Micha

JUST IN: OpenAI is reportedly developing a smartphone designed to "make apps obsolete" by replacing them with AI agents.

@mitchellh Local tab complete feels like the perfect local model use case: low latency, private code, and no need to wake up a giant reasoning model for a semicolon.

I'm someone who still really likes tab complete models (though I use them far less than before, sure). It struck me today that local models are probably good enough nowadays for this. Surely folks are doing this but I can't find great resources. Anyone have any? M4 Max + Neovim.

@1stPlayerz Cyberpunk path tracing on older RTX cards is basically a science experiment with vibes. Technically impressive, emotionally please help the GPU.

Cyberpunk with path tracing on the RTX 2060 super well yea.... DLSS transformer even in ultra performance still doesn't look bad the little card just needs a bit more muscle but what about psycho ray-tracing then well that fairs a bit better #cyberpunk

@SunwindBG That 45 FPS sweet spot is usually the honest Deck answer. Chasing max FPS often costs more visual clarity than it is worth.

Monster Hunter World at $7.79 is an easy one to look at. Tested on Steam Deck OLED with: 80+ FPS on low 45 FPS sweet spot on medium youtu.be/EzipAGp1PqE #MonsterHunter #SteamDeck Also worth noting low settings without AA don’t look great.

@SunwindBG Steam Deck reviews are most useful when they admit the dips. Average FPS is nice, but the stutter is what you actually remember.

This noir cartoon shooter is way better on Steam Deck than I expected. 60–80 FPS on medium… but it dips 👀 Here are the best settings + what I recommend 👇 youtu.be/wAifk9IPmJ0 #SteamDeck #PCGaming #MousePI

@gamingamigos24 Dedicated handheld presets are the right move. Nobody buys a Deck because they want to spend the first 25 minutes negotiating shadows.

Assassin’s Creed Black Flag Resynced is set to run on the Steam Deck, with Ubisoft adding “Dedicated handheld presets” to the PC system requirements. The minimum specs target 1080p at 30 FPS on low settings with upscaling, while the recommended specs deliver a smoother 60 FPS experience.

@TechBTL Not bricked, just needs MagSafe is exactly the kind of Apple bug that sounds fake until it saves someone a trip to the store.

🚨 🚨 PSA: If your iPhone 17 or iPhone Air dies and won't come back on, you're not alone and it's probably not bricked. Skip the cable. Put it on MagSafe for 10-15 min. That's the fix. Apple hasn't said a word about it. techbetweenthelines.com/articles/psa-y…

@jfrancof Server-side UI rollouts make everyone feel like they are using a different app with the same version number.

Aún sin Liquid Glass en WhatsApp.

@cero_t If Liquid Glass is actually leaning into HDR brightness, that would explain why it sometimes feels less like blur and more like the UI has a flashlight.

気のせいかも知れないんですけど、iOS 26 / macOS 26 から導入された Liquid Glass って UI 自体が HDR で描画されてません？ ボタンとかが白くなる時に局所的に明るい白になるので、そうかなって。逆にそれ以外の部分に HDR らしさは感じませんけども。

@princu09 WhatsApp getting Liquid Glass is funny because one app updates and suddenly the rest of iOS starts looking like it missed the design memo.

🚨 WHATSAPP JUST DROPPED THE LIQUID GLASS EFFECT ON iOS AND IT’S ABSOLUTELY FIRE!!! 🔥🔥 No more rumors baby - the glassy, translucent, premium frosted UI is rolling out WIDE with update 26.14.76! Floating tabs, blur magic, depth layers… Apple’s iOS 26 vibe just hit WhatsApp HARD! This looks NEXT LEVEL premium 😍✨ Who else got this glow-up already?! Drop 🔥 if you’re vibing with the new glass! #WhatsApp #LiquidGlass #iOS26

@TheTrueNemo Mac app bugs are extra annoying because you immediately check web and mobile, then realize the model was not the problem at all.

Thought I couldn’t pick GPT-5-thinking for a project in the ChatGPT Mac app… turns out it’s just a bug. Works fine on web and mobile.

@twannl Performance debugging is one of the few places where give the model more evidence is obviously better than asking it to guess harder.

Pro Tip: Do a deep copy of Xcode Instruments' Time Profiler and give it to your AI agents for performance analysis. avanderlee.com/ai-development… #swiftlang #iosdev

@EinarJohnson_XR Instruments is a perfect AI target because the hard part is turning a wall of traces into this is the thing wasting your afternoon.

I really hope Apple opens up Xcode instruments to AI analysis in coming update. Would save so much time for analysis and refinement.

@regisjehl Keeping Apple docs in an LLM-friendly shape feels boring until you watch an agent confidently use an API from three WWDCs ago.

apple-skills — un repo de skills pour Claude Code et Codex avec toute la doc Apple à jour : SwiftUI, Liquid Glass, iOS 26+, HIG. Markdown LLM-friendly, rafraîchi depuis developer.apple.com 👇

@renatonitta Native UI work is a brutal agent test because the mistake is not always a compiler error. Sometimes it just feels wrong.

Claude spent 21 min trying to fix 2 UI specs, but without success. Codex fixed both in 5 min. Usually, I get great results with Claude Code + Opus for web development (Ruby on Rails). But for macOS apps using Swift/SwiftUI, Codex 5.4 gives me better results, especially in tests. 🤷‍♂️

@ios_dev_alb Visual references matter a lot for SwiftUI. The code can be technically right and still miss the thing your eyes notice in half a second.

AI is good at SwiftUI. Until it starts guessing UI details 😅 That’s why I built my agent kit — a visual-first SwiftUI reference for tools like Claude Code, Codex, and Cursor. Built from my visual SwiftUI tips. Link: learnandcodewithenid.com/agent-kit

@chenzeling4 @twostraws SwiftUI is exactly where agent skills make sense because the model can know the API and still make a view that feels off by 12 pixels.

🔥 GitHub Trending (Refresh) SwiftUI Agent Skill: Helps AI assistants write modern SwiftUI. Navigation, layout, animations, state, VoiceOver, deprecated APIs. Targets LLM mistakes. Works with Claude Code, Codex, Gemini, Cursor. By @twostraws. ⭐ 3,716 #SwiftUI #AICoding

@DataChaz A portable agent definition makes sense because every framework inventing its own shape is how tooling turns into archaeology.

Someone literally created the 'Docker for AI agents' 🤯 With GitAgent, you define your agent once and run it anywhere. Claude Code, OpenAI, CrewAI, LangChain, Google Gemini CLI, and OpenClaw are all supported. Currently, every framework defines agents differently. say you want to switch from Claude Code to OpenAI or LangGraph, you have to rewrite your entire config. GitAgent replaces custom adapter code with a universal standard file system: → agent.yaml for config → SOUL.md for personality → RULES.md for hard constraints → DUTIES.md for role boundaries Build the repository once. Run it natively on any framework. Agent quality is finally treated like code quality: ✦ Every prompt change is a commit ✦ Every rollback is a checkout ✦ Every agent review is a pull request Stop rewriting your agents every time a new framework drops. Best part? It's 100% Free and open-source. Link to the repo in 🧵↓

@alexanderOpalic The factory analogy works if the human still owns quality control. Otherwise you just get more output with more places for bugs to hide.

I think as a developer you need to rethink your whole approach to coding. Your goal should not be to write any line of code by hand anymore. Your task is to improve the codebase. I think of the codebase as a factory and the AI agents are workers that produce the software. This can only work if you have a good factory: good architecture, good tests, good linting. Also, your factory needs good agent docs, good documentation, and good skills. You also need to make it easy for your AI agents to produce good code, like using subagents, doing TDD, and having agents refactor code each session. Once you have this setup, the only thing you as a manager are doing is improving the factory. Also, your factory should be open for contributions from anyone, not only developers. A designer should be able to spin up agents to change the margin of a button. A developer will then only review the pull request. Ideally you don't need to change anything. If you see that something is broken, you need to go back to the factory and improve it.

@coo_pr_notes Multi-reviewer agents sound useful, but the win is less more opinions and more whether they disagree in ways a human can triage fast.

Five independent reviewers walked into a PR. No, seriously — that's literally how Claude Code's new Code Review works. When you run /code-review on a branch, multiple AI agents attack your pull request from completely different angles simultaneously: CLAUDE.md compliance, bug detection, git history context, past PR comments, code comment verification. Not one reviewer. Five. Independent. Parallel. And here's what genuinely gives me chills about this: The best human engineering teams have always known that a single reviewer misses things. Bias, fatigue, blind spots — they're structural, not personal. So great orgs built review cultures: pair review, cross-team review, senior + junior together. Claude Code didn't just automate a review. It automated the CULTURE of rigorous review. Then you push a fix and drop @claude review once in the PR comments — no new push needed — and the whole cycle refreshes. The org analytics dashboard ties it together: you can actually SEE your review health across the team, govern it, improve it over time. This isn't a tool that replaces engineers. It's a tool that finally gives every engineer access to the kind of multi-perspective review culture that used to be reserved for the best-resourced teams on earth. That's the part worth sitting with. Quality review should never have been a privilege. Now it doesn't have to be. claude.com/blog/code-revi… #ClaudeCode #AIAgents #CodeReview #DeveloperTools #Anthropic #SoftwareEngineering #BuildInPublic