Daily digest

9 items · ~9 min · Week 2026-W31

Must-read (1)

HiFi-UMI: Learning Deployable Manipulation Policies from High-Fidelity UMI Data Alone

Simple AI (Simple World Lab)
Research official + media 2 src. ~1 min

HiFi-UMI is a portable, robot-free data-collection rig (head-mounted stereo-inertial SLAM plus synchronized wide-FoV cameras) that raises the fidelity of Universal Manipulation Interface demonstrations enough that policies trained solely on this data deploy directly to real robots, matching in-domain teleoperation across three VLA backbones. The team open-sourced HiFi-UMI-2K, 2,000 hours of microsecond-synchronized demonstrations.

Why it matters
Topped HuggingFace Daily Papers for July 29, 2026 with 137 upvotes, the highest vote count in the window; removes the need for costly real-robot teleoperation data collection at scale.

Worth knowing (5)

1,100+ Employees at OpenAI, Anthropic, Google, and Meta Sign "Pacing the Frontier" Letter

Industry official + media 3 src. ~1 min

More than 1,100 employees across OpenAI, Anthropic, Google, and Meta — including Dario Amodei, OpenAI chief scientist Jakub Pachocki, and Google's Anca Dragan — published "Pacing the Frontier," asking the US government to help build international technical and governance tools for a coordinated, verifiable slowdown of automated AI development. OpenAI and Anthropic subsequently endorsed the statement as companies.

Why it matters
First cross-lab employee-driven push (rather than company PR) for a government-backed pacing mechanism, and it landed with two of the frontier labs endorsing it institutionally within a day.

Yandex launches "75/75/75" program to push AI-native software development

Yandex
Industry official + media 3 src. ~1 min

Yandex announced the "75/75/75" initiative: by the end of 2026 it wants at least 75% of its engineers regularly using AI to write code, AI involved in preparing at least 75% of code changes company-wide, and each such change to be at least 75% AI-generated. The push leans on Yandex's internal coding agent GENA, already connected to over a thousand projects, and cites cases like an ~80% AI-generated video-translation rewrite finished in two days instead of two to three weeks.

Why it matters
One of the largest Russian tech employers is formalizing company-wide AI-coding adoption targets and opening the underlying tooling to external developers, signaling how fast agentic coding is being institutionalized outside the US/China axis.

Google DeepMind Disbands Dedicated AlphaFold Team, Shifts Researchers to Gemini

Google DeepMind
Industry media only 3 src. ~1 min

The Financial Times reported on July 29 that Google DeepMind dissolved the dedicated team behind its Nobel Prize-winning AlphaFold protein-structure system over the past year, reassigning most researchers to Gemini and other scientific programs (enzyme design, nuclear fusion, genomics). DeepMind confirmed the personnel moves to press; some AlphaFold co-authors, including lead author John Jumper, have left for Anthropic.

Why it matters
Signals DeepMind is folding its landmark single-purpose science model into a Gemini-centric, multi-agent research strategy, and marks a notable talent outflow toward Anthropic.

A New Role for Relevance: Guiding Corpus Interaction in Agentic Search

Tencent
Research official + media 2 src. ~1 min

Tencent researchers propose RARG (Relevance-Aware RipGrep Search Agent), which uses relevance scores not just to pick top documents but as an execution prior that orders search, seeds entry points, and reranks excerpts during multi-step corpus interaction. It outperforms existing retrieval and interaction-based search agents on QA and reasoning benchmarks.

Why it matters
Reached 86 upvotes on HuggingFace Daily Papers (July 29, 2026); addresses a real bottleneck in agentic RAG systems where naive top-k relevance filtering fails on complex, multi-hop questions.

HumanCLAW: Can Vision-Language Models Act Through a Body?

Meta Research
Research official + media 2 src. ~1 min

Meta introduces HumanCLAW-Bench, 1,218 long-horizon egocentric find-navigate-interact episodes across 41 indoor scenes, and a framework that separates decision-making from motor execution to isolate a VLM's 'action intelligence.' The best of nine tested state-of-the-art models achieved only 16.8% success, with failures traced to lost embodied self-awareness rather than poor target recognition.

Why it matters
Featured on HuggingFace Daily Papers (July 30, 2026, 43 upvotes); a stark benchmark result showing current VLMs struggle to track their own body state during embodied tasks, relevant to the broader agentic/robotics push across labs.
For reference (3)

Pass the Baton: Trajectory-Relayed On-Policy Distillation

Zhejiang University
Research official + media 2 src. ~1 min

A new on-policy distillation method where a teacher and student model relay segments of a reasoning trajectory to each other, aiming to keep the student's training distribution closer to what it will see at inference than standard full-trajectory distillation.

Why it matters
Landed on HuggingFace Daily Papers (July 29, 2026, 27 upvotes); on-policy distillation is an active area for cheaply transferring reasoning ability from large to small models.

GitHub Copilot CLI 1.0.76 adds sessions sidebar, enterprise sandbox floor

GitHub
Tools official 1 src. ~1 min

GitHub Copilot CLI v1.0.76 (July 30, 2026) adds a new Sessions sidebar for managing multiple concurrent sessions, lets enterprise administrators enforce a restrictive sandbox floor that managed settings can tighten but never loosen, adds grok-4.5 model support, enforces sandbox denied-path rules for relative and symlinked entries on macOS/Linux, and caps hook output at 10 MiB to prevent a memory-exhaustion vector from unbounded hook responses.

Why it matters
The enterprise-enforced sandbox floor and hardened hook-output bound are notable security/governance additions as Copilot CLI competes with Codex CLI and Claude Code for autonomous-agent trust in managed environments.

OpenCode 1.18.6-1.18.9 fixes MCP compatibility and desktop stability

SST
Tools official 2 src. ~1 min

Across four releases from July 27-28, 2026 (v1.18.6 through v1.18.9), OpenCode restored compatibility with legacy MCP SDK clients, fixed MCP servers failing to reconnect after expired SDK sessions (including under concurrent requests), honored configured MCP OAuth callback ports in mcp debug, and fixed several desktop-app issues including branch-specific repository cache corruption and a Solid.js cleanup crash breaking navigation.

Why it matters
The rapid cadence of MCP-compatibility fixes reflects churn as the broader MCP ecosystem adapts around the new spec landing the same week.