Google DeepMind launches Diffusion-GEMMA for 8× faster inference

#1 𝕏

Google DeepMind launched Diffusion-GEMMA, a diffusion-based text-generation model that parallelizes token sampling to cut inference latency by up to 8× while matching autoregressive quality. They’ve open-sourced the code and benchmarks for developers.

#2 𝕏

Claude launched public beta for scheduled deployments and vault-based environment variables in Claude Managed Agents, and brought dynamic workflows in Claude Code to general availability—enabling agents to run on schedules, use tools securely, and tackle more complex tasks.

#3 📝 OpenAI News

Access OpenAI models and Codex through your Oracle cloud commitment - Announces that OpenAI models and Codex are available through Oracle Cloud commitments, enabling customers to access OpenAI's offerings via Oracle's cloud infrastructure. The announcement describes a partnership to broaden access options for enterprise users.

#4 📝 Anthropic News

Policy on the AI Exponential - Anthropic proposes two policies—a detailed Advanced AI Framework and an Economic Policy Framework—to govern rapidly advancing AI, with the Advanced AI Framework giving governments the legal authority to block or deter dangerous deployments, requiring frontier developers to test models, publish safety frameworks and regular risk reports, engage independent evaluators, secure model weights and training infrastructure, and face civil penalties tied to global annual revenue that escalate with repeated violations. The rules would apply to models trained with more than 10^25 FLOPs or developed by companies with >$500M in AI revenue or >$1B in AI R&D spend, cite examples like Claude Mythos Preview finding thousands of high‑severity vulnerabilities across major OSes and browsers, identify four catastrophic risks (biological, cyber, loss of control, automated R&D), and recommend resilience measures such as gene synthesis screening, biosurveillance, PPE stockpiles, hardening critical software, and a government function to track frontier cyber capabilities.

Also covered by: @Anthropic

#5 📝 Claude Code Blog

The evolution of agentic surfaces: building with Claude Managed Agents - Examines how to design and build agent-driven interfaces using Claude Managed Agents, covering patterns for agentic surfaces, platform integrations, and developer workflows to deploy agentic experiences.

#6 𝕏

Sundar Pichai unveiled DiffusionGemma in Gemma 4—an open experimental text diffusion model that achieves up to 4× faster inference by generating entire text blocks simultaneously instead of token-by-token.

#7 𝕏

xAI launched Grok Voice, an API delivering human-like timing, tone, and warmth with Think Fast 1.0 performance on the EVA-Bench Pareto frontier—all at a fraction of competitors’ prices.

#8 𝕏

Santiago reports that a leading voice AI provider has cut its TTS, STT, and LLM API prices by 50%, with additional volume-based discounts that drive costs even lower as you scale.

#9 𝕏

Santiago announces you can now run a trillion-parameter model on personal hardware and launches Builder’s Arcade, a 5-day NVIDIA-accelerated Azure AI Foundry mini-program.

#10 𝕏

Dario Amodei unveiled Anthropic’s policy proposal to help governments address frontier AI risks and a financially backed framework for managing job displacement.

Also covered by: @Anthropic

#11 𝕏

Google Research launched Regularized f-Divergence Kernel Tests, a framework for auditing machine unlearning and differential privacy. It detects localized data shifts with fewer samples than traditional methods.

#12 📝 OpenAI News

How an astrophysicist uses Codex to help simulate black holes - Astrophysicist Chi-kwan Chan uses Codex to generate, implement, and test new numerical schemes that change how particle motion is tracked so simulations no longer need to compute every tiny corkscrew of electrons and ions spiraling around magnetic field lines near a black hole, a bottleneck that currently forces extremely small timesteps. If the testable algorithms Codex proposes work, they could eventually enable simulations of trillions of particles, helping the Event Horizon Telescope team move from the 2019 black hole image toward producing videos and studying previously inaccessible plasma physics.

#13 𝕏

Philipp Schmid released a Developer Guide for Diffusion Gemma, Google’s open-source diffusion model, with end-to-end code to train, fine-tune and deploy on Vertex AI. It also walks through best practices for high-res image synthesis, inpainting and prompt engineering.

#14 𝕏

Philipp Schmid unveils an interactive blog that shows how one API call spins up an isolated sandbox for Gemini Managed Agents to reason, call tools, execute code and parse outputs.

#15 𝕏

Madhu Guru warns that since Gemini’s early days enterprises often picked the cheapest models and sacrificed quality.

#16 𝕏

Boris Cherny shows that the /usage view breaks down exactly which skills, MCPs, and plugins consume your tokens, so you can fine-tune and optimize your token usage.

#17 𝕏

Garry Tan provides an idempotent patch script to update OpenClaw’s dist folder so it properly recognizes and sends Claude Fable 5’s new adaptive-thinking (thinking.type:"adaptive" + output_config.effort) params via API key.

#18 𝕏

xAI explains how eToro’s agent Tori now taps Grok models and real-time SpaceXAI data feeds to deliver on-demand market sentiment analysis for consumers.

#19 𝕏

LlamaIndex 🦙 Day 0 Anthropic Fable 5 in ParseBench scored 90.02% content faithfulness (vs 86.19% Gemini 3 Flash; 86.81% GPT-5.5) and 72.62% semantic formatting (vs 58.35%/60.

#20 𝕏

Google DeepMind’s latest research in Sierra Leone explores AI teaching assistants as partners to help educators manage surging student numbers, amplifying their reach without replacing teacher expertise.

#21 📝 Simon Willison

Anthropic Walks Back Policy That Could Have ‘Sabotaged’ AI Researchers Using Claude - Anthropic announced it will make Fable 5's safeguards for frontier LLM development visible after backlash over previously invisible interventions that could limit model effectiveness without notice. The company says flagged requests will visibly fall back to Opus 4.8 and that API responses will include reasons for refusals, acknowledging the prior choice of invisible safeguards was the wrong tradeoff.

#22 𝕏

Aravind Srinivas introduced Claude Fable 5 as the new orchestrator model inside Computer—now available to all Pro and Max users—and highlighted its strength in handling long-running agentic workflows.

#23 𝕏

Cursor made its code review agent over 3× faster, 22% cheaper, and 10% better at finding bugs. You can also use the new /review command to run Bugbot locally and catch issues before pushing code.

#24 𝕏

Sundar Pichai open-sourced Diffusion GEMMA’s faster text-generation model weights on Hugging Face, available under an Apache 2.0 license.

#25 𝕏

Dario Amodei warns that transparency mandates for frontier AI, once critical given unclear risks, are no longer enough and must be backed by stronger regulatory measures.