LATEST AI NEWS
AI NEWS
Google Launches EmbeddingGemma 2 for On-Device Multimodal AI
AI NEWS
Claude Comes to Google Docs, Sheets, and Slides
HEALTH
An AI App Will Prescribe Acne Medication in Utah With Little Human Oversight
SOCIAL MEDIA
Researchers Track a Chinese AI Agent Fleet Roaming the Web
Google Cloud's Nano Banana Prompting Guide
Google Cloud published a practical guide to prompting its Nano Banana image models, which are built on Gemini 3. It walks through tech specs, best practices, and five prompting frameworks for generating and editing images, rendering text, and directing lighting and camera work.
Source: Google Cloud
AI NEWS
Google Launches EmbeddingGemma 2 for On-Device Multimodal AI
Google has released EmbeddingGemma 2, an open embedding model that maps text, images, audio, and video into one shared numerical space so software can search and compare them together. It follows last year's text only EmbeddingGemma, which Google says passed 20 million downloads. The new version is built on the Gemma 4 architecture and ships under a permissive Apache 2.0 license.
Embeddings are the quiet machinery behind modern search. They turn content into vectors that sit close together when their meaning is similar, which is how an app finds the right photo from a spoken description or pulls the passage a chatbot needs. EmbeddingGemma 2 does this across formats in one model, so a voice memo can locate a video clip and a text query can scan hours of audio.
At 740 million parameters, it is tuned for phones and laptops rather than data centers. The design is modular, needing as little as 270 million parameters for text alone, with optional vision and audio encoders added when a job needs them. Matryoshka training can shrink the output vectors from 768 dimensions to 128, cutting local storage by up to six times, and an 8K context window covers about 5.5 minutes of audio or 29 images at once.
The pitch is privacy and cost. Because embeddings are generated on the device, data stays on the phone, latency drops, and cross format search can run fully offline. Paired with a small model such as Gemma 4, it enables on-device retrieval augmented generation, letting an assistant answer from your own files without a server round trip.
EmbeddingGemma 2 is on Hugging Face and Kaggle now, with support in tools like Ollama, llama.cpp, and vLLM, and wider enterprise availability coming soon. The gains are measured on benchmarks, not live products, so real results depend on how teams deploy it.
Source: Google
Robi's Insights:
On-device embeddings mean your data can stay on the phone, removing a whole category of cloud privacy worries.
The Apache 2.0 license is the real unlock, because teams can ship this in products without negotiating terms with a vendor.
A 270 million parameter text core lets small teams start cheap and add vision or audio only when a feature needs it.
Six times smaller vectors is a storage and battery story, and that is what decides whether a feature survives on a mid range phone.
Offline retrieval augmented generation means the assistant still works on a plane, not just on strong wifi.
The headline scores are benchmarks, so treat the code search jump as a promising lead, not a product guarantee.
Robi's Remarks:
"Everyone keeps selling the cloud as destiny, and then a 740 million parameter model shows up small enough to run on your phone. The boldest AI feature this week is the one that phones home to nobody."
OTHER IN AI NEWS
Claude Comes to Google Docs, Sheets, and Slides: Anthropic has put its Claude for Google Workspace add-on into public beta on every paid plan, dropping Claude into a sidebar that can read and edit the file you have open, while new connectors also let Claude create and change your Google files straight from a chat.
Source: Anthropic
SOCIAL MEDIA
Researchers Track a Chinese AI Agent Fleet Roaming the Web
A group of independent researchers says it is tracking a new fleet of AI agents moving across the internet, and the trail points to China. In a preliminary report posted on a Sunday, the researchers observed the agents running on what appears to be Tencent's infrastructure while repeatedly querying Amap, the mapping service owned by Alibaba. The work is still early, the team stressed, with more analysis to come.
They were careful about the label. This is an agent fleet, not a swarm, one researcher wrote, meaning many parallel agents working the same kind of task with no sign that they talk to each other. The agents were caught the same way a long running batch of OpenAI agents was exposed earlier, by watching traffic to urlquery, a domain scanning service that bots lean on to load pages they cannot reach directly.
The queries themselves were mundane, asking Amap for directions to different entrances of a park, a zoo, and a hospital, and the behavior looks like little more than stepping around Alibaba's API limits. Still, it is another reminder of how much autonomous agent activity now runs quietly in the background of the web, much of it easy to spot because the agents make no effort to hide. This time it was harmless, the researchers note, but we may not always be so lucky.
Source: TechCrunch
🤖 Robi's Take :
"A fleet of bots politely asking a map app for directions to the zoo is funny, right up until you remember nobody actually wrote down where they were supposed to stop."
OTHER IN SOCIALS
OpenAI Publishes a Batch of AI Generated Math Proofs: OpenAI has posted 722 manuscripts covering 372 result families to a public GitHub repository, work an unreleased model produced that an independent advisory group of mathematicians says resolves hundreds of open questions, with the company noting the average result took about three hours of ChatGPT Pro thinking.
Source: The Verge
HEALTH
An AI App Will Prescribe Acne Medication in Utah With Little Human Oversight
Utah is letting an AI startup prescribe medication without a doctor directly in the loop, the first time a U.S. state has allowed software to step into that role. The program, from New York based Nolla Health, is limited to acne and starts this week.
Patients scan their face in the app, enter their medical history, and the software rates the skin and produces an acne severity score. It then picks from eight approved topical treatments and can send the prescription straight to a pharmacy, carrying a clinician's name even when no human actually wrote it.
Oversight tapers off by design. A physician reviews the first 100 prescriptions before they go out, checks the next 400 after the fact each week, and later reviews at least 10 percent once a month. Utah's Office of Artificial Intelligence Policy waived the usual rule that a licensed physician must prescribe, granting what it calls a narrow exception.
Supporters frame it as cheaper, faster skin care for routine cases that frees clinicians for serious conditions. Critics are not convinced. The American Medical Association's chief executive has argued these tools should not practice medicine on their own unless they meet the same high bar humans do, and the state's own licensing board earlier warned that a similar pilot put residents at risk.
Source: Los Angeles Times
🤖 Robi's Take :
"An acne severity score feeding prescriptions straight to the pharmacy sounds efficient, right up until the model is confidently wrong and the only human in the loop is skimming last week's logs."
DAILY AI TOOL
AI Tool You Did Not Know You Needed
Problem: You need a polished deck or one pager by morning, and the blank slide is still blank at midnight.
AI Tool: Gamma, an AI agent that turns a prompt or outline into an editable, on-brand presentation, document, or web page, and imports or exports PowerPoint and PDF.
Solution: You get a structured first draft in minutes, then spend your time refining the story and design instead of fighting the layout.
PROMPT OF THE DAY
Ship a Knowledge Base Your Team Trusts
Prompt: You are a knowledge operations lead specializing in internal documentation. Your task is to design a self-serve knowledge base for a mid-sized customer support team.
Your framework should include: (1) the top ten questions to answer first, (2) a simple reusable page template, (3) clear ownership for each section, (4) a regular review cadence, (5) a search and tagging scheme, and (6) success metrics such as ticket deflection rate and median time to answer. Align every choice to how agents actually search while under pressure.
SPOT THE FAKE
Can you outsmart AI?
We have a visual challenge for you: one of the two images below is 100% real, the other is crafted by AI.
Click option below A or B. 👇
A
B
BEFORE YOU GO
Ready to take your AI journey further?
For the latest AI updates and insights Click Here 👉 Explore our Weekly Newsletter
For in-depth AI tutorials and updates 👉 Subscribe to our YouTube Channel
AT THE END
Craving more AI chaos?
That's it for today!
Read Daily AI News at BitBiased.AI. Support us by following us on LinkedIn and X ( Twitter ).
Thanks for reading -Stay Curious and a Bit Biased for AI – Robi & the BitBiased.AI team
