Welcome back! OpenAI finally launched hardware… sort of. Its new Codex Micro is a $230 button pad built with Work Louder that lets developers monitor Codex agent threads, accept or reject changes, and see task status through glowing keys.
It’s not the mysterious Jony Ive device everyone’s been waiting for, but it is a funny little signal of where coding agents are headed: less “chat window,” more cockpit.
Just Because AI CAN Do Something Doesn't Mean It SHOULD
I tried to use AI to replace a huge part of my workflow recently, and it was a big bust.
The goal: automate my YouTube thumbnail workflow. Thumbnails are a real bottleneck for me… the design, the iterations, the back-and-forth with my designer. So I figured if any part of my creative process was going to fall to AI, it would be this one.
At first, some of the AI-generated thumbnails looked genuinely impressive. But they weren't consistent. Getting to a usable one took a ton of iterations. And when I lined them up next to what my human designer actually produces, the AI versions just didn't perform as well. (Also, one of them gave me a truly heroic forehead. We don’t speak of it.)
It was a great reminder that just because AI CAN do something doesn't mean it SHOULD.
The people getting the biggest productivity gains aren't the ones trying to automate everything. They're the ones who know:
Which tasks should be fully automated (repetitive, high-volume, low-judgment)
Which should be augmented with AI (research, first drafts, ideation, cleanup)
Which still need to stay fully human (taste-driven decisions, relationship work, anything where "close enough" isn't good enough)
That framework isn't sexy. But it's the difference between people who look busy with AI and people who are actually shipping more because of it.
Which brings me to something I'm excited to share. I've partnered with Teachable for their AI Academy, and I'm hosting a webinar where I'll walk through the practical framework I use to decide what AI should own with real examples from my own workflows (including the ones that flopped, thumbnails very much included).
If you'd like to join, you can register right here.
Hope to see some of you there.
— Matt
P.S. If enough of you hit reply asking for it, I'll share the forehead thumbnail. Maybe…
Thinking Machines Releases Its First Open Model
Via TechCrunch
Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati, released Inkling, its first in-house AI model. This is the company’s first public model release after spending more than a year building mostly out of view.
What it is: Inkling has 975 billion total parameters and 41 billion active parameters (by comparison, frontier models like Anthropic's Claude Opus are estimated at up to 5 trillion total parameters). It was trained on 45 trillion tokens across text, images, audio, and video. It can reason natively across multiple modalities, though its outputs are currently limited to text, code, structured data, and styled artifacts.
The angle: Thinking Machines doesn’t claim Inkling is the strongest model overall. Instead, it’s positioning Inkling as a customizable foundation model that companies can fine-tune. Users can also adjust “thinking effort” up or down depending on whether they want better performance or lower cost and latency.
Early proof points: The company says Inkling can match Nvidia’s Nemotron 3 Ultra on one coding benchmark while using about a third as many tokens. It also published demos showing Inkling building a web app, creating styled artifacts, and iterating on a multiplayer game through a long feedback loop.
Why it matters: This is the first real test of Thinking Machines’ thesis: that the future of enterprise AI is not just bigger general-purpose chatbots, but models that organizations can adapt to their own knowledge, workflows, and constraints. The question now is whether Inkling can compete with the heavyweights—or whether its real strength is giving companies a model they can make their own.
Hack Reveals Suno Scraped Millions of Songs
A hacking incident reportedly revealed that AI music generator Suno scraped millions of songs and lyrics from platforms including YouTube Music, Deezer, Freesound, the International Music Score Library Project, and more
What leaked: According to 404 Media, materials shared by the hacker included Suno source code from 2023 and 2024, plus scraping instructions for pulling audio and lyrics from protected platforms. One file reportedly showed Suno had consumed more than 2 million YouTube Music clips.
Legal backdrop: Suno is already facing lawsuits over claims that it trained on copyrighted music. In one case brought by the Recording Industry Association of America, Suno admitted it trained on copyrighted materials, but argued that doing so is permitted under fair use. The new leak appears to support allegations that Suno didn’t just train on music from the open web, but intentionally ripped tracks from platforms with copyright protections.
The bigger picture: AI music is forcing the industry to confront the same question that keeps coming up across generative AI: when does training become theft? Suno can argue fair use in court, but if the hacked files are accurate, they make the optics much harder.
Run Local AI on Your Mac
Via Osaurus
Osaurus is a free, open-source native Swift app for Apple Silicon Macs that lets you run AI models fully on-device using Ollama, MLX, or LM Studio. It requires no account or subscription, works offline, and keeps inference running directly on Apple’s M-series chips.
How you can use it
Run local AI models without sending data to the cloud
Build autonomous agents with voice control and folder watchers
Run parallel AI jobs on-device
Optionally connect ChatGPT, Claude, or Gemini when needed
Pricing: Free
Find the Right Name for UI Elements
Via NameThatUI
NameThatUI is a visual reference dictionary for user interface elements across macOS AppKit, SwiftUI, and web UI patterns. It helps developers and designers identify the correct technical names, framework symbols, and HTML or API identifiers for the UI components they want to describe.
How you can use it
Find the correct name for macOS, SwiftUI, and web UI elements
Write more precise prompts for coding agents
Translate plain-language UI descriptions into technical terms
Browse visual examples of common interface components
Pricing: Free
Jobs, announcements, and big ideas
xAI open-sources Grok Build, its coding agent and terminal UI, inviting developers to extend it freely.
Perplexity launches SPACE, a sandbox platform that isolates AI agents performing tasks on your computer.
OpenAI backs a "reverse federalism" approach to building a national AI safety framework in the US.
Chinese startup Moonshot AI announces open-weight AI model Kimi K3, rivaling US models at lower price.
George Lucas embraces AI in filmmaking, telling audiences "there's nothing you can do about it."
What dropped in the world of AI this week? From Claude’s new browser to Spotify’s AI, I break it down ⤵
That’s a wrap! See you next week for more.
