Anthropic Hosts "Code with Claude" in Tokyo; RHO Method Lets LLM Agents Self-Improve from Past Failures; The Frontiers of CPU Inference

Anthropic Hosts "Code with Claude" in Tokyo — Three Guidelines for Developers

On June 10, Anthropic held "Code with Claude," a developer event in Tokyo. With its new model "Claude Fable 5" — which entered general availability (GA) that same day — front and center, the event offered guidelines for incorporating high-performance AI into services.

According to ITmedia AI+, the event presented three key guidelines for building AI-powered services. The release of Claude Fable 5 means the model, previously available only by invitation, is now broadly accessible — a step that has effectively lowered the implementation barrier for Japan's developer community.

That Anthropic held an event in Tokyo itself, rather than only overseas, is seen as a sign of how much the company values the Japanese market. The company also recently drew industry attention with the debate over Claude consciousness, making its commitment to improving the developer experience increasingly clear.

RHO: LLM Agents That Self-Improve from Past Failures

"RHO (Retrospective Harness Optimization)," featured on Hugging Face Daily Papers, is a method that improves an LLM agent's harness (skills, tools, and workflows) using nothing but the agent's own past trajectories. Its defining characteristic is that it requires no external ground-truth dataset at all.

Concretely, it uses a DPP (Determinantal Point Process) to select a coreset of varied difficulty from past tasks, then re-runs each task in parallel. Failures are diagnosed through self-verification and self-consistency, and candidate harness updates are evaluated via pairwise self-preference.

The results are striking: a single optimization round lifted the SWE-Bench Pro pass rate from 59% to 78%. Consistent improvements were also confirmed on Terminal-Bench 2 and GAIA-2. Reaching this level of improvement without any external grading points to the promise of agent self-improvement.

MemDreamer: Separating Perception and Reasoning in Long-Video Understanding

Also featured on Hugging Face Daily Papers, "MemDreamer" proposes a new approach that separates perception from reasoning in long-video understanding. By combining hierarchical graph memory with an agentic retrieval mechanism, it reportedly sidesteps context-length limits and eases attention dilution.

Long-video understanding is a problem domain where the context window is often too small to process an entire video at once. This separation — structuring information at the perception stage and retrieving only what is needed at the reasoning stage — is gaining attention as a scaling direction for multimodal understanding going forward.

The Frontiers of CPU Inference: MoE Models Take Center Stage

On Reddit's LocalLLaMA community, a lively discussion unfolded around LLM inference on CPUs. One user reported running the Q4_K_M quantization of Qwen3.6 35B A3B with llama.cpp on an Intel Core Ultra 7 165H (64GB RAM) — a chip with AVX2 but no AVX512 support — and getting roughly 10 tokens per second.

MoE (Mixture of Experts) models activate only a fraction of their total parameters, which makes them an excellent fit for CPU environments where memory bandwidth and compute resources are limited. The strategy of using 64GB of RAM to host large MoE models has become a practical option for users without a GPU.

That said, some point out that 10 tps falls short of practicality in thinking mode, so challenges remain for CPU inference. Community members continue to exchange information on other models, quantization settings, and forks of llama.cpp.

Automating "Work Logs" in Development with Coding Agents

On Qiita, a developer shared their experience putting automated "work log generation" into practice for development with coding agents.

Tell a coding agent "I want to implement authentication" and it will move fast, from proposing a file structure through implementation and testing — but "why this design" and "which trade-offs were chosen" rarely leave a record. Nuances that colleagues would fill in verbally during human-to-human review tend to be lost as tacit knowledge in AI-assisted development.

To address this, a practice has been reported that builds automated work log generation into the development process to record and reuse the rationale behind design decisions. As AI-assisted development accelerates, how to design the recording and retrospective review of the development process itself is becoming a key issue that can make or break team productivity.

---