Hey buddy,

Just wanted to give you a quick rundown on some of the wild stuff popping up on Hacker News today, Thursday, August 13, 2026. Grab a coffee, this is gonna be quick!

Gemini 3.7 Flash is Out

First up, Google dropped Gemini 3.7 Flash. People are saying it's super fast and way cheaper than the older 3.6 model for simple tasks. Someone in the comments even mentioned Deepseek v4 Flash is like, 5 times cheaper than 3.6, which is wild. Sounds like the AI model race is all about speed and cost right now.

DeepSeek Harness Dev Preview

Speaking of DeepSeek, they also launched a developer preview for something called DeepSeek Harness. It sounds like a platform to get LLMs working with other tools. But some folks in the comments weren't too hyped, saying a "plugin system" isn't exactly groundbreaking in 2026. They're expecting more "agentic" stuff, you know, self-improving AI. Still, others see it as a sign of a "great convergence" in how we build AI apps.

Spaghettifying DRAM

This one's pretty technical and kinda wild – "Spaghettifying DRAM". It's a low-level hardware hack where someone figured out how to mess with DRAM memory controllers. One comment gave an "ELI5" explanation, saying it's like finding a secret way to shuffle around your food jars (memory) that your parents (the OS) don't expect. Sounds like you need deep kernel access for this, so not something your average app can do.

Accelerating GPT-5.6 Sol Ultrafast

Then there's Cerebras and OpenAI teaming up to make GPT-5.6 Sol Ultrafast even faster. People are already thinking about how this insane speed can be used for things like SRE (Site Reliability Engineering) tasks, where AI agents could quickly review logs and even propose fixes. The cost of these closed-source models is a big topic too – how much does it *really* cost to run them versus what they charge?

Codex in ChatGPT Desktop App for Linux

Big news for Linux users: the ChatGPT desktop app with Codex (that's their coding AI) is now in preview for Linux! Folks are happy about the in-app browser features for feedback. There's a lot of talk in the comments about whether AI is genuinely useful or just "advertising slop," but having a native app is definitely a step up for many.

Choose Boring Technology (2015)

An oldie but a goodie came up again: "Choose Boring Technology" from 2015. It's all about sticking to stable, proven tech instead of jumping on every new shiny thing. The comments are still buzzing about it, with most agreeing: use boring tech at work, save the experiments for your personal projects. Someone had a funny take, saying "push all your innovation tokens into agents" is a good move, OR just don't use agents at all and stick to boring tech to make stuff that actually works.

McDonald’s Loyalty Program Data

And finally, this one made me raise an eyebrow: a Wired journalist requested their data from the McDonald’s loyalty program and got a whopping 515-page dossier! People are debating if it's creepy that companies collect so much on you, or if it's just normal business practice, like a shoe store keeping receipts. But others are worried that companies use this data not just to give you better service, but to figure out how much they can hike prices or lower quality before you finally bail. Also, apparently, employees sometimes game the system to make service times look faster, which is a whole other thing.

Alright, that's the gist of it for today. Talk soon!