Good Morning! Google I/O 2025 just unleashed a tidal wave of AI features—from the snappy new Gemini 2.5 Flash Preview to agentic Colab and Jules the async coding sidekick—so you can prototype and ship at warp speed. A fresh Lawrence Berkeley Lab report warns that by 2028 AI workloads could gulp 165–326 TWh annually, enough to power nearly a quarter of U.S. homes. And MIT Technology Review’s “Power Hungry” deep dive shows how a single query can burn hundreds of joules once you factor in servers, cooling, and carbon intensity—courtesy of MIT.

— Forrest Knight & Meghanadh Vasireddy

Google I/O 2025’s Fresh AI Models & Agentic Tools

Context: Google I/O 2025 was a full-on AI bonanza for devs. Mat Velloso (VP of AI Developer Platform at Google DeepMind) reminded us that “developers are the architects of the future,” and then unleashed a slew of model and tooling upgrades to supercharge your workflows. Whether you’re building chatbots, vision apps, or full-stack AI services, there’s something here to speed up your next proof-of-concept.

What’s new:

• Gemini 2.5 Flash Preview: Faster, leaner coding and reasoning, with thought summaries to help you manage token costs.

• Gemma 3n & PaliGemma: Lightweight open multimodal models for on-device audio/text/image/video, plus a vision-language champ for captioning and VQA.

• Agentic Power-Ups: A sneak-peek at agentic Colab that auto-fixes your notebooks, plus Jules—the async coding sidekick that spins up VMs, clones repos, and preps PRs.

• Workflow Boosters: Firebase Studio now auto-provisions backends from Figma, and Stitch can whip up UI designs and front-end code from plain English or images.

Dive into the previews today in Google AI Studio and Vertex AI—Flash goes GA in early June, Pro follows soon after. Grab the keys, build faster, and let these tools handle the heavy lifting so you can focus on the fun stuff.

MIT’s Uncovering the Carbon Footprint of Every Query

Context: You’ve probably noticed AI is gobbling up cycles and data-center juice like never before. But what’s the real cost of that chat prompt, image render, or video snippet?

What’s new: MIT Technology Review’s latest “Power Hungry” analysis (May 20, 2025) crunches the numbers from expert interviews, open-source benchmarks, and grid data to map AI’s carbon footprint down to the joule:

• Per-query footprints:

  • Text LLMs: ~114 J for Llama 8B → ~6 700 J for Llama 405B

  • Stable Diffusion (1 024×1 024): ~2 300 J

  • Open-source video (5 s @16 fps): ~3.4 MJ

They forecast that by 2028, AI could gulp 165–326 TWh annually—enough to power 22% of US homes—driving a data-center boom that breaks past decade-long efficiency gains. Since most centers rely on fossil-heavy grids, regional carbon intensity swings wildly from ~70 gCO₂/kWh in sunny afternoons to >300 g at night or in coal-dependent areas.

Takeaway: A single AI call feels negligible, but at billions of daily hits, the cumulative toll is massive. The piece urges closed-box AI providers to open their energy books and for smarter grid planning to keep our AI ambitions sustainable.

Sponsored

Snyk Launch 2025 | Secure the Future of AI Application Development

AI is definitely speeding up app development and letting us build bigger, more complex things. However, security needs to evolve just as quickly, or we risk new vulnerabilities. Balancing innovation with protection is a growing challenge.

Snyk is holding an event, Snyk Launch 2025, on May 28th. They'll be sharing their latest innovations to help development teams tackle this security gap, especially with AI in the mix.

Here's a quick look at the agenda:

• Peter McKay (CEO): Discussing security risks in AI enterprises & leadership strategies.

• Danny Allan (CTO): A deep dive into Snyk's new AI security technology.

• Manoj Nair (CIO): Talking about expanding secure AI through ecosystem partnerships.

• Live Demos: Showcasing their AI-driven security tools in action.

If you're working on keeping software secure in this fast-paced AI world, this could be insightful.

Register for Snyk Launch: https://snyk.plug.dev/tDSUQbj

🔥 More Notes

• World’s First Petahertz-Speed Phototransistor in Ambient Conditions: Researchers built a phototransistor that manipulates electrons with light pulses under real-world conditions at petahertz speeds—breaking the trillion-cycles-per-second barrier and opening doors for ultrafast optoelectronics.

• Microsoft Simplifies AI Training on Companies’ Own Data: At Build 2025, Microsoft unveiled “Copilot Tuning,” a low-code tool that lets non-engineers fine-tune AI models on their own datasets, plus multi-agent orchestration features to streamline enterprise AI development.

• Russian GRU Cyber Actors Targeting Western Logistics & Tech Firms: A joint CISA/FBI/NSA advisory warns that Russian GRU hackers are actively exploiting supply-chain networks and tech infrastructures—urging firms to harden endpoints and monitor for IOCs.

📹 Youtube Spotlight

Was this forwarded to you? Sign Up Here