Will Hawkins

Aaaand Gemini 3 is officially here! We worked tirelessly on its capabilities across the board. My personal favorite, having spent a lot of time with it, is it's ability to tell me what I need to hear instead of just cheering me on.

Frontier Safety Framework report for Gemini 3 Pro, presenting risk assessments and evaluation results in CBRN, Cybersecurity, Harmful Manipulation, Machine Learning R&D and Misalignment domains. storage.googleapis.com/deepmind-media…

check out our updated frontier safety framework - now including a harmful manipulation CCL!

Today marks the start of #ACMFAccT2025 with @oiioxford researchers and DPhil students presenting research on the societal challenges posted by AI. First up is Will Hawkins ‘Deepfakes on Demand' in the Responsible System Development session, today, 11.09am bit.ly/4lirTjJ

Really happy to announce the launch of SynthID Detector - a verification portal based on @GoogleDeepMind's SynthID watermarking technology that helps users determine if content, or part of it, was generated by Google’s AI tools. Read more in our blog: blog.google/technology/ai/…

🚨 New paper at @FAccTConference 2025! See thread for an overview or the full paper for details: arxiv.org/abs/2505.03859

🚨 New research led by @WillHawkins3 @oiioxford reveals a surge in publicly available deepfake image generators with nearly 35,000 model variants downloaded almost 15 million times since late 2022. Many are used to create non-consensual explicit images of women.

Per our Frontier Safety Framework, we continue to test our models for critical capabilities. Here’s the updated model card for Gemini 2.5Pro with frontier safety evaluations + explanation of how our safety buffer / alert thresholds approach applies to 2.0, 2.5, and what’s coming. storage.googleapis.com/model-cards/do…

Our incredible team built many models announced here, including image, voice, music and video generation! And: I'm moving to London this summer, and I'm hiring for research scientist and engineering roles! Our focus is on speech & music in Zurich, Paris & London. DM/email me.

We've written a paper (145 pages!!) about our approach for AGI safety at @GoogleDeepMind. It's not just scalable oversight and interp -- so much more needs to come together. deepmind.google/discover/blog/…

Just released GDM’s 100+ page approach to AGI safety & security! (Don’t worry, there’s a 10 page summary.) AGI will be transformative. It enables massive benefits, but could also pose risks. Responsible development means proactively preparing for severe harms before they arise.

ShieldGemma 2 has arrived 🛡 The turnkey solution for image safety checks is based on the new Gemma 3 4B model, and key features include: -Classifications across 3 safety categories -Easy integration of safety checks into image-based applications -Customizable policy templates

Check it out on: Hugging Face: huggingface.co/google/shieldg……Kaggle: kaggle.com/models/google/…… Huge credit to Dana, Wenjun, Ryan and the team who have worked tirelessly to ship 🚢 Please give us feedback to help us build maximally useful models for the community :)

🚀 ShieldGemma 2 is now out! We’re launching a 4B parameter image safety classifier built on the wonderful new Gemma 3 to help open model developers & users flexibly build their safety requirements into systems. Read more, and access links below! developers.googleblog.com/en/safer-and-m…

ShieldGemma 2 goes beyond text and now supports images. We continue to drive open source tooling to help make use of AI models safer. developers.googleblog.com/en/safer-and-m…

Gemma 3 is here and its the best open non-reasoning model on LMSYS! 🚀 @GoogleDeepMind Gemma 3 is an open, multimodal (text + vision), multilingual LLM with a context of 128k tokens and comes in 4 sizes! TL;DR: 4️⃣ Four sizes with 1B, 4B, 12B, 27B as pre-trained and instruction-tuned versions 🥇 #1 open, non-reasoning model in LMArena with a score of 1338, above o1-mini 🖼️ Text and image inputs and text output for 4B, 12B, and 27B sizes 🌍 Multilingual with support of 140+ languages 📈 Increased context window of 128K tokens (32K for 1B model) 🧠 Post-training combines RLHF, RLMF, and RLEF and model merging 😮 Gemma 3 12B outperforms Gemma 2 27B, same for 4B and 9B 🔍 Vision encoder based on SigLIP with adaptive window algorithm for high-resolution images 🧪 Trained on up to 14T tokens (27B) on Google TPUs using JAX 🤗 Available on @huggingface, @Kaggle and AI Studio to try

Welcome Gemma 3, our new open-weight LLM from @GoogleDeepMind. All sizes (1B, 4B, 12B and 27B) excel on benchmarks, but the key result may be the 27B reaching 1338 on LMSYS. For this, we scaled post-training, with our novel distillation, RL and merging strategies. Happy building!

Google DeepMind has just released Gemma 3 family of models 😍🔥 > six new long context multilingual vision LMs (4B, 12B, 27B, 128k context length, 140+ languages) > two new small LMs (32k, 1B) > ShieldGemma for multimodal safety Based opinion: they're good 🤩 Keep reading 👀

I’m so happy to announce Gemma 3 is out! 🚀 🌏Understands over 140 languages 👀Multimodal with image and video input 🤯LMArena score of 1338! 📏Context window of 128k Available in AI Studio, Hugging Face, Ollama, Vertex, and your favorite OS tools 🚀Download it today! developers.googleblog.com/en/introducing…

@NVIDIAAIDev 🛡 Also launching: ShieldGemma 2, a powerful 4B image safety checker built on Gemma 3. Developers can customize ShieldGemma 2 to suit their safety needs.