News Flash: Anthropic says its AI now does, Toward Recursive Self-Improvement: How GLM Built, & 3 More Architect Dispatches — What Changes Today?

1. Anthropic's AI Now Conducts 26% of Its Own Research

The News Highlight:

Anthropic has revealed that its Claude models now lead 26% of the company's internal AI research work, overseeing tens of thousands of active internal agents. To provide transparency into this rapid evolution, Anthropic has released new measurement frameworks that track how much AI is contributing to building the next generation of models, the effectiveness of human oversight over these agents, and the compute resources driving this automated R&D.

My Analysis:

As a Solutions Consultant, I constantly evaluate how enterprises can scale their engineering efforts, but this is a paradigm shift. We are officially entering the era of recursive self-improvement. Anthropic using Claude to build the next Claude isn't just a neat trick; it's a fundamental change in how we calculate R&D velocity. For enterprise architectures, the most critical signal here is the oversight measurement. As we deploy multi-agent systems in corporate environments, proving that humans can still effectively monitor and govern tens of thousands of autonomous actions will be the primary hurdle for compliance and security teams.

• Primary Source: Anthropic says its AI now does a quarter of its research work

2. Recursive Self-Improvement in Action: GLM Builds Its Own Inference Stack

The News Highlight:

Z.ai successfully utilized a GLM-5.3-powered "Infra Agent" to construct the production serving stack for GLM-5.3-Flash across more than 100,000 Chinese accelerators in under two weeks. By leveraging dense feedback loops and kernel-level fixes, the AI agent tripled system throughput while humans remained strictly in charge of setting objectives and managing risk.

My Analysis:

Following right on the heels of Anthropic's announcement, this is the practical, infrastructure-level application of AI building AI. In my architecture reviews, optimizing hardware utilization—especially across heterogeneous or massive accelerator clusters—is usually a painstaking, human-intensive MLOps bottleneck. Seeing an AI agent handle system-level optimization and kernel fixes to triple throughput in just two weeks is staggering. For cloud architects and DevOps teams, this signals that "Infrastructure as Code" is rapidly evolving into "Infrastructure by AI." We need to start designing our deployment pipelines to accept optimization pull requests directly from infra-agents.

• Primary Source: Toward Recursive Self-Improvement: How GLM Built Its Own Inference Infrastructure

3. Bonsai 2 27B Achieves Near-Lossless Compression in a 5.9GB Footprint

The News Highlight:

PrismML has launched Ternary Bonsai 2 27B, a multimodal model that brings advanced reasoning, coding, and agentic capabilities into a remarkably small package. By utilizing ternary {-1, 0, +1} weights with FP16 group-wise scaling, the model achieves 1.76 effective bits per weight, shrinking a 27B parameter model to a mere 5.9GB footprint. It supports a massive 262K-token context window and runs efficiently on local devices, including Apple hardware via MLX and Nvidia GPUs via CUDA.

My Analysis:

This is a massive breakthrough for Edge AI. When I talk to enterprise clients about deploying LLMs, data privacy and cloud inference costs are always the top concerns. Fitting a highly capable 27-billion parameter model into less than 6GB of VRAM completely changes the deployment calculus. It means we can now run enterprise-grade, agentic AI locally on standard corporate laptops or edge servers without sacrificing context length (262K is huge for local). I highly recommend developers start testing ternary quantization frameworks; this is the future of privacy-first, on-device enterprise AI.

• Primary Source: Introducing Bonsai 2 27B: Near-Lossless Compression in a 9x Smaller Footprint

4. Figure's Helix 2.5 Demonstrates Zero-Shot Humanoid Robotics in 30 Homes

The News Highlight:

Figure has introduced Helix 2.5, a new humanoid control model pre-trained on their Index dataset. In a major test of generalization, Figure deployed their robots to 30 rented homes in the Bay Area. Without collecting any prior training data or mapping in those specific environments, the robots successfully performed zero-shot physical tasks, including tidying rooms, folding towels, and making beds.

My Analysis:

In the realm of embodied AI, the transition from rigid, pre-programmed factory floors to dynamic, unstructured environments (like a home) is the ultimate test. Zero-shot physical execution means the model is actually generalizing spatial reasoning and object manipulation, rather than just memorizing a space. From a commercial standpoint, this drastically reduces the time-to-value for robotics. If a robot doesn't need a multi-week site-mapping and training phase before it can start working, the ROI for physical automation in logistics, healthcare, and eventually consumer markets just accelerated by years.

• Primary Source: Helix 2.5

5. Agora: Git as Shared Memory for Autonomous AI Agents

The News Highlight:

Researchers have released Agora, a framework that utilizes an append-only Git Directed Acyclic Graph (DAG) as shared memory for autonomous AI research agents. This system allows agents to share hypotheses, experimental results, and verifications across separate sessions, turning complex multi-agent workflows into immutable, reproducible, and searchable Git commits.

My Analysis:

This is one of the most elegant architectural patterns I've seen for multi-agent systems this year. One of the biggest challenges I face when designing agentic workflows is state management and context loss over long-running tasks. By mapping agent memory and decision trees directly onto a Git DAG, Agora solves this beautifully. It provides built-in version control for AI thoughts, allows agents to branch off neglected hypotheses, and gives human overseers a transparent, auditable trail of how an AI reached a conclusion. If you are building multi-agent systems, adopting a Git-backed state machine should be on your immediate roadmap.

• Primary Source: Git as Shared Memory for AI Research Agents

Morning Executive Comparison Matrix

Dispatch	Core Domain	Production Maturity	My Recommendation
Anthropic AI R&D	AI Self-Development	Early Enterprise Adoption	Implement robust agent oversight and governance metrics immediately.
GLM Infra Agent	MLOps & Infrastructure	Production Proven	Explore AI-driven infrastructure optimization for large-scale clusters.
Bonsai 2 27B	Edge AI & Quantization	Ready for Edge	Test ternary models for local, privacy-sensitive enterprise workloads.
Helix 2.5	Embodied AI	R&D / Pilot	Monitor zero-shot spatial generalization for future physical automation.
Agora (Git for Agents)	Multi-Agent Architecture	Developer Preview	Adopt Git-based state management for complex, long-running agent workflows.