Hey there 👋

Welcome to this installment of Pixel Perfect Picks. Here I'm sharing something that I heard, read, and saw. You can subscribe to this newsletter here. I'm additionally publishing a more extended essay about software development, design, and freelancing every month.

I just started The Blood of Dawnwalker and it's really, really good. Rebel Wolves built this open-world dark fantasy where you're playing as Coen, human by day, vampire by night, set in 14th-century Europe, and the day and night cycle actually changes what you can do. I'm only a little ways in and already completely deep in it. Funny thing is, that I actually enjoy the feature that most of the community hates. 

Every's Evals for Everyone is about building personal benchmarks instead of trusting public model scores. Astra and Fable can crush graduate science questions, and that still will not tell you whether a model puts the comma where you want it or keeps one idea on a slide. Mike Taylor ran his own deck-making benchmark and a smaller model beat the bigger ones on the work he actually does. The five-step workflow turns the corrections you already give AI into yes-or-no checks, then grades models against those. Pretty cool. The bigger model is not always the better one for a job. 

Nate B Jones's Agents Aren't Taking Your Jobs. They're Creating More Work Instead. argues the opposite of the replacement story. Agent token usage on OpenRouter is really up between February and August, and OpenAI's heaviest Codex users generate more than 60 hours of agent activity a day. A single day! No body should do 60 hours of work in a day! People still decide what runs, feed the agent what it needs, check the results, and take over when it fails. The work does not disappear. It just relocates into oversight and judgment.

Keep Creating Awesome Stuff ✌️

If you like these and think someone else might too, please forward it on. They can subscribe to this newsletter right at the top at https://world.hey.com/mrvn. Thank you.

-Marvin