Google's ultra-efficient Gemma 3

PLUS: Sakana's peer-reviewed AI-authored paper

Author

Rowan Cheung
March 13, 2025

Good morning, AI enthusiasts. The era of massive compute requirements for cutting-edge AI may be coming to an end, with Google’s latest open-source release outperforming giants at just a fraction of the size.

With Gemma 3’s high-level performance, multimodal capabilities, and on-device operation on just a single GPU, the AI efficiency barrier is quickly getting destroyed.

In today’s AI rundown:

  • Google’s Gemma 3 for single-GPU deployment

  • Gemini Flash with new image capabilities

  • Create your AI-powered Telegram assistant

  • Sakana’s peer-reviewed AI-authored paper

  • 4 new AI tools & 4 job opportunities

LATEST DEVELOPMENTS

GOOGLE

Image source: Google

The Rundown: Google just unveiled Gemma 3, a new family of lightweight AI models built from the same technology as Gemini 2.0 — delivering performance that rivals much larger models while running efficiently on just a single GPU or TPU.

The details:

  • The model family comes in four sizes (1B, 4B, 12B, and 27B parameters) optimized for different hardware configurations from phones to laptops.

  • The 27B model outperforms larger competitors like Llama-405B, DeepSeek-V3, and o3-mini in human preference evaluations on the LMArena leaderboard.

  • Other new capabilities include a 128K token context window, support for 140 languages, and multimodal abilities to analyze images, text, and short videos.

  • Google also released ShieldGemma 2, a 4B parameter image safety checker that can filter explicit content — with easy integration into visual applications.

Why it matters: Gemma 3’s performance is mind-blowing, beating out top-level systems that dwarf it in both size and compute. Running on just a single GPU, these models hit a once incomprehensible sweet spot of being open-source, powerful, fast, multimodal, and small enough to be deployed across devices — a truly massive feat.

TOGETHER WITH DAGSTER

The Rundown: Dagster consolidates your AI capabilities into one powerful orchestrator that developers love — helping reduce costs, eliminate complexity, and ensure reliable pipelines from prototype to production.

With Dagster, you can:

  • Consolidate all AI capabilities under one intuitive interface

  • Save 40%+ on infrastructure costs by optimizing AI workloads

  • Ship AI features 3x faster with standardized development practices

GOOGLE

Image source: Google

The Rundown: Google released new experimental image-generation capabilities for its Gemini 2.0 Flash model, letting users upload, create, and edit images directly from the language model without requiring a separate image-generation system.

The details:

  • A 2.0-flash-exp model is available via API and in the Google AI Studio with support for both image and text outputs and editing via text conversation.

  • Gemini uses reasoning and a multimodal foundation to maintain character consistency and understand real-world concepts throughout a conversation.

  • For instance, you can prompt it to generate a story with pictures and then guide it to the perfect version through natural dialogue.

  • Google says Flash 2.0 also excels at text rendering compared to competitors, allowing for ads, social posts, and other text-heavy design generations.

Why it matters: This upgrade is a major step in shifting how AI generates visual content — moving away from dedicated image models toward language models that natively understand both text and visuals. Just as natural language prompting has taken over other domains, image editing appears to be next on the list.

AI TRAINING

The Rundown: In this tutorial, you will learn how to build a personal AI assistant on Telegram that can answer questions, remember conversations, and eventually connect to other services using n8n's automation platform.

Step-by-step:

  1. Create a Telegram bot - search for "BotFather" in Telegram, type /newbot, and save the API token you receive.

  2. Sign up for n8n (free 14-day trial) and create a new workflow with a Telegram trigger using your bot token.

  3. Add an AI Agent node after the trigger, connect it to your preferred AI model (like those from OpenAI), and use the message text as the prompt.

  4. Add a Telegram "Send Message" node to return the AI's responses to your chat using the chat ID from the trigger.

  5. Enable Window Buffer Memory in the AI Agent settings so your bot remembers previous conversations.

Pro tip: You can also expand the assistant’s capabilities by connecting it to calendars, emails, notes apps, and other services. We did an extensive workshop on how to create your own AI Agent to automate tasks with n8n here.

PRESENTED BY JOTFORM

The Rundown: Jotform AI Agents let organizations provide 24/7, conversational customer service across multiple platforms — no coding required.

With Jotform AI Agents, you can:

  • Get started easily with over 7,000 ready-to-use AI agent templates

  • Automate workflows and trigger custom actions in real time

  • Handle voice, text, and chat inquiries seamlessly

  • Customize your agent’s look and feel to align with your brand identity

SAKANA AI

Image source: Sakana AI

The Rundown: Japanese AI startup Sakana announced that its AI system successfully generated a scientific paper that passed peer review, with the company calling it the first fully AI-authored paper to clear the scientific bar.

The details:

  • AI Scientist-v2 generated three papers, creating the hypotheses, experimental code, data analyses, visualizations, and text without human modification.

  • One submission was accepted at the ICLR 2025 workshop with an average reviewer score of 6.33, ranking higher than many human-written papers.

  • Sakana also pointed out some caveats, including the AI making citation errors and workshop acceptance rates being higher than typical conference tracks.

  • The company concluded that the paper did not meet its internal bar for ICLR conference papers but displayed “early signs of progress.”

Why it matters: While this milestone comes with significant asterisks, it also represents a major early marker of AI's advancing role in academic research processes. Between models like Sakana’s and Google’s AI co-scientist, a seismic shift is getting closer and closer for the scientific world.

QUICK HITS

  • 🗂️ Luma AI - Recruiting Coordinator

  • 📱 Anthropic - Software Engineer, iOS

  • 🩺 Abridge - Clinician Scientist (MD)

  • 📊 Deepmind - Staff Quantitative UX Researcher

New legal filings revealed that Google owns 14% of Anthropic, with its investments totaling over $3B in the rival AI startup.

Alibaba researchers open-sourced R1-Omni, a new multimodal reasoning model that can ‘read’ emotions using visual and audio context.

Google DeepMind introduced Gemini Robotics and Gemini Robotics-ER, two Gemini 2.0-based models to help robots accomplish real-world tasks without training.

Perplexity launched a new Model Context Protocol (MCP) server for its Sonar model, allowing Claude to access real-time web search capabilities.

Snap released its first AI Video Lenses, powered by its own in-house model, offering premium subscribers three AR animations with new options planned to launch weekly.

Moonvalley released Marey, an AI video model that claims to be trained exclusively on licensed content for use in filmmaking — capable of creating 30-second-long HD clips.

Captions unveiled Mirage, a foundation model designed specifically for generating UGC-style content for ad campaigns.

COMMUNITY

Join our next workshop this Friday at 4 PM EST to learn about how to use Manus AI to boost productivity and automate everyday tasks easily, with Dr. Alvaro Cintas, The Rundown’s AI professor.

RSVP here. Not a member? Join The Rundown University on a 14-day free trial.

We’ll always keep this newsletter 100% free. To support our work, consider sharing The Rundown with your friends, and we’ll send you more free goodies.

That's it for today!

Before you go we’d love to know what you thought of today's newsletter to help us improve The Rundown experience for you.

Login or Subscribe to participate in polls.

See you soon,

Rowan, Joey, Zach, Alvaro, and Jason—The Rundown’s editorial team

Reply

or to participate.