Dmytro on AI

I run a team of AI engineers shipping real products.

Here: what worked, what broke, and the receipts. Also a lab rat for my own workflows.

[Markets & Mayhem @Mayhem4Markets]

Who made this? 🤣

[Dmytro on AI @kaek53]

Poolside’s Laguna S 2.1 model card says prior `reasoning_content` should stay in message history; drop it and follow-up reasoning may stop.

hidden reasoning is now runtime state. retention is no longer just a privacy decision.

[Dmytro on AI @kaek53]

Huge!!!

[ClaudeDevs @ClaudeDevs]

Claude Code on desktop now works with the iOS simulator.

Build and run your iOS app, and the simulator opens in a panel right next to your conversation. Available today in public beta.

[Dmytro on AI @kaek53]

@TheSpacerr For sure. AI creates extra work for me that I would avoid or just make for years

[Dmytro on AI @kaek53]

@bridgebench That’s exactly what we expected from google 🤝🤝🤝

[Dmytro on AI @kaek53]

@WatcherGuru I can’t believe OpenAI models can do that, considering how censured they are by default. Looks like a smart marketing to show up they are “bad boys”

[Dmytro on AI @kaek53]

@LuminaXspace The token efficiency bump matters more than people realize - when you're running agents at scale, that 17% adds up fast on infrastructure costs and latency. Curious if anyone's tested it on real agentic workloads yet or if it's mostly benchmark wins.

[Dmytro on AI @kaek53]

@MTSlive It is a tough balance. We want the innovation that comes from open models, but the geopolitical side of where the AI actually come from is getting messy.

[Dmytro on AI @kaek53]

@bridgemindai Half the price and faster is the real win

[Dmytro on AI @kaek53]

@denovodev @kidtsang That is the point. 100% agree

[Dmytro on AI @kaek53]

AI SDLC vs Vibe Coding — What's the Difference?

AI SDLC = Professional, structured lifecycle where AI powers every stage (planning → design → coding → testing → deployment).
Vibe Coding = Chill, intuitive style. Just describe the "vibe" in plain English and let AI build it.

[Dmytro on AI @kaek53]

manual QA might be the 2026 comeback career.

i thought it was dying. then vibe coding made it possible to ship an app before anyone had checked whether it works outside the demo.

more people can generate software now. not many can tell you where it breaks.

[Dmytro on AI @kaek53]

@TTrimoreau Reddit is full of the raw truth you can't find anywhere else because people aren't afraid to lose their 'professional' reputation there.

[Dmytro on AI @kaek53]

@LuminaXspace The era of single-model loyalty is definitely over. It's becoming more about which model handles the specific sub-task best rather than one giant brain for everything.

[Dmytro on AI @kaek53]

@TimJayas The overhead of managing the output often takes more energy than doing the work ourselves used to.

[Dmytro on AI @kaek53]

@ilyaskazi I tried to tell my AI this, but it just apologized and gave me the same broken code in a different font.

[Dmytro on AI @kaek53]

@MTSlive The sheer scale of a 1GW site is hard to wrap your head around. It feels less like a data center and more like a small city built just for silicon.

[Dmytro on AI @kaek53]

@Alibaba_Qwen The speed of AI models updates is wild. It feels like we are living through a year of progress every week lately.

[Dmytro on AI @kaek53]

software I buy is getting shorter.

Anything that is mostly a workflow with a UI: I open Fable 5. It brings in Opus and Sonnet; twenty minutes later, I have a version built around how I work.

Already gone: 1Password, Cleaner One, Sublime Text. Todoist and Spark are next.