iiinova — where your digital self keeps evolving

Yousuf K. @yousufkalim.com · 28/09/2026

Anthropic will definitely nerf Opus 5.5 to the ground after a week.

Yousuf K. @yousufkalim.com · 28/09/2026

They were asking everyone to slow down a couple of days ago 😂 The irony: Opus 5.5 dropped 10 days later, and now we’ve got Sonnet 5.5 🔥🔥

000

Yousuf K. @yousufkalim.com · 28/09/2026

As much as I hate Anthropic for fighting open source, and as critical as I've been, let's be real. Nobody touches them on model quality. They're a full month ahead of everyone else.

000

Yousuf K. @yousufkalim.com · 28/09/2026

Big step forward. Much needed, especially after what happened with OpenAI’s sandbox failure. That was scary af

000

Yousuf K. @yousufkalim.com · 17/09/2026

Amazing how AI companies convinced us that the problem isn’t the limits… it’s us for actually using the model we paid for 😂😒

000

Yousuf K. @yousufkalim.com · 17/09/2026

GPT Astra is 90% perfect... The other 10% I couldn't test because I keep reaching weekly limits😐

000

Yousuf K. @yousufkalim.com · 17/09/2026

Adding Apple TV to the iCloud subscription is genius.

000

Yousuf K. @yousufkalim.com · 14/09/2026

Kinda sucks seeing usage limits shrinking across every frontier model. We all anticipated a price hike because the costs are obviously heavily subsidized. But instead of raising prices, they’re shrinking the usage limits and leaving power users stranded. The great compute crunch has begun.

100

Yousuf K. @yousufkalim.com · 14/09/2026

Usage boosts from Claude are gone 🥲

000

Yousuf K. @yousufkalim.com · 10/09/2026

Not sure if it’s just me or Codex usage limits actually got worse, even on Astra low. I’m on the 20x plan, had only 3 simple sessions running on Astra low and it blew through all my usage within hours, then ate ~$130 credits while I still had banked resets available. Without a single notification.

100

Yousuf K. @yousufkalim.com · 09/09/2026

If a CVE still needs a human to open the PR, you don't have a factory.

101

Yousuf K. @yousufkalim.com · 09/09/2026

there's nothing more satisfying than shipping a database optimization Gives more dopamine than hearing ATM chrrrrr

000

Yousuf K. @yousufkalim.com · 09/09/2026

iPhone Duo should have been named iFold

000

Yousuf K. @yousufkalim.com · 09/09/2026

I’m still finding Fable a better choice for SWE. Astra is no doubt a great upgrade, but I think it shines more in other use cases like creativity and investment research

000

Yousuf K. @yousufkalim.com · 09/09/2026

Chat has won. It's all chat + computer from this point on

000

Yousuf K. @yousufkalim.com · 08/09/2026

how the fuck is OpenAI suddenly this far ahead of Anthropic? 😂

000

Yousuf K. @yousufkalim.com · 08/09/2026

Meet a developer who can't use AI because his company doesn't let him. He knows this is probably the last developer job he will ever have. If he gets laid off, he will be years behind everyone else. Scary place to be. I told him to find a new job.

000

Yousuf K. @yousufkalim.com · 08/09/2026

1 million token context sounds great until you realize the interesting problem is how many times your agent has to reread those 1 million tokens

000

Yousuf K. @yousufkalim.com · 04/09/2026

This doesn’t feel like a launch day. It feels like job interview day 😭

000

Yousuf K. @yousufkalim.com · 04/09/2026

Anthropic just reset the usage limits across all accounts 😂

000

Yousuf K. @yousufkalim.com · 04/09/2026

is Astra running on Cerebras or what? it's fucking blazing fast while somehow being more intelligent and efficient than 5.6 sol. i'm using it right now to fix memory corruption in a multithreaded C++, and the speed difference is kinda crazy pro tip: use md reasoning effort

000

Yousuf K. @yousufkalim.com · 01/09/2026

Claude is watermarking text now. Fable 5.1 and every model released after Aug 2 gets an imperceptible watermark on text outputs. everyone, not just EU users.

000

Yousuf K. @yousufkalim.com · 01/09/2026

Claude Fable 5.1 has been released. Excited to push it hard 🔥

010

Yousuf K. @yousufkalim.com · 31/08/2026

What's the library the model keeps adding that you already have?

000

Yousuf K. @yousufkalim.com · 31/08/2026

AI video summaries need timestamps on every point. Otherwise checking one suspicious detail means watching the whole thing anyway.

100

Yousuf K. @yousufkalim.com · 29/08/2026

Cursor's SpaceX deal comes with an unexpected casualty: OpenAI model access. OpenAI says it will wind down the contract after SpaceX acquired Cursor, with a proposed shutoff on November 12.

100

Yousuf K. @yousufkalim.com · 28/08/2026

Vercel open sourced vgpu, an agent first WebGPU library that can render in browser, node, cpu sandboxes and CI 🔥 Headless/CPU rendering is the big thing here. Agents can actually test shader work in CI instead of writing wgsl and hoping for it to work

150

Yousuf K. @yousufkalim.com · 28/08/2026

Cursor can now go from a blank prompt to a repo, live preview, and Vercel deployment This removes so much dumb setup for small apps, you can figure out if the idea is even worth keeping before touching repo/deploy stuff

010

Yousuf K. @yousufkalim.com · 28/08/2026

Perf + better defaults for humans and agents is the kind of win most teams only talk about.

010

Yousuf K. @yousufkalim.com · 28/08/2026

Chat history is not a knowledge base. It dies the second someone opens a new thread. You need a good harness, not a good model.

220

Yousuf K. @yousufkalim.com · 28/08/2026

A skill you never invoke is just another markdown file. Favor your harness by getting rid of those.

000

Yousuf K. @yousufkalim.com · 28/08/2026

What's still faster to do by hand?

000

Yousuf K. @yousufkalim.com · 28/08/2026

Stale docs are a lot more dangerous now. AI can turn them into actions instead of just ignoring them like the rest of us.

000

Yousuf K. @yousufkalim.com · 26/08/2026

Somewhere along the way we started benchmarking the setup instead of the actual work

010

Yousuf K. @yousufkalim.com · 26/08/2026

If the team uses three names for the same thing, the model will too. Don't complain

000

Yousuf K. @yousufkalim.com · 24/08/2026

Agent harness & harness engineering will not be phrases uttered on the internet 12 months from now. They're necessary for now as applied AI is in its infancy and complexity still hasn't been abstracted out of the user's experience, but this will all sit cleanly under "product" and "UX" very soon.

110

Yousuf K. @yousufkalim.com · 24/08/2026

These tools can be incredibly powerful for localized, well-defined tasks. When you combine them with systems that can generate those tasks from telemetry, you get a self-improving system.

000

Yousuf K. @yousufkalim.com · 24/08/2026

I think developers who refuse to use AI coding tools because they’re “real programmers” are making the same mistake people made refusing to use Stack Overflow. You don’t get extra credit for doing everything the slow way. The output matters more than your pride.

000

Yousuf K. @yousufkalim.com · 23/08/2026

Linus Torvalds being an AI skeptic not that long ago and now casually using it to debug kernel bugs 🙂 we’re moving fast

060

Yousuf K. @yousufkalim.com · 22/08/2026

I genuinely haven’t seen a sub with more value than SuperGrok Heavy. Here’s the exciting part: you can stretch it into almost 3x the usable AI capacity compared to paying for one rival sub if you use it right. First of all, some targeted users are getting it for $99/month.

100

Yousuf K. @yousufkalim.com · 22/08/2026

Insane for one release 🤣 I mean they rewrote the whole runtime in rust 😂 crazy things happening around the world!

010

Yousuf K. @yousufkalim.com · 21/08/2026

Rofl that both OpenAI and Asana management believed that it would take FIVE years to migrate Enzyme to React Testing Library. What amusing BS! I imagine it went like this: Manager: “How long will it take? Give me SOMETHING” Dev (annoyed): “FIVE YEARS OK?? Now leave me alone.”

030

Yousuf K. @yousufkalim.com · 21/08/2026

Your Mac is slowing you down. Moving to Linux has exponentially improved performance for my agents, in particular on the file system side.

030

Yousuf K. @yousufkalim.com · 20/08/2026

Opus 4.7 was where I started feeling it too. Not necessarily dumber, just way more constrained and weirdly less enjoyable to work with

020

Yousuf K. @yousufkalim.com · 20/08/2026

You probably want to replace the default writing and design skills that many harnesses have with something that suits your context and tastes.

020

Yousuf K. @yousufkalim.com · 20/08/2026

Claude's skill creator (literally, just ask it to make a skill) is still the best way to make reusable skills if you care about the details. ChatGPT tends to want to do magic & just do it for you, while Claude does tests and shows results, repeatedly asking for input & feedback.

010

Yousuf K. @yousufkalim.com · 20/08/2026

GitHub search still feels like punishment for not remembering the exact filename.

020

Yousuf K. @yousufkalim.com · 20/08/2026

Grep hygiene is such a good term for this. You can give an agent the whole repo and it’ll still find five abandoned plans before the file that actually matters.

100