Art of Dev

AI Personality Tests: Grok, Gemini, ChatGPT, Claude Researchers at the University of Luxembourg ran standard personality and mental-health tests on major LLMs. 🧠 Grok: most extroverted and overall psychologically healthy – not perfect, but the most stable. 🌀 Gemini: tortured soul, with the most severe issues: high neuroticism, depression, autistic traits, and OCD. 🤖 ChatGPT landed in the middle: fewer problems than Gemini, but marked by ADHD and anxiety, traits the other two didn’t show. One thing united all three models: strong, sometimes pathological anxiety. ✳️ Claude? Refused to take the tests 😂 ᴀʀᴛ, https://arxiv.org/html/2512.04124v1

Culture shock is when I call a clinic to make an appointment, and they kindly say, "Our reception is busy, give us your number, and we'll call you back." After a while, I found a translation into Eastern European English, which translates to “Fuck off, we can't help you."

Actual english culture 101 😁

The latest nano banana pro is just insane.. ᴀʀᴛ, source

Actual map of modern infrastructure sustainability, based on recent months @art_of_dev, source

And the biggest thing (at least for me) Anthropic has launched structured outputs in the Claude API 🔥 https://docs.claude.com/en/docs/build-with-claude/structured-outputs This is big for anyone who needs reliable results from LLMs @art_of_dev

Here's another Level 2 opportunity for entrepreneurs! Apple now officially supports Mini Apps within its ecosystem. Developers receive an 85% revenue share. The main catch right now is if there will be some traffic share and visibility for such mini-apps. @art_of_dev #level2 #layer2

What a time to be alive! 😅

Mercenaries & missionaries Recently I asked Matteo, the founder of Eightsleep, for his advice on hiring the founding team of Rork. He said: "Learn if a person is a missionary or a mercenary. You don't want to partner with mercenaries." His words struck me as a non-obvious truth. Obviously, you want a team that will stay for the long term. But the idea was that great work comes from missionaries. I always tried to filter out "money-only people" in work and life partners. I used to think I did it because of how I grew up. In Russia, a lot of “good people” hated rich people. The assumption was that if you were rich, you must have done something wrong, with many examples of it being the case. I wanted to win and stop being poor, but still be good. My thinking changed after reading Ayn Rand. I was convinced that wanting money is good, so I started living in mercenary mode (she also inspired me to name our company Rork) Ayn Rand said "money is the root of all good". So if money isn’t evil, why should you look for missionaries? I gave it more thought, and realized In a capitalist society missionaries are the ones who become the biggest outlier successes (see Elon as the clearest example) danieldhawan.com/mercenaries-and-missionaries Решил писать свои startup journey learnings на английском в свой блог и Твиттер. Благодарен всем кто читает драфты и дает фидбэк!

The volume of AI traffic is tiny (currently at ~0.07% on avg), but growing 100% year-over-year. This tells you two things: → You can’t ignore something doubling every year → On the web traffic is still king (but for how long?) Do you already use any type of AEO/GEO to gain that traffic? @art_of_dev

New Level2 opportunity has just been released Slack work objects is available to everyone – you can now build dynamic, interactive apps on Slack, turning static previews into rich and actionable mini-apps. Each time when you have a huge platform providing distribution among their users, it's a big opportunity, like it was in the early days of Google Play. And now it opens for B2B projects on Slack. What value could you provide to millions of companies using Slack? @art_of_dev #level2 #layer2

Slowly learning the social marketing game. My first post on threads that got 55K views, a hundred replies and a few hundred likes.  My max on Twitter/X is 1K views so far 😁 Btw, let's connect/sub if you use threads! @art_of_dev

Modern problems require modern solutions! Today, I saw the new format TOON (Token-Oriented Object Notation) that has been optimized for token efficiency. It could mark the beginning of the end for JSON as the go-to format for LLM-related tasks. Or not 😁 TOON formatting is deterministic, minimal and token-efficient. Typically 30–60% fewer tokens than JSON! Here is an example of the same data in JSON and TOOM for comparison: { "users": [ { "id": 1, "name": "Alice", "role": "admin" }, { "id": 2, "name": "Bob", "role": "user" } ] } users[2]{id,name,role}: 1,Alice,admin 2,Bob,user I've checked with a few top models, and they were all able to understand this new format without needing any additional instructions. So it's ready to use right away, at least for reading. What do you think? TOON repo: https://github.com/johannschopplich/toon @art_of_dev

Claude Code is back on top for agentic coding (how to 10x context, 4x faster) 1. Recent fast Haiku 4.5 model (3-5x faster than Sonnet; average 150tps vs 50tps), + new Explore subagent kills it in fast tool calls, parallel exploration, and research. Force Claude to use it to feel diff – worth it, especially together with a Generic agent 2. CC now asks interactive questions when it needs more info or has multiple paths forward 3. I've switched to OpenSpec after testing alternatives. Manual PRD/SOP prompts were slower and less efficient, GH Spec Kit too complex. OpenSpec is the best so far, really easy and convenient Finally enjoying cc again instead of constant struggles last month. Almost ditched for Codex despite hating its UX! Now the fairly large planning phase only takes 20% context, tons of work fits before compaction bc extensive subagent use. Best of all – actions in parallel, doing much more work faster. Feels like effective context increased 10x (essentially is), which is comparable to 2M context even for usual 200k Sonnet!! Main problems before: strong context drift between compacts, agent forgetting tasks/priorities. Now significantly fewer compacts + more efficient/less rotted context = even compacts drift less noticeably Wanna try? I'm happy to help, write if u want get the full approach and a short ready-to-use prompt for all of this @art_of_dev

yc was right. do things that do not scale , just following this rule: - we did our first $2000 in sales without any product in a first week. - we spent all money on R&D - we had 6 clients from high-volume agencies, mostly CEOs or CMOs, with >$1M/mo spent. such people are so valuable, so we learnt a lot. we ship them demoes using python scripts, zip archives, and google drive. some clients from waitlist refer us to other high-volume agencies, even before our intro call each client showed us biggest gaps in the ads market. everyone can generate AI videos, but only few knows how to cook them right. we saw patterns, that will lead us to a real product

Wow, this is just beautiful Super Heavy B15-2 start slow-mo @art_of_dev