Posts

Google’s Gemma series just leveled up in a very practical direction. Gemma 4 12B, released June 3, is a unified, encoder-free multimodal model that handles text, images, and audio — and does it on a 16GB laptop. For anyone building agentic workflows who’s been waiting for an open-weight multimodal option that doesn’t require a data center, this is the release to pay attention to. What “Encoder-Free” Actually Means Traditional multimodal models bolt separate encoder components onto a language model — a vision encoder, an audio encoder — and then fuse their outputs. This creates architectural complexity, increases parameter counts, and often results in modality-specific failure modes. ...

Here’s a genuinely clever piece of engineering: what if instead of building a whole new coordination infrastructure for multi-agent systems, you just used Git? That’s the core idea behind h5i v0.1.5, released today by Hideaki Takahashi. The new Agent Radio feature (h5i msg) turns your existing Git repository into an asynchronous message bus for AI coding agents — specifically demonstrated with Claude Code and OpenAI Codex communicating with each other in real time, no external server required. ...

If you’ve been waiting for the US open-weight model scene to close the gap with the frontier, today is a significant day. NVIDIA dropped Nemotron 3 Ultra 550B — a 550-billion-parameter Mixture-of-Experts model with 55B active parameters — live on Hugging Face and NVIDIA NIM, and it’s the most intelligent US-based open-weights model available right now. What Just Happened NVIDIA CEO Jensen Huang announced Nemotron 3 Ultra during Computex 2026, and the team wasted no time getting weights into the wild. As of June 4, both BF16 and NVFP4 quantized variants are live under the nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4 namespace on Hugging Face. Full weights, training recipes, and datasets have been released under a permissive license. ...

OpenAI’s Codex started as a tool for software developers. That framing is now officially obsolete. On June 2, 2026, OpenAI announced a major Codex expansion: six role-specific plugins, a new Sites feature for creating and deploying interactive web apps, and Annotations for in-place document refinement. The headline: more than 5 million people use Codex weekly, and 20% of them aren’t developers — and that non-developer segment is growing more than 3x faster than the developer segment. ...

Researchers have published what might be the most important AI security paper of 2026 so far: a proof-of-concept worm that doesn’t just use AI to find vulnerabilities — it is AI, running autonomously, propagating through networks, and harvesting compute from its victims to sustain itself. The paper is arXiv:2606.03811, from researchers at the University of Toronto’s CleverHans Lab (Prof. Nicolas Papernot’s lab), the Vector Institute, and the University of Cambridge. It was responsibly disclosed before publication. ...

Anthropic has formalized the enterprise consulting side of its Claude ecosystem. Three months after launching the Claude Partner Network with a $100 million investment commitment, the company today introduced a structured Services Track and a Partner Hub portal that brings the program into a recognizable enterprise channel partner model. The Services Track: Three Tiers The Services Track is structured around three tiers for consulting and systems integration firms deploying Claude in enterprise contexts: ...

If you’re running multi-agent Claude Code pipelines in production, today’s v2.1.162 release has two changes that directly address operational pain points: you can now see why a session is waiting, and a failed tool call in a parallel batch no longer takes its siblings down with it. waitingFor in Agent JSON Output Running claude agents --json now returns a waitingFor field for sessions that are blocked. This field shows what the session is waiting on — for example, a pending permission prompt, a human approval request, or an upstream dependency. ...

The exec approval problem in autonomous agent systems has always had the same uncomfortable shape: either you interrupt the agent constantly with approval prompts (safe, annoying, productivity-killing), or you give it free rein to run whatever it wants (fast, terrifying in production). OpenClaw’s new auto exec mode, announced May 31, introduces a third option — and it’s a genuinely useful one. The Three Tiers Auto mode implements a three-tier decision pipeline for exec approval: ...

The threat model for AI agent skills is not the same as the threat model for traditional software packages. And the security tools catching up to that gap are finally starting to look like they belong in the same category. OpenClaw announced today that every ClawHub skill now ships with a Skill Card and is scanned by SkillSpector — an NVIDIA-developed tool built specifically to detect AI-native attack vectors inside agent skill files. ...

Something subtle but significant landed in OpenClaw v2026.6.2 today: agents can no longer write skills directly into production. Instead, they propose them — and humans decide whether those proposals become part of how the agent works from that point forward. That one change reframes a whole category of risk in autonomous agent systems. What Is Skill Workshop? Skills in OpenClaw are reusable procedural instructions agents load and execute. A skill might describe how to do invoice follow-up, how to run a repo health check, or how to process a specific kind of request. Once a skill is active, it shapes every future run that loads it. ...