EN RU

Daily digest

7 items · ~7 min · Week 2026-W30

Worth knowing (1)

RESOURCE2SKILL: Distilling Executable Agent Skills from Human-Created Multimodal Resources

Microsoft Research
Research official 2 src. ~1 min

Microsoft Research proposes a method to distill reusable, executable agent skills from human-created multimodal resources such as tutorial videos, organizing them into a hierarchical wiki-based skill library that agents can retrieve, inspect, and compose across software domains like slide design, spreadsheets, CAD, and music production.

Why it matters
Reached 121 upvotes on HuggingFace Daily Papers (2026-07-20); reports +11.9pp average improvement over skill-free agent baselines across seven authoring benchmarks.
#agents #skill-learning #multimodal
For reference (6)

Musk says xAI's 2-trillion-parameter Grok 4.6 has finished initial training

xAI
Industry media only 2 src. ~1 min

Elon Musk announced that xAI's next large model, expected to be called Grok 4.6 and built with roughly 2 trillion parameters, completed its initial pre-training run the week of July 20, 2026, days after Moonshot AI released the 2.8-trillion-parameter Kimi K3.

Why it matters
Marks xAI's direct response to Chinese open-weight frontier models like Kimi K3, though post-training, safety evals and red-teaming reportedly still take xAI 4-6 weeks, so no near-term public release is confirmed yet.
#grok #frontier-model #training #competition

Anthropic opens applications for AI for Science rare disease research grants

Anthropic
Industry official 1 src. ~1 min

Anthropic is accepting applications through August 2, 2026 for its AI for Science program's new rare disease track, offering up to $50,000 in Claude credits over six months across a clinical-research track (with partners like the Monarch Initiative) and a biotechnology track for drug development.

Why it matters
Extends Anthropic's Claude Science initiative into a funded grant program, signaling a deeper push into applied scientific/biomedical use cases beyond general-purpose assistant products.
#grants #healthcare #rare-disease #claude-science

Understanding Reasoning from Pretraining to Post-Training

Research official 1 src. ~1 min

Using chess as a controlled testbed spanning the full pretraining-to-RL pipeline, the authors show that RL post-training does not simply sharpen the supervised policy: on easy problems it amplifies moves the model already preferred, while on hard problems it surfaces correct moves nearly absent after SFT, and longer-pretrained checkpoints benefit more from RL.

Why it matters
Featured on HuggingFace Daily Papers 2026-07-20 with 20 upvotes; gives a mechanistic account of what RL post-training actually changes in a model's policy, connecting pretraining scale to RL returns.
#reasoning #rl #pretraining

Recursive Harness Self-Improvement (RHI): Refining Agent Harnesses from Execution Feedback

Sakana AI
Research official 2 src. ~1 min

Sakana AI and UC Berkeley researchers introduce Recursive Harness Self-Improvement, an iterative algorithm that refines user-constructed agent harnesses using pairwise feedback drawn from the agent's own execution history, without retraining the underlying model.

Why it matters
Extends Sakana AI's recently launched Recursive Self-Improvement research program, part of a broader industry push toward AI systems that improve their own tooling rather than relying solely on more compute.
#agents #self-improvement #reasoning

Claude Code v2.1.216 adds filesystem sandbox opt-out and fixes worktree git isolation

Anthropic
Tools official 1 src. ~1 min

Claude Code v2.1.216, released July 20, 2026, adds a sandbox.filesystem.disabled setting to skip filesystem isolation while keeping network egress control, fixes a quadratic message-normalization slowdown in long sessions, and patches several worktree-isolation bugs where subagents could redirect git operations into a shared checkout.

Why it matters
The worktree git-isolation bug could let a subagent's git commands leak into the wrong repository checkout, and the sandbox opt-out gives teams finer control over filesystem vs. network isolation.
#claude-code #cli #coding-agent #sandbox #bug-fix

OpenCode v1.18.4 adds adaptive thinking controls for Kimi models

SST (OpenCode)
Tools official 1 src. ~1 min

OpenCode v1.18.4, released July 20, 2026, adds adaptive thinking controls for Kimi models on Anthropic-compatible providers, synchronizes the desktop app's embedded terminal theme with the overall app theme, reduces OpenAI provider header timeouts, and restores Azure Cognitive Services support.

Why it matters
Incremental but steady progress for the open-source OpenCode coding agent, expanding provider/model compatibility that developers rely on for multi-provider agent workflows.
#opencode #cli #coding-agent #kimi