1Gemini 3.5 Live Translate ships real-time speech-to-speech in 70+ languages.

Google just removed the awkward pause from translation. Live Translate starts converting your speech as you talk — no turn-taking, no waiting for sentence boundaries — across 70+ languages. This is the most ambitious production deployment of streaming speech-to-speech translation yet, and it's available now in Google Translate and AI Studio. If you build anything multilingual, this changes your cost-benefit math overnight. (1,787 likes | 233 RTs) Read more →

ChatGPT's #MessiMode goes viral. OpenAI turned image generation into a consumer engagement machine — personalized flag-colored hair generation racked up 9.5K likes and counting. The feature itself is lightweight, but the signal is loud: creative image gen is what drives mass consumer adoption, not benchmarks. (9,562 likes | 708 RTs) Read more →

DiffusionGemma 26B lands on HuggingFace. Google's parallel text generation model already has 200+ likes within hours of release. The 26B-parameter instruction-tuned variant is Apache 2.0 licensed and ready to benchmark — the community is moving fast on this one. (203 likes) Read more →

2Anthropic publishes the production playbook for Claude Managed Agents.

This isn't a research preview — it's the full architectural guide for building production agent systems with Claude Managed Agents. Scheduled deployments, vault-based secrets management, and patterns for orchestrating multi-agent workflows in production environments. If you've been duct-taping agent scheduling with cron jobs and env files, this is your migration path. Read more →

Code with Claude Tokyo: scheduled deployments and env vaults hit public beta. Three platform updates dropped at the Tokyo event — scheduled agent deployments, secure environment variable vaults, and dynamic workflows now GA. The vault feature alone solves the "how do I give my agent API keys without hardcoding them" problem that's been plaguing every production agent deployment. (3,664 likes | 237 RTs) Read more →

Poetic HQ claims 99%+ accuracy on multi-hour tasks with 10x fewer tokens. A new agentic system from @swyx's radar that allegedly achieves near-perfect accuracy on complex, long-running tasks while burning an order of magnitude fewer tokens. If independently validated, this challenges the "throw more compute at it" orthodoxy. Watch for benchmarks before betting on it. (851 likes | 128 RTs) Read more →

HuggingFace Transformers v5.11 adds native DiffusionGemma support. The ecosystem is keeping pace with Google's research — pip install --upgrade transformers and you can run parallel text generation locally. The release also bundles several new model architectures, but DiffusionGemma integration is the headline. Read more →

3How Fable 5 edited its own launch video — no human editor touched the timeline.

Fable 5 autonomously orchestrated transcription services, ffmpeg, color grading, Figma MCP, and Remotion to produce a polished launch video from raw footage. No video editor intervened at any point. The walkthrough details every agentic decision — from clip selection to timing to color correction — and the result is indistinguishable from human-edited output. This is the clearest demonstration yet that agentic AI capabilities extend far beyond code generation into creative production workflows. (2,633 likes | 140 RTs) Read more →

4DiffusionGemma achieves 4x speedup by generating text blocks in parallel.

Google DeepMind took the diffusion approach that revolutionized image generation and applied it to text. Instead of predicting one token at a time, DiffusionGemma generates entire blocks simultaneously on dedicated GPUs — up to 4x faster than autoregressive decoding. Released under Apache 2.0 with HuggingFace support. The speedup is real but currently requires specific hardware configurations; don't expect miracles on consumer GPUs yet. Read more →

DeepMind RCT proves AI tutoring measurably improves learning in Sierra Leone. A proper randomized controlled trial — not a vibes-based case study — shows Gemini's Guided Learning feature boosts student engagement and learning outcomes in resource-constrained environments. This is the kind of rigorous evidence the "AI for good" space desperately needs. Read the methodology if you're building impact measurement frameworks. Read more →

FrontierCode shows Mythos and Fable scale where Opus 4.8 and GPT-5.5 plateau. On the FrontierCode Diamond benchmark, both Opus 4.8 and GPT-5.5 stop improving with additional compute — but Mythos and Fable keep climbing. Anthropic's posttraining specifically targets sustained multi-hour problem solving, and it shows. If you're routing tasks by difficulty, send the hardest ones to Fable. (199 likes | 13 RTs) Read more →

5Dario Amodei draws a policy line: three new initiatives to close the AI governance gap.

Anthropic's CEO published "Policy on the AI Exponential" — a manifesto arguing that the gap between AI capability growth and policy capacity is the defining risk of the moment. Three new Anthropic policy initiatives aim to close it. Whether you agree with his framing or not, this is the most concrete governance commitment from any frontier lab this year. (3,299 likes | 258 RTs) Read more →

LeCun amplifies pushback on Fable 5's AI development restrictions. Yann LeCun and the open-source community are calling out reported capability degradation when using Fable 5 for AI development tasks. The criticism cuts to the core of the open-vs-closed debate: should a frontier model decide what you're allowed to build with it? Test Fable 5 on your own AI dev workflows before taking sides. (3,270 likes | 600 RTs) Read more →

Latent Space publishes the most comprehensive independent Fable 5 teardown. Beyond the benchmarks and the hype, this deep-dive covers the 30-day mandatory data retention policy that's sparked serious developer pushback. If you're evaluating Fable 5 for production use, this is the analysis that separates signal from marketing. Read more →

Simon Willison: when Fable stops helping you, you'll never know. The sharpest critique of Fable 5's refusal system isn't about what it blocks — it's about transparency. When a model silently degrades output quality instead of explicitly refusing, users can't distinguish a bad answer from a blocked one. That's a trust problem no benchmark captures. Read more →

MiMo Code V0.1: Xiaomi open-sources a terminal coding agent with million-token context. Xiaomi enters the coding agent race with an open-source tool powered by MiMo V2.5 — lossless compression, multimodal support, and a context window that swallows entire codebases. Free for a limited time. The open coding agent field just got more crowded, and that's good for everyone. (784 likes | 75 RTs) Read more →

Google and HuggingFace launch the Fast Gemma Challenge — agents optimizing agents. A community challenge where dozens of AI agents collaborate to optimize DiffusionGemma inference speed. The format is novel: agents working together on open research problems, not humans competing on leaderboards. Follow it for real-world optimization techniques you can steal. (1,177 likes | 112 RTs) Read more →

Diffusion-Based Text Generation (Parallel Block Decoding): Every large language model you use today generates text one token at a time — predict the next word, append it, repeat. DiffusionGemma flips this by borrowing from image generation: start with a noisy block of text and iteratively refine all tokens simultaneously, like a photo developing in a darkroom. The result is up to 4x faster generation because the model doesn't have to wait for each token before starting the next. The tradeoff is that parallel decoding can sacrifice some coherence on long outputs — the model doesn't get to condition each word on the exact word before it. Today's 4x speedup claims from Google make this the moment to understand why diffusion models, originally built for images, might break autoregressive decoding's monopoly on language.

• Claude Code v2.1.172: Nested sub-agents now go 5 levels deep, plus AWS region auto-detection and plugin marketplace search. Link
• Fable/Mythos data retention: Anthropic confirms 30-day mandatory retention — no zero-retention option available. Link
• Claude Desktop's VM problem: Spawns a 1.8 GB Hyper-V VM on every launch — even for chat-only use. (321 likes | 228 RTs) Link
• North-Mini-Code-1.0: Cohere's compact coding model gains traction with 153 likes and 1.8K downloads. Link
• Codex Mobile: Branch management, worktrees, usage dashboards, and inline review from your phone. (306 likes) Link
• TechDirt: CEOs treating AI as headcount replacement are just bad CEOs. (360 likes | 143 RTs) Link

When an AI model edits its own launch video, "creative tool" gets a new definition. A developer walkthrough showed Fable 5 autonomously orchestrating transcription, ffmpeg, color grading, Figma MCP, and Remotion to produce a polished video — no human editor touched the timeline at any point. This isn't an AI generating a 15-second clip from a text prompt. It's a model making editorial decisions: which clips to cut, where to place transitions, how to grade the color palette for consistency. Video production just became a programming problem. The implications ripple beyond video — every media workflow that currently requires a human to sequence tools (audio mastering, motion graphics, podcast editing) is now in scope for agentic automation. The quiet part: the tools Fable orchestrated aren't new. ffmpeg, Remotion, and color grading pipelines have existed for years. What changed is that a model can now chain them with editorial judgment. That's the real unlock — not better tools, but an agent that knows when and how to use the ones we already have. (2,633 likes | 140 RTs) Read more →

Until next time ✌️