Good morning! Here's what's happening in AI today:
xAI launches Grok Bot, AI teammates that sign into your tools and do real work
OpenAI brings the ChatGPT desktop app to Linux with Projects and Codex
Unsloth releases the first desktop app to run and train models locally
How to train Claude to write in your actual voice
4 new AI tools worth trying today
AI MODEL
xAI unveiled Grok Bot in early beta, framing it as an AI teammate that signs into a person's existing tools, uses them the same way a human would, and comes back with finished work rather than just a chat response.
Grok Bot signs into the tools someone already uses instead of requiring a new integration or API setup for every task, lowering the barrier to actually deploying it.
The framing positions Grok Bot as a teammate rather than an assistant, meaning it's meant to complete a task independently and return with a result, not wait for step-by-step direction.
Early beta access means functionality and tool support will likely expand over time, with the current release serving as a first look rather than the finished product.
Most AI agents so far have needed custom integrations or specific APIs to interact with a person's actual tools. Grok Bot's pitch is signing in and working the same way a human employee would, which if it holds up removes a major setup barrier to real agentic work. The "hired a teammate" framing signals xAI wants this positioned as genuine delegation, not just automation. Early beta means expectations should stay tempered, but the direction points toward agents doing complete tasks with far less setup friction.
AI TOOL
OpenAI released a preview of the ChatGPT desktop app for Linux, giving supported systems access to ChatGPT, ChatGPT Work, and Codex alongside existing projects and browser workflows in one native application.
The Linux preview brings feature parity with ChatGPT Work and Codex, rather than launching as a stripped-down or limited version for the platform.
Existing projects and browser workflows carry over into the desktop app, so users don't lose context moving from browser-based ChatGPT to the native Linux client.
This closes a long-standing gap for Linux users, who previously had to rely on the browser version while macOS and Windows had dedicated desktop apps.
Linux has historically been the last platform to get first-class treatment from major AI companies, so a full-featured preview rather than a limited build is a meaningful signal. Developers and technical users, who make up a disproportionate share of the Linux user base, get direct access to Codex and Work features without browser overhead. It's a smaller-audience release than a typical consumer launch, but it closes a real functionality gap for a group that skews heavily toward the exact workflows ChatGPT Work and Codex are built for.
AI MODEL
Unsloth AI introduced Unsloth Desktop, the first open-source desktop app to run and train AI models locally, supporting Mac, Windows, and Linux with connections to Claude Code and Codex for working directly with local LLMs.
The app supports a wide range of local formats and modalities, including MLX, diffusion image and video models, audio, and GGUF, covering far more than typical local-inference tools.
Training runs twice as fast with 70% less VRAM required, making it practical to fine-tune models on consumer hardware that couldn't previously handle the workload.
Unsloth Desktop works across CPU and multi-GPU setups spanning NVIDIA, AMD, Intel, and Mac, avoiding the vendor lock-in that limits many local AI tools to a single hardware ecosystem.
Running and training models locally has typically required stitching together multiple separate tools, each with its own hardware quirks and format support. A single open-source desktop app that handles both, across every major OS and GPU vendor, removes a lot of that friction for developers who want to keep models and data off the cloud. The VRAM reduction and speed gains matter concretely too, since they put fine-tuning within reach of hardware that couldn't handle it before. For anyone serious about local AI work, this consolidates a fragmented setup into one app.
HOW TO AI
Anthropic confirmed this week that Claude now weaves an invisible watermark into everything it writes. Not a banner, not a disclaimer, a signal built into the text itself, there to comply with EU rules that took effect this month, applied everywhere Claude is used, not just in Europe.
That's not really the point of this guide, though. The mark doesn't prove Claude wrote something, and it doesn't prove a human did either. What it actually does is sharpen an old question: is this writing actually yours, in substance and voice, or is it just text a model produced that happens to have your name under it.
Most people who use Claude to write never get past "make it sound more human." That's the wrong target. The real fix is teaching Claude what your voice specifically is, tested against real evidence, not vibes.
1. Meet the method
"Write like me, casual but professional" doesn't work, not because the instruction is too short, but because everyone's version of casual but professional collapses to the same output. Voice isn't a tone adjective. It's a set of measurable patterns: the specific words you reach for, how long your sentences run, how you open and close a piece, where you hedge and where you just state things.
2. Build your voice profile
Aim for 100 or more real samples, messages, emails, posts, whatever exists in volume. Paste a large batch into Claude and ask it to extract patterns, not adjectives:
I'm going to paste a large sample of my own writing, unedited, real
messages, posts, or emails I actually wrote. Read all of it and
extract:
- Words and phrases I reach for often, and ones I clearly avoid
- My sentence rhythm: short and choppy, long and layered, or a mix
- How I open a message or post, and how I close one
- Where I hedge versus where I state things flatly
- What I do with humor, sarcasm, or emphasis, if anything
- Formatting habits: punctuation choices, capitalization, list use
Don't summarize me in adjectives like "casual" or "direct." Show me
the actual patterns, with quotes from what I pasted as evidence.
[paste your writing samples]What comes back should read like a linguist's notes, not a personality quiz result.
3. Test it blind
Reading a draft and thinking "yep, sounds like me" is close to worthless as a test, since you already know what you meant to say. The real test: give Claude a real situation you've already responded to, without showing it your actual answer, and have it draft a response using only what it's learned. Then compare.
Read the full guide.
P.S. You can access all my prompts, workflows and AI trainings if you upgrade here
OpenAI Developers added the ability to keep work from other agents in sync with ChatGPT Work and Codex, letting users import projects, chats, skills, and plugins, review import history, and opt into automatic updates from Settings.
LTX.io introduced LTX-2.5, a major upgrade to its world model already powering film, robotics, and real-time workflows, adding higher pixel fidelity, multishot scenes that hold together across cuts, and a new Diffusion Fidelity Rendering approach built for cinema-level frame quality.
Linear launched live cards in Slack for Linear Agent's coding sessions, letting teams follow along as the agent investigates an issue and makes changes, then open the diff directly from Slack once it's ready for review.
🖼️ Higgsfield Layers: an image editor with layer decomposition, flawless text rendering, and native 4K, split any image into editable layers. (URL needed)
🎙️ Soniox TTS v2: frontier text-to-speech with expressive control, voice cloning, and 60+ languages, at $0.70 per generated hour.
🎬 Wan CLI: call Wan directly from your agents to generate images and video programmatically, credits sync across the CLI and your account. (URL needed)
🔍 Firecrawl: a keyless web search option for opencode, live results with no API key or signup required.
THAT’S IT FOR TODAY
Thanks for making it to the end! I put my heart into every email I send, I hope you are enjoying it. Let me know your thoughts so I can make the next one even better!
See you tomorrow :)
- Dr. Alvaro Cintas
✓ Full archive of premium guides with ready-to-use prompts
✓ Structured AI courses (step-by-step, start-to-finish)
✓ Every upcoming premium tutorial