Latest Update
Why Google’s Gemma 4 12B Matters
2026.06.04Hey — it’s the Carat team. 🤗
Today, we’re unpacking the week’s AI news around Gemma 4 12B, the new AI model Google just opened up.
Why Google’s Gemma 4 12B Matters
Google just released Gemma 4 12B. Gemma is Google’s family of AI models that outside developers can download and use. Here, 12B means roughly 12 billion parameters. Parameters are close to the decision rules an AI builds up during training: the higher the number, the more patterns it can hold, but the more expensive it gets to run.
What makes this model worth watching is that it targets a size that can run on a 16GB laptop while supporting image understanding, audio processing, and agent-style reasoning. The bigger story: AI is moving beyond large cloud services and toward models people and companies can run on personal devices or inside private environments.
📌 3 AI stories today
1️⃣ Runway introduces Aleph 2.0 for video editing
Runway opened up Aleph 2.0 through its API this week, after first introducing the model on May 21. Aleph 2.0 is an upgraded version of Runway’s main video-editing model, released alongside its new Edit Studio.
With Aleph 2.0, you can provide a reference frame showing the change you want, and the model applies that change across the relevant parts of the video. Runway also says it can carry edits through videos with multiple cuts, not just a single isolated shot.
2️⃣ Amazon Search can make product images that don’t exist yet?
Amazon is testing a feature that shows AI-generated product images inside search results. According to The Verge, when users search for a style or vibe, Amazon may show not only real product photos but also AI images shaped around that query.
Even if the exact product does not exist yet, Amazon can show the look the user has in mind first, then guide them toward similar products. Shopping search is shifting from “find an existing item” to “visualize the scene you want first.”
3️⃣ Google wants to catch AI voice scam calls
Google is adding fake-call detection to the Android Phone app. The feature is designed to reduce AI voice scams that impersonate family members, workplaces, or institutions.
According to TechCrunch, the feature analyzes call conversations on the device and shows a warning when it detects scam-like requests, such as bank transfers or gift-card purchases. Google says the call content is processed on-device and is not sent to its servers.
🧪 Today’s prompt tip
A prompt shared by X creator wish4ku has been gaining traction, with around 13,000 bookmarks.
It lets you create a selfie that looks like you casually snapped a photo with your favorite person while having drinks at a cafe.
Here’s a result we made on Carat using a Yoo Jae-suk photo.
Here’s the prompt. Add your photo and a photo of your favorite person to make your own version.
[Core rules]
• First image = you / second image = the other person
• Preserve both people’s face identity 100%
• No face blending / averaging / AI reinterpretation
• Keep them looking like the exact same real people
• Do not change: face shape, eye size, eye shape, eyelids, nose, lips, eyebrow shape, face proportions, asymmetry, spacing between facial features, skin texture, hairline, hairstyle, overall impression
• Strictly forbidden: face beautification, AI idol makeover, stylization, jawline correction, enlarged eyes, skin smoothing, added makeup, beauty filters, glossy skin, oversharpening, influencer look, transformation into a Korean model type, age change
[Situation]
A natural Korean couple iPhone front-camera selfie taken late at night in a quiet cafe.
They feel close but a little awkward, without forced posing.
An everyday snapshot that feels accidentally captured, like an Instagram Story.
[Composition]
• Landscape 4:3
• iPhone fixed on the table for a selfie
• Low angle + slightly tilted frame
• Awkward crop / imperfect off-center composition
• Two large drinks in the foreground
• The cups naturally cover small parts of the faces, but both faces remain identifiable
• Hands and cup positions overlap in a realistically awkward way
[Expression]
• No forced smiles
• Natural, unpolished expressions
• Imperfect real-couple selfie mood
[Lighting & texture]
warm dim cafe lighting iPhone front camera feeling
• Dark yellow cafe lighting
• No flash
• Slightly underexposed
• subtle sensor noise
• mild grain
• slight softness
• slightly blurry focus
• compressed social media quality
• low-light phone camera texture
• Natural low-quality phone-camera look
• No HDR / cinematic look / DSLR look / studio lighting
• No editorial-photo feel / ultra-high definition / heavy retouching
[Core priority]
Face identity comes first.
Preserving the original faces matters more than the mood.
They must look like the exact same real people.
[Overall mood]
- 2000s digital-camera vibe
- Rough iPhone selfie feel
- Realistic Korean couple photo
- Natural, lived-in mood
- Looks like a real person took it on a phone
- Feels like a random saved gallery photo
- Ready to post as an SNS Story
Try making it on Carat. 👉 Create it on Carat
That’s it for today. We’ll be back tomorrow with more AI news. ☺️
Was this helpful? Share with your friends
Share
Get daily AI trend updates
