A. Khansari

A. Khansari

@akhansari.tech

Gardening where people, tech, product, and AIs meet. Passionate about the terminal, sustainable software, and product thinking. https://akhansari.tech/ 🇪🇺🇫🇷

242 followers188 following393 posts

If I owned a company, I wouldn't let junior developers use LLMs to generate code. It's the only way to learn properly. If coding is truly solved, then we don't really need developers in the first place. We just need product owners with system and software architecture background.

Soon AI will answer to the ultimate question of life, the universe, and everything. But we won't have any clue what to do with it. True story.

It will be funny. Soon, programming languages or frameworks have to wait several months or even over a year for LLMs to learn their new features and releases.

Will those with infinite tokens be the oligarchs of tomorrow? Will open source stop them and make it more decentralised, or will (geo)politics and lobbying get in the way?

Dift has now commit history navigation. Open a commit to review its changes, or switch to the full file tree to see every file as it was at that point. You can also select two commits to compare. gitlab.com/akhansari/dift #git #diff #tui

An emerging pattern I see in companies: Specs are Gen AI, docs are Gen AI, codes are Gen AI, feedback is Gen AI, ... Replace meetings by AI agents and we're good.

Someone should use Jev to create a top-tier music recommendation system. It would be even better if it were based on a audio-scrobbler.

Do you also use Git staging to save good LLM outputs while you experiment? It lets you iterate on prompts and easily throw away untracked changes until you get a result you're happy with for your task/commit.

A. Khansari reposted

Oh, I think this will be my favourite nvim tip for discovering CLI usage. Run a command and replace the current buffer with its output: :%!your_command

Now that OpenAI, Claude, and xAI are all down at the same time, I have a Black Mirror scenario: In a few years, when skill atrophy hits hard and the US has enforced closed AIs, imagine there's a major outage at AI providers and nobody can use AI to fix the AI.

If the LLM supports it, prompt the narrative, business logic, and structural rules in your native lang instead of English, but keep the output in English. Expressing yourself in your mother tongue allows for maximum clarity and nuance, reducing ambiguity and quality degradation. It's a game changer.

The Designing with Types series is finally complete. Using #TypeScript and Effect to model domain rules, make illegal states unrepresentable, and turn primitive values into validated domain types. Inspired by Scott Wlaschin's original F# series. akhansari.tech/series/desig...

akhansari.techSeries: Designing with Types: A TypeScript + Effect ApproachUsing TypeScript's type system and Effect's functional primitives to model domains, express business rules, and make illegal states unrepresentable.

I know I keep saying this, but man, I'm in love with Fedora Atomic. It's the most stable and boring OS I've ever tested. It's just perfect. Plus, Cosmic Desktop makes it look beautiful.

Ok, I've switched completely to nubjs.com. It seems pretty good so far, so I'll evaluate it over the next few weeks. +nub -pnpm -tsx -concurrently

nubjs.comNub — an all-in-one toolkit for Node.jsNub is a TypeScript-first toolkit for Node.js: run TypeScript files on stock Node, a faster npm run, a pnpm-compatible package manager, and a built-in Node version manager. No lock-in.
A. Khansari@akhansari.tech

Is it just me, or ever since the pnpm devs started vibe coding, every new release comes with pretty annoying basic bugs?

With so many new LLM development workflows popping up and none truly proven yet, it has created complete chaos, making team progress difficult and software quality uncertain and unpredictable.

I've always found it ironic that many open-source projects are hosted on GitHub, which isn't open-source itself. And after all this time, no maintainer has bothered to change that.

I used to think LLMs would kill JavaScript, but I never imagined it would end up becoming their storefront (via Bun). Humans are shaping LLMs after all...

Oh there is now an official page for this: opensource.org/ai

opensource.orgOpen Source AITHE OPEN SOURCE AI DEFINITION 1.0 We have released the first stable version of the Definition. Read version 1.0 What’s Open Source AI? Following the same idea behind Open Source Software,…
A. Khansari@akhansari.tech

Open-Weight != Open-Source Open-Model means nothing. Open-Weight could be proprietary and having license. Open-Weight > Closed-Weight because you can own your data and infra.

I'm still new to this, but I find this insane that I can easily cross-compile #Rust from Linux to Windows, but not to macOS. And yet, people still claim macOS is the ultimate dev-friendly OS.

People give an LLM a couple of basic prompts to generate a generic task or project (something already baked into the training data, like a game or any open-sourced app), and think they've unlocked magic...

The real power is offloading boilerplate, repetitive, and generic work to the LLM so we can focus on novel problems and real added value.

Token sellers (OpenAI, Anthropic, etc.) will become more and more like hyperscalers (AWS, Google, MS). Just like solution architects, they will have an army of tech people attached to the sales department who will implement AI in companies and build customer loyalty to drive token consumption.

To escape market volatility, their only path is offering turn-key tooling, and leveraging higher-quality services to create vendor lock-in.

A. Khansari reposted

The impact of AI across the entire economy is nontrivial. Look at the just-announced $3,000 price increase on a fully loaded MacBook Pro. That's entirely an AI tax triggered by data-center-induced chip scarcity. 1/6