Friday, July 3, 2026

Fable harness optimization reportedly matches Sonnet 4.6 performance at 7× cost reduction on legal tasks via improved prompting harness. Source

Z.ai launches ZCode. an agentic coding IDE built on GLM-5.2, entering a crowded field against Cursor, Claude Code, and GitHub Copilot. Source

audio.cpp brings native C++/GGML music generation. support for ACE-Step, HeartMuLa, Stable Audio 3, RoFormer, and HTDemucs in a single stack, with claims of 10-minute music in 60 seconds. Source

Self-hosted AI gateway. MIT-licensed open-source project offering one endpoint for 237 providers (90+ free) with auto-fallback and token compression. Source

WorldDirector. a controllable world simulator with persistent dynamic memory aimed at long-horizon generation. Source

Multi-Resolution Flow Matching. training-free staged sampling to accelerate diffusion/flow models without retraining. Source

Hierarchos (232M). a recurrent memory-augmented model combining RWKV with hierarchical loops and slot-based memory, reporting stable training. Source

Reasoning effort drives agentic reliability. observational study finds reasoning effort, not tool access, is the key lever for first-try code generation success. Source

OpenAI reportedly releases GPT-5.6, including a 'GPT Soul' model claimed to match Claude Mythos Preview at 3× lower cost. Source

Anthropic in talks with Samsung on a custom AI chip, signaling deeper vertical integration into compute hardware. Source

OpenAI in talks to grant the Trump administration a 5% equity stake, per FT report. Source

Zuckerberg tells staff AI agents have not progressed as quickly as hoped. a rare candid internal signal from Meta on agentic AI pace. Source

AI #175: The Fable Continues. Zvi Mowshowitz's weekly digest covering Fable 5 re-deployment and broader AI developments. Source

Using DSPy to evaluate Datasette Agent's SQL prompts. Simon Willison walks through systematic prompt optimization with DSPy. Source

VulcanBench v2 launches. 10 real-world tasks drawn from merged PRs, built using Fable to combat benchmark saturation. Source

RAG benchmark on synthetic healthcare data. finds that document shaping beats query rewriting and reranking for most retrieval gains. Source