macOS. You have 500GB free. me. Deletes and empties trashcan of 200GB of Xcode stuff macOS. You have 400GB free. macOS 2026 everyone!
Yousuf K.
@yousufkalim.com
Building @HQForWork • Staff Engineer @LiveRecover Software Engineer. Investor. Husband. Dad. Writing about AI, software, stocks, and growth.
The Open Secure AI Alliance launched on July 27 with Nvidia, Microsoft, IBM, Hugging Face and 30+ others. No OpenAI, Google or Anthropic. Quite a missing row.
The new US framework for reviewing frontier AI models is reportedly secret and voluntary. That’s a strange combination for something meant to build trust 🤯
Opus 5 is one of first models that you dislike the more you use it. Almost every time I give it tasks it either gets super lazy or it over works the entire project. Fable 5 is smarter and GPT 5.6 is best for everyday work. I don’t use Opus 5 anymore
On 4th of August, Chatgpt started turning pastes over 10k characters into attachments instead of dumping them into the composer. Small change but much nicer default for logs and long docs.
The most dangerous button in an AI app is still “apply all” Change my mind
Vercel Functions added WebSocket support in public beta on June 22. Standard Node.js libraries, no special setup. Finally.
AI can build a spreadsheet in 30 seconds and then spend ten minutes explaining why it changed the formulas.
Apple’s Foundation Models framework now exposes on-device and cloud models through one Swift API. Clean abstraction. The app shouldn’t care where the model runs.
Software engineering hasn't changed at its core. It's still hard to learn and it still takes years. What changed is that you can now produce software without understanding how it works.
16 GB RAM felt like plenty until every app became Chromium with an AI sidebar.
Maturing is realizing you need to start investment in hardware to run open source because frontier models will jack up the prices in the next 10 months and one shot everyone.
You should try and learn to give more attention to the process and the work itself rather than the end result. Way of working is always better to think about than the reward.
Millions of weekly npm downloads. Still no normal way to explain to relatives what I built.
If a usage reset is close and there’s plenty left, fast mode makes perfect sense. Finish the feature now without burning the next window.
Claude Science launched on June 30 with Jupyter, R, cluster access, reusable workflows and auditable artifacts in one workbench. Scientists already have enough tabs.
Voice AI has a brutal UX curve. Half a second of silence feels slower than five seconds in chat.
OpenAI’s July 8 audit found ~30% of SWE-Bench Pro tasks were broken. We’re reading way too much into tiny leaderboard jumps.
A CTO built a free tool over Christmas break. It now replaces software that costs $60,000/year.
The EU’s AI transparency rules started applying yesterday. Machine readable marks make sense. Visible AI labels can easily become the new cookie banners though.
AI engineer is becoming one of those job titles where you need another sentence to know what the person actually does 😂
GitHub Models was retired on July 30. We’re going to need a graveyard page for AI products at this rate.
I used to obsess over the perfect tech stack. Now I know the only stack that matters is the one that solves a real problem today.
Opus 5 As usual I have no take, other than: If you've designed your harness/environment well, and not over-optimised around a specific model, today should feel like any other day... ...with a slightly lower failure rate
“Models barely hallucinate anymore” sounds true until the answer is about something you actually know well.
The author of Clean Code doesn’t review code written by AI. He trusts the guardrails. You can’t really review AI generated code seriously. You need strict guardrails, a good harness and a clear plan. That’s it
