Jul 31, 2026

Thinking Machines released Inkling-Small, a 276-billion-parameter open-source AI model that outperforms its larger predecessor on several benchmarks, cutting compute requirements and cost for enterprises. Meanwhile, OpenAI engineers detailed the multi-layered agent loop behind Codex and ChatGPT Work, showing how a harness, API, and inference layer work together and how reducing duplicated work across iterations can significantly lower latency and cost.

NVIDIA’s AI Red Team identified four universal failure modes after six months of assessing AI agents: missing access controls, tools that allow arbitrary code execution, absent network egress restrictions, and plaintext secrets. Their tests showed that even with LLM-based review or allowlists, command-line tools remain high-risk when attacker-influenced input can trigger malicious execution.

In performance engineering, an engineer running a CPU-first llama.cpp server found that reformatting prompts outperformed coarser quantization: Q4_0 gave a 37% speedup but caused factual errors in 25% of answers, while prompt reformatting preserved accuracy and boosted prefill speed. GitHub’s Blackbird code search team achieved a 15x speedup in case-folding by removing an ASCII fast-path optimization that accidentally blocked compiler vectorization. A detailed breakdown of Docker build caching revealed three distinct caches with separate invalidation rules, and showed that manifest-first layer ordering drastically reduces rebuild times when source code changes.

On the robotics front, OpenNeoData published the largest tactile-robotics dataset yet: over 200,000 trajectories and 5,000 hours of visuotactile data across six embodiments, accompanied by two models pretrained on tactile input at scale. Other teams sidestepped expensive sensors entirely—one infers contact fields from 500,000 hours of human video, while others use affordable hardware to make tactile sensing more accessible.

In creative defenses against AI scraping, ShieldFont introduced an open-source web font that uses OpenType GSUB rules to serve human readers the original text while giving scrapers a semantically different but fluent decoy, raising the collective cost of large-scale ingestion. In hardware hacking, an engineer reverse-engineered the Attack Shark X86 keyboard protocol by patching the vendor’s Electron app to log HID reports, reconstructing the full 64-byte feature report protocol and enabling configuration from Linux.

Finally, ZeroFS shipped a native Rust kernel module that replaces its FUSE-based client, speaking the 9P2000.L.Z protocol directly and integrating with netfslib to eliminate userspace round trips for network filesystem operations.