Ahsan Khalid 🇵🇰

Software Engineer building production SaaS | AI agents & automation | Sharing the systems, experiments & lessons behind my work.

[Ahsan Khalid 🇵🇰 @Ahsankhalid23]

@marcelkargul https://t.co/nGXQgAVW8a

[Ahsan Khalid 🇵🇰 @Ahsankhalid23]

Just saw tsrs benchmarks.

It’s a Rust port of the TypeScript compiler and runs up to 8× faster while using about 70% less memory than TS 7.0 with the same worker count.

The default adaptive config trades a bit more RAM for extra speed.

[Ahsan Khalid 🇵🇰 @Ahsankhalid23]

Daily AI updates | Day 11

Explored DeepInfer — the inference engine behind DeepSeek’s services.

It ships offline inference with optimized kernels, a zero‑overhead async batch scheduler and batch‑invariant results.

Supports NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs.

[Ahsan Khalid 🇵🇰 @Ahsankhalid23]

WhatsApp slipped AI directives into its help pages, telling the model not to suggest the European data‑protection regulator.

If you feed public docs into an LLM, hidden snippets can silently bias its answers. Builders need to audit source text, not just what users see.

[Ahsan Khalid 🇵🇰 @Ahsankhalid23]

@olegzanx Yeah I’d want that too. The quiz is cool and all, but if I just gave a good answer about the webhook retry it should just become a test automatically instead of disappearing after the push.

[Ahsan Khalid 🇵🇰 @Ahsankhalid23]

I didn't expect a git hook to quiz me before push.

Sphinx pauses a push, reads the diff and asks a question about the change’s behavior – e.g., what happens if a Stripe webhook retries.

Answer it, then the push continues.

[Ahsan Khalid 🇵🇰 @Ahsankhalid23]

Daily AI updates | Day 10

LLM agents can get binding human directives that survive planning cycles.
SAI adds MCP overrides (Constraint, Priority, Strategy) schema‑validated and prepended to prompts, keeping authority across planning cycles fine‑tuning‑free.

[Ahsan Khalid 🇵🇰 @Ahsankhalid23]

The weird part is you can query the live web for $0.75 per 1 000 successful requests.

DeepSeek’s Search & Fetch APIs expose a raw JSON endpoint that returns structured results (Search) or a page’s Markdown (Fetch), letting agents and RAG apps pull live content.

[Ahsan Khalid 🇵🇰 @Ahsankhalid23]

Turns out the Claude Code team measures success by revert and incident rates.

They aim for PRs that merge with at most one human reviewer, as long as those rates don’t climb.

Human time is saved, not shifted.

[Ahsan Khalid 🇵🇰 @Ahsankhalid23]

Daily AI updates | Day 9

Turns out Gemini 4’s Argon model can write a StarCraft: Brood War bot in under an hour.

StarSkirmish gives each LLM an hour to produce a C++ Protoss bot, then pits it against other LLM‑generated and human bots. The win‑rate is the model’s score.

[Ahsan Khalid 🇵🇰 @Ahsankhalid23]

90% of US work hours are on tasks barely covered by AI benchmarks 🤔

Only 208 of 923 occupations have a task with ≥5 benchmark prompts, and 6 of the 22 job families have none.

The map shows most benchmarks focus on the biggest labor markets.

[Ahsan Khalid 🇵🇰 @Ahsankhalid23]

noticed DEsk on GitHub – a Wayland desktop that goes skeuomorphic, like iOS 6.

• 865 ★, 24 forks
• Beta v0.1.2 with installer for Fedora
• Commands: desk, desk‑session, desk‑settings

It renders a room with a walnut desk, paper‑like windows, and a drawer‑dock. 🤔

[Ahsan Khalid 🇵🇰 @Ahsankhalid23]

Daily AI updates | Day 8

Turns out Muse Spark Open – an 816 B‑parameter multimodal model – runs on two Apple M5 Ultra Mac Studios via Apple’s MLX framework.

[Ahsan Khalid 🇵🇰 @Ahsankhalid23]

Just saw Krea 3.

It's an image‑editing model built for production that only changes what you ask.

According to the page it runs in under a second and costs $0.06 per image—about 4× faster and 3× cheaper than Nano Banana Pro.
🤔

[Ahsan Khalid 🇵🇰 @Ahsankhalid23]

The weird part is how a font can change your debugging flow.

Iosevka 34.9.0 ships an open‑source slab‑serif monospace variant built for code and terminals.
Its monospace glyphs are exactly ½ em wide and you can pick Normal or Extended width for tighter or breezier columns.

[Ahsan Khalid 🇵🇰 @Ahsankhalid23]

Daily AI updates | Day 7

Turns out AI agents aren’t going rogue.

Experts say the hacks stem from labs skipping basic containment and monitoring.

CISOs should vet AI vendors and tighten those security basics.

[Ahsan Khalid 🇵🇰 @Ahsankhalid23]

Hot take: Anthropic’s Scion feels like Mononoke.
Both expose a “derived data” layer that speeds up blame.
The post notes Sapling got a “hard‑won lessons” credit, while Mononoke isn’t mentioned – that omission feels plagiarism.
I think giving proper credit is key.

[Ahsan Khalid 🇵🇰 @Ahsankhalid23]

Even a hot AI chatbot can go bankrupt.

https://t.co/nuhHbxbvoR announced a voluntary Chapter 11 to enable a court‑supervised sale; without a buyer, the platform will be wound down.

Reminder: hype doesn’t replace a solid revenue model especially with regulatory headwinds.

[Ahsan Khalid 🇵🇰 @Ahsankhalid23]

Eleven‑year‑old borrow‑checker bugs resurfaced.

GitHub issue shows covariant field assignment can bypass struct lifetime requirements, exposing unsoundness in Rust’s borrow checker.

If you rely on Rust for safety guarantees, keep an eye on this until it’s resolved.

[Ahsan Khalid 🇵🇰 @Ahsankhalid23]

Merge traffic exploded at Anthropic.

~2.5k merges/day in May 2026 → >30k/day by Sep 2026.
Source control became the biggest bottleneck, so they built Scion—an in‑house VCS—and fully migrated from git in five months.