
Full story, 5-act design breakdown, gameplay GIFs, and what stood out testing Gemini 4 Carbon: rkrants.blogspot.com/2026/11… Source (zero external 3D models/audio): github.com/rabimba/dino-3d @GeminiApp @googledevs

Working on #WebVR. ex-@mozilla . Opinions are my own. Googe Developer Expert - Web. Connect with me at @rabimba@toot.cafe

Full story, 5-act design breakdown, gameplay GIFs, and what stood out testing Gemini 4 Carbon: rkrants.blogspot.com/2026/11… Source (zero external 3D models/audio): github.com/rabimba/dino-3d @GeminiApp @googledevs

Courtesy of @GoogleDeepMind, I got early access to Gemini 4 Carbon (@GeminiApp). Alongside heavier engineering tests, I one-shotted Chrome’s 2D Offline Dino into a 152KB 3D WebGL runner, Wi-Fi satellite boss, 2s time-rewind & meteor headbutts. Play: rabimba.github.io/dino-3d/

Also fun fact the picture above and this one, both are not real and generated by it :D Image inception? :P

Its biggest strengths: 🧠 Built-in "thinking" & reasoning capabilities 🖼️ Native image generation (all the way up to crisp 4K!) ⚡ Incredible speed/latency at a super developer-friendly price point Huge congrats to the teams at @GoogleAIDevs and @GoogleAIStudio on this launch!

The secret is finally out! 🍌🚀 I’ve been testing Gemini Nano Banana 2.1 for quite some time now as part of the Gemini API Early Access Program, and it is officially live today! I am honestly so impressed by this model.

Peak tok/s is a vanity metric. If your serving stack violates TTFT & TPOT SLOs, production goodput is near zero. I broke down how to benchmark vLLM fairly on @OpenXLA /TPUs vs CUDA, covering interconnects, compilation costs, & MBU: youtu.be/-AjLjr5kPqY?si=JJns… @GoogleDevsNA

I built ConferenceRank: 980+ CORE-ranked venues, 300+ live AoE deadlines with 1-click calendar sync, acceptance rates, and an AI abstract-to-venue suggester. Free, no login 🚀 rabimba.github.io/Conference…

The video for my DevLab 2026 talk is finally out! 🎬 If you couldn't make it live, here is the full breakdown of "The Throughput Trap." We look past mirage-like public benchmarks to explore real-world LLM serving. 📺 Watch: youtu.be/-AjLjr5kPqY📝

Tired of tracking CS conference deadlines and doing AoE timezone math at 2 AM? I built ConferenceRank: 980+ CORE venues, 300+ live AoE countdowns, 1-click Cal sync, acceptance rate stats, and an AI venue suggester. Free & zero-login: rabimba.github.io/Conference…

From “What can I get you?” to “I know exactly what you’ll like.” ☕🤖 Today at #RenderATL 2026, I’m leading a hands-on lab on building a Coffee Recommendation Agent with Google ADK + RAG. 📍 BLD ATL, Floor 3 – Expo Hall ⏰ 11:30 AM ET

Presenting our paper on passive prompt injection in LLM-based SOC log analysis at @USENIXSecurity on Aug 12! 🛡️ A collab by @PayPal AI Labs, @UHouston & @KentState. Catch our poster too! In the area? Let's meet up! 📄 arxiv.org/abs/2607.14493 🔗 usenix.org/conference/usenix…

Excited to speak at @Google's Dev Lab Summer 2026 (invited by the Google ML & PyTorch teams)! "Benchmarking vLLM on XLA Backends: From CUDA to OpenXLA for LLM Serving" @JeffDean also has a talk! Come say hi! rsvp.withgoogle.com/events/d… @GoogleDevsNA @googledevs @PayPal

"How good is this driver, right now, this lap?" If your AI coach doesn't answer that, it's guessing. In Part 2 of my Racecraft series, I explain how we use input jerk and coasting ratios to build a continuous, skill-adaptive DriverModel. 🏎️ rkrants.blogspot.com/2026/06…

"How good is this driver, right now, this lap?" If your AI coach doesn't answer that, it's guessing. In Part 2 of my Racecraft series, I explain how we use input jerk and coasting ratios to build a continuous, skill-adaptive DriverModel. 🏎️ rkrants.blogspot.com/2026/06…

If an AI coach gives you feedback 800ms late, it’s worse than silence. At 130 km/h, a late cue is a safety hazard. 🏎️ I’m breaking down the 5-part technical series behind "Racecraft", my on-device, real-time AI coach built on Gemma 4. rkrants.blogspot.com/2026/06…

Are your AI benchmarks just giving you a false sense of security? As models scale, they hit "phase transitions" where your metrics break. If you're relying on static tests, you're just measuring how fast you're running toward a cliff. rkrants.blogspot.com/2026/06…