Guillaume Laforge (@glaforge) header
Guillaume Laforge (@glaforge) profile picture

🥑 Developer Advocate @GoogleCloud 🧠 Focusing on GenAI 🤖 ⭐️ @ApacheGroovy co-founder 🏆 @java_champions 📣 @lescastcodeurs podcast 🦋 @glaforge.dev

Paris, France
Joined April 2008
Photo by Guillaume Laforge (@glaforge)Photo by Guillaume Laforge (@glaforge)Photo by Guillaume Laforge (@glaforge)Photo by Guillaume Laforge (@glaforge)Photo by Guillaume Laforge (@glaforge)Photo by Guillaume Laforge (@glaforge)Photo by Guillaume Laforge (@glaforge)Photo by Guillaume Laforge (@glaforge)Photo by Guillaume Laforge (@glaforge)Photo by Guillaume Laforge (@glaforge)
Before heading on vacation, I built an unofficial @Java SDK for #Google @antigravity 🚀 If you want to embed AI agents directly into enterprise Java applications, #GitHub Actions, or custom #CLI tools, here is how it works and what you can do with it 🧵 👇
2
6
1
30
3,356
Guillaume Laforge retweeted
The clever thing to know about @antigravity Orchestrator 💡 is that it is not a router to choose a model for a task! It dynamically switches between models within a single task!! 🔄🚀 That's so much better. You don't have to have your big model in charge, handing off tasks to smaller models. What if one of those tasks needs some extra oomph, it dynamically switches to the most appropriate model at the right time!! The combination of Orchestrator and /swarm is 🤯🤯
We’re rolling out Orchestrator in @Antigravity to Google AI Ultra subscribers, letting you run multiple models in a single session instead of routing the entire task to one model. Orchestrator pairs a frontier model with a fast execution model: Carbon sets the plan, hands off code edits and testing to Hydrogen, and steps back in to repair the plan if it hits friction — reaching near-Carbon task success rates on Terminal-Bench 4.0 while consuming 24% less quota and completing 19% faster on average.
1
3
10
933
Guillaume Laforge retweeted
Introducing Gemini 4 Neon! Pushing the frontier across domains and punching above its model class in a meaningful way. To me, Neon feels like the right balance of intelligence, speed, and usability. Available now across @GoogleAIStudio, the API, and more!
593
696
517
10,051
646,255
I'm super happy to echo this release, as I contributed the #Gemini Decision 1 model support in #LangChain4j 1.24.0!
2
1
7
435
You can read more about the 𝚐𝚎𝚖𝚒𝚗𝚒-𝚍𝚎𝚌𝚒𝚜𝚒𝚘𝚗-𝟷 model support in #LangChain4j in the documentation: docs.langchain4j.dev/integra…
1
116
Rather than looking at a wall of text, I really like @antigravity's /𝚐𝚎𝚗𝚎𝚛𝚊𝚝𝚒𝚟𝚎-𝚞𝚒 command to visualize things. Here, I'm updating an agent that generates illustrated Sci-Fi stories, and I clearly see the structure & components.
1
2
7
418
My favorite way and place to deploy my #java applications these days!
By default, source-based deploys on Cloud Run use a Cloud Build VM to build and push an image to Artifact Registry. With --no-build, you can skip Cloud Build and deploy in 7 seconds by overlaying your local files onto a Google-managed base image. Blog: wietsevenema.eu/blog/2026/cl…
3
594
Guillaume Laforge retweeted
Nouveau mode ajouté à ScanPaname, la carte du métro parisien en direct Les tunnels apparaissent sous la surface, chaque ligne à sa profondeur et les rames bougent en direct.🏄
37
113
9
1,077
99,919
Guillaume Laforge retweeted
Android's source code is huge. 1k git repos, 1.62 million files, and it takes 300GB for a built workspace. How did we reduce checkout time from 36 minutes to 8 seconds? Full builds from 80 minutes to 2 minutes? We just open sourced the solution. developers.googleblog.com/ac…
14
74
6
628
25,085
Guillaume Laforge retweeted
Text, code, images, audio, and video mapped into a single, unified embedding space. Watch this five minute explainer on how EmbeddingGemma 2 uses modality specific encoders, a bidirectional transformer, and a two stage training recipe built on Gemma 4 and Gemini Embedding 2 to run efficiently in resource constrained environments. Dive into the full architecture and evaluation results in the technical report: arxiv.org/abs/2610.37364
23
65
2
707
29,238
Guillaume Laforge retweeted
Honestly, I was so blown away by the quality of the talks this year! If you want to catch up on the awesome slides and demos, the full WebGPU Paris Meetup 2026 presentations are up on YouTube: 👉 youtube.com/playlist?list=PL…
The WebGPU Paris Meetup is NEXT WEEK! 🇫🇷 The lineup is wild: Unity's new renderer, Hugging Face web kernels, WebLLM, William Candillon, and Hexagon's geospatial tech! 100% technical, pure engineering love. Seats are limited, so secure yours now! 👇 🎟️ rsvp.withgoogle.com/events/w…
5
16
1,593
Guillaume Laforge retweeted
The @CopilotKit crew wrote up a great summary of A2UI 1.0, the standard protocol for generative UIs. Get details on what's new, what's changed, and how to get started. copilotkit.ai/blog/introduci…
2
15
50
3,569
Guillaume Laforge retweeted
🚨Google is preparing a full line of Gemini 4 According to current tests and releases, the models are divided like this: ▪ Hydrogen is a light workhorse. The fastest and cheapest, sharpened for everyday dialogues, simple tasks and agency cycles. ▪ Carbon is a station waggon. The balance of speed and strength, well pulls conventional development and multimodal tasks. ▪ Argon is a model of deep reasoning. Strong in complex logic, refactoring and long chains. ▪ Krypton is a heavy flagship. Positioned as a competitor of Fable 5.2 and GPT-6.2 Sol, focussed on complex applications, simulations and high-load tasks. The heavier the element, the more computing power and the higher the capability ceiling. Google's strategy is clear: give simple tasks to light models, complex ones to difficult ones, and dynamically route between them. One lineup covers everything - from regular chat to serious engineering work.
21
22
3
494
41,368
Guillaume Laforge retweeted
Typing "/btw how's it going" in @antigravity is so handy. Like tapping shoulder of your sub-agents for quick check-in. Sometimes they even give me a rough ETA.
5
1
49
3,105
Guillaume Laforge retweeted
Our decisions API supports multimodal input across text, images, audio, video, and PDFs. It is the only one on the market that supports this. It is super powerful, check out the real-time call router below powered by the Gemini Decisions API.
15
19
6
249
16,502
Guillaume Laforge retweeted
Since Google launched Carbon and Hydrogen last week.. my feed has literally turned around.. The whole "Gemini isn't good" narrative has almost disappeared I agree Carbon is slow.. it's good but pretty slow. Hydrogen on the other hand surprised me a lot.. after my past experience with Flash models I didn't expect to like it, but it's so proactive, which previous Flash models really lacked Hallucinations aren't fully gone either.. Carbon can be a bit untrustworthy, but it's pretty low if you're careful with your prompts and keep an eye on it. Hydrogen feels much better than Carbon here tbh.. it's fast and smart I think their successors (Krypton and Oxygen) will be the moment people stop complaining and start switching to Gemini in masses.. save this for when those drop And Google is THIS close to a batch of models that feel like Opus 5.5.. where you can confidently hand it any task and be sure it'll figure out the best way to do it None of this is just opinion.. it's based on the tests I've been running, and I'm already working on review threads for both. It's pretty interesting tbh and they'll back up every word above!! And parallely the Antigravity team has been updating the app and adding features like crazy.. and i hope this doesn't stop happening as there's a lot of improvements to be made.. Congrats to GDM and Agy teams for the amazing releases last week!
46
19
7
647
29,495
Guillaume Laforge retweeted
What if prototyping sandboxed AI agents on your Mac felt frictionless - 3s boot, no Docker, no local K8s - and the exact same manifests scaled straight to Kubernetes? ate-local + AX Central for Google's AX & Agent Substrate: Apple container microVMs, Podman & K8s.
2
1
4
234
Guillaume Laforge retweeted
What if your code just needs a fast, calibrated judgment to branch on? 🤔 Google recently introduced the Gemini Decisions API (powered by Gemini Decision 1) to evaluate up to 1,000 independent questions over multimodal inputs in a single ~200 ms pass with 0 output tokens billed. In my latest blog post, I break down: 🔹 Yes/no (noul), multiple choice, and ordered rubric (score) primitives with calibrated confidence 🔹 Practical gotchas (v1beta setup, 0.01–0.99 bounds, and option key sorting) 🔹 Cross-modal fusion across text, images, audio, video, and PDFs in one pass 🔹 Building a support triage & dynamic Gemini model router control plane in Python Blog post: medium.com/google-cloud/gemi… GitHub samples: github.com/meteatamel/genai-… #GenAI #Gemini #AIAgents #MachineLearning #GoogleCloud
4
10
478
Guillaume Laforge retweeted
We are cooking at Google. Look at the last few weeks: Models • Gemini 4 Carbon: now GA in the API and AI Studio, rolling out to Ultra • Gemini 4 Hydrogen: ~2.5x faster than 3.8 Flash, 1M context, $1.50/$7.50 per 1M tokens flat • Gemini Omni 1.2 Flash: 15-second clips, 30% cheaper • Gemini Robotics 2.5: new vision-language-action model • Genie 4: real-time interactive world model • Gemini Decisions API: score up to 1,000 questions in one call Agents and dev tools • Antigravity: Orchestrator, Swarm, Dynamic Workflows, Voice mode, Browser Use, Knowledge Work layout and a plugin marketplace • ADK 2.14: execution environments, A2A and eval go stable • A2UI 1.0: open protocol for generative UI • Elephant Goldfish: open-source plugin to design and review with coding agents Hardware • Gemini glasses: pre-orders open with Warby Parker and Gentle Monster • XREAL AURA: first wired Android XR glasses with Gemini • TPU 8tx and 8i Max, plus TPU Sync open-sourced • Gemma 4 running on a TPU in space • Pixel Tag and Gemini Intelligence on Pixel 10 Bullish on Google.
164
103
31
1,735
152,704
Guillaume Laforge retweeted
We're rolling out Dynamic Workflows in Google Antigravity. Use /run-workflow and the agent writes a Python script to orchestrate parallel and sequential subagents with structured handoffs, giving you deterministic control flow instead of prompt drift. Review the script before it runs, and save it to re-run the same flow anytime.
242
248
118
3,259
257,836