Good Morning! Here's what I have for you in today's newsletter:
OpenAI launches ChatGPT Images 2.5, faster with comment-based edits
Gemini 3.5 Transcribe now uses your voice and screen context on macOS
Meta introduces Muse, a personal agent that works directly in WhatsApp
Turn Fable 5.1 into a researcher with Stanford's peer-reviewed method
4 new AI tools worth trying today
AI TOOLS
OpenAI launched ChatGPT Images 2.5, faster and sharper than before, with improved fidelity, consistent details across multiple edits, and comment-based edits that change only what's specified.
Comment-based edits let users point at a specific part of an image and describe only the change wanted there without regenerating the whole image for one adjustment.
Consistent details across multiple edits keep a character or object's appearance the same through a series of changes
Faster generation speed is called out specifically to support rapid iteration, generating multiple attempts without a long wait between each one.
Comment-based edits isolate a change to the specific part of an image being addressed, leaving the rest untouched. It's a small mechanic, but it's definitely better than trusting the tool with a finished piece and having to regenerate from scratch over one wrong detail.
AI AGENTS
Meta introduced Muse, a personal agent powered by Muse Spark 1.3 that works directly inside WhatsApp, messaged just like texting another person, currently available only in the US to those 18+.
Muse operates entirely inside WhatsApp's existing messaging interface, rather than requiring a separate app or platform to interact with.
The agent is built specifically to get tasks done on a person's behalf, distinct from a general-purpose chat assistant answering questions.
It can be set up once and then messaged like any other WhatsApp contact, with no separate onboarding flow to open first.
Muse Spark 1.3 already handles longer, multi-step tasks and asks clarifying questions before acting, so Muse inherits real capability, not just a chat wrapper around a basic model. Restricting it to US adults first is a cautious rollout for something that texts like a real person, the kind of feature worth watching closely before it reaches everyone else.
AI MODELS
Google brought Gemini 3.5 Transcribe to the Gemini app for macOS, using voice and screen context together to summarize local files, plan and research, and create images through natural language.
Screen context lets the model reference what's actually open on the desktop without relying only on typed or spoken input.
Local file summarization works directly on files already on the machine, without requiring them to be uploaded elsewhere first.
Voice and screen context combine into one input method, letting users speak a request while pointing at what's on screen.
Reading actual screen content lets users reference something hard to describe in words, like a specific chart or file layout. Voice input paired with what's visibly on screen is a more natural way to work than typing out a description of something already right in front of you.
HOW TO AI
Ask Fable 5.1 one question and you get one answer, a single lens. Stanford has a published method for this exact problem, STORM, peer-reviewed at NAACL 2024. No install, no repo, just five prompts run in order.
Step 1: Discover the right five perspectives for your topic
The real Stanford paper doesn't default to the same five personas every time. Practitioner, Academic, Skeptic, Economist, Historian is a strong general default, but a medical topic wants a Clinician, not an Economist.
I'm researching [your real topic]. Before generating any expert answers, tell me which five distinct perspectives would genuinely surface the most disagreement on this specific topic, and why the generic set might miss something here.Step 2: Generate five disagreeing experts
I'm researching [your real topic]. Answer as five experts who genuinely disagree with each other: [either the generic five, or the topic-specific five from Step 1]. Give each one 3 to 4 real sentences. Don't make them agree just to be polite.
Check honestly, do the five actually disagree, or did they just restate the same point five ways?
Step 3: Map where they contradict each other
Go through the five expert answers above and list every real point where two or more of them actually contradict each other. For each contradiction, name which experts disagree and what the actual disagreement is.Step 4: Synthesize a brief that keeps the disagreements
Write a research brief on this topic that draws from all five perspectives. Where they contradict each other, name the disagreement directly instead of picking a side or averaging them into a vague middle position. End with two or three genuinely unresolved questions.
Step 5: Make it peer-review its own work
Review the brief you just wrote as a skeptical peer reviewer. Find the three weakest claims, anything stated with more confidence than the evidence actually supports. Rewrite those specific parts.P.S. You can access all the AI trainings (including the full version of this one), prompts and workflows if you upgrade.
Inception AI introduced Mercury 2.5, the most capable diffusion LLM on the market, offering a 40% jump in intelligence over Mercury 2 and running over 1,100 tokens per second on widely-available NVIDIA GPUs.
Perplexity launched its Search API in Hermes Agent, giving it access to an index of more than 400 billion URLs with real-time results ranked by relevance.
Lovable introduced drafts, letting a team create parallel versions of a project to explore ideas freely and only apply changes once they're happy with the result.
🔐 Grok Bot: just added support for filling out forms and logins directly inside the chat, working with any password manager, no more switching to a browser mid-task.
🎬 Krea: just launched Realtime Director, letting anyone direct high-quality video generations live as they play, powered by fal's H3 Max.
🎙️ Gradium: just launched Voice Design, generating a brand-new synthetic voice from a text description in seconds, free on every plan including the free tier.
🎮 Buzzy: just added its 3D Director Console, powered by GPT-6 Astra, turning a written prompt into a fully interactive 3D world and then into finished video.
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