29 May 2026 00:00:00 GMT

Building a CMYK editor.

This feels oddly satisfying to work on.

I can't imagine doing this back in the day when people's jobs were to prepare images for print and tweak all these controls by hand.

Must have been painful.

This version is fun though. Adding it to the Garden soon.

Tweet by claudeai

Thu, 28 May 2026 00:00:00 GMT

Anthropic dropped Claude Opus 4.8 today and it's so good.

I'm on Pro so the session runs out in like 7-8 prompts, but the problem solving and task execution make up for it. Visual tasks especially, it actually gets what I'm pointing at.

Was using GPT 5.5 a few days back and thought it had pulled ahead of Opus 4.7. Anthropic caught up fast.

Tweet by claudeai

Tweet by darkroomdevs

Sun, 24 May 2026 00:00:00 GMT

Found Lenis.dev today: an open-source JS library for smooth, fluid scrolling. Built by Darkroom Engineering.

Tweet by darkroomdevs

Tweet by rsms

Sat, 23 May 2026 00:00:00 GMT

Wish someone had told me this earlier.

Tweet by rsms

The way they handle empty states. No illustrations, no copy that tries too hard. Just a quiet line of text.

Linear's empty states

Started learning from the two courses I bought from Josh Puckett's Inter...

Wed, 20 May 2026 00:00:00 GMT

Started learning from the two courses I bought from Josh Puckett's Interface Craft and Emil's animation course.

Soooooo good !!!!!!!!!!!!!! learnt a lot so cookin up my portfolio rn.

Ended my internship at Woods Creative today. 10 months, Aug 2025 to May ...

Fri, 01 May 2026 00:00:00 GMT

Ended my internship at Woods Creative today. 10 months, Aug 2025 to May 2026. Had a wonderful time working with Amy, Chris, Susan, and Ella. Amazing people.

Chris left me with his Irish pearl of wisdom: "nothing will go as you planned, but it will go better than you imagined, and I promise you'll feel that someday."