🧠 Meta releases Mind Reading AI

Good morning! Here's what's happening in AI today:

Meta AI released the full training code for Brain2Qwerty v1 and v2, its non-invasive brain-to-text decoder, with research partner BCBL releasing the v1 dataset alongside it.

Meta is framing this as a step toward restoring communication for people with brain lesions or disorders that prevent speech. What stands out is the decision to open-source the actual code and dataset on the same day as the research reveal, which is rare for frontier neuroscience work and meaningfully speeds up how fast the rest of the field can build on it.

Cursor introduced a native iOS app that lets developers build from anywhere using always-on cloud agents, or remotely control agents already running on their computer directly from their phone.

Most coding agents have assumed you are sitting at a desktop. Cursor's iOS app changes that assumption directly, letting developers check in on long-running agent work or kick off new tasks from anywhere. For a tool already popular among professional developers, mobile access removes one of the last remaining points of friction.

Anthropic announced that Claude in Microsoft Foundry is now generally available, hosted on Azure, giving customers direct access to Claude Opus 4.8 and Claude Haiku 4.5.

A large share of enterprise AI spending runs through whichever cloud a company has already standardized on. By becoming generally available inside Microsoft Foundry, Claude removes a real adoption barrier for Azure-first organizations that previously had to set up a separate vendor relationship just to use it.

I ran the same prompts on 4.7 and 4.8 for two weeks. Same tasks, same context, same everything.

On 4.8, they came back padded. Longer than needed, hedged, the actual answer buried in the middle.

4.8 didn't regress. Anthropic trained it to be more careful, and for hard reasoning tasks that's great. For everything else, careful is the wrong default.

Every problem is fixable. Most fixes take under a minute. The last one solves it permanently.

Most prompts look like this:

“Write a summary of this article for my newsletter.”

4.8 has no idea who reads your newsletter, what tone you want, or how long the summary should be. So it picks the safest version of all of that.

Correct. Forgettable.

The fix:

“Write a 3-sentence summary of this article for my newsletter. Readers are professionals who use AI at work. Direct tone. Lead with the most surprising part, end with why it matters to them.”

Same task. Every gap is closed. The output shows it immediately.

When writing any 4.8 prompt, always include three things: who the output is for, what tone you want, and what format you want. Those three details alone cut the padding by half.

4.8 has default habits it falls back on when nothing overrides them. Padding with context you didn't ask for. Disclaimers. Summary sections. Bullet points by default.

Turn them all off in one line at the bottom of your prompt:

Rules: no disclaimers, no summary at the end, be direct. If anything is unclear, make a decision and tell me what you chose.

Here's the version I use for writing tasks:

Rules: no disclaimers or caveats, no "it depends" answers, no summary, no bullet points unless I ask. If unclear, make a call and note it at the bottom.

[Image]

That last part matters. Telling 4.8 to decide and note the assumption means you get an answer instead of a back-and-forth. Correct it in one message if it was wrong. Keep moving if it wasn't.

4.8 is not the best model for every task. Most people default to it because it's the newest. That's the mistake.

I tested Sonnet 4.6 and Opus 4.8 across every task I run regularly. Here's what I actually use:

Sonnet 4.6 is faster, follows tone instructions better, and is less cautious. For writing it beats 4.8 consistently.

Change the model before you start, not after you get a bad output.

Become a paying subscriber to get access to all premium content.

✓ Full archive of premium guides with ready-to-use prompts

✓ Structured AI courses (step-by-step, start-to-finish)

✓ Every upcoming premium tutorial

OpenClaw launched native iOS and Android apps, bringing agents into your pocket with channels, tasks, and replies you can manage on the go.

X Developers announced a hosted X MCP, giving agents like Grok and Cursor direct access to the X API with zero setup required.

Gamma launched inside the ChatGPT app, letting users generate presentations directly in any chat and continue editing them in Gamma.

🎬 Magnific: Now available as a plugin inside Premiere Pro, After Effects, DaVinci Resolve, and Final Cut Pro. Generate, upscale, relight, and remove backgrounds directly in your edit, with access to 250 million stock assets.

🚀 Base44: The first app-creation platform with its own trained model, built on tens of millions of real builder interactions so it understands the shape of a real app from the first prompt.

🎯 Giga AI Scout: Tell it the KPI you care about and Scout builds the agents, learns from every conversation, tests each change, and keeps improving that number on its own.

💻 Replit Desktop: A dedicated desktop app for Windows and Mac that lets you multitask and get the full Replit experience outside the browser.

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