The Groundy archive · Page 3 of 43
Browse Groundy's complete archive of 1027 articles on AI, developer tools and infrastructure. Page 3 of 43.
49–72 of 1027 articles · Newest first
Meta reports GPU-based ad retrieval at one billion ads per second, but independent research shows GPU vector search only pays off when the index is resident on the accelerator
Vercel Sandbox mounts Box folders using a separate sync user to keep credentials out of the agent context, a vendor-stated design that requires independent verification before
Self-reported preprint data shows cross-layer Java and PHP deserialization defenses cut attack success to 6.2%, outperforming single-layer baselines that leaked up to 55%.
Preprint evidence suggests computer-use agents misclick due to input resolution limits. Adaptive zooming and verifier abstention improve grounding accuracy without changing or
A Hausa preprint shows alignment techniques fail to balance safety and fluency, requiring native review and per-language evidence for launch.
Simulation evidence shows priority dispatch for premium riders conditionally slows others, with harm scaling by demand regime and member share, requiring distribution-level, 1
A preprint proposes compiling subtask state to reduce coding agent drift, reporting improved scores and lower costs on specific benchmarks compared to native harnesses.
VLA-HWBench shows action decoding strategy, not model size, determines latency on Jetson and AMD platforms, with flow-matching heads outperforming token-by-token decoders.
A study of 120 agent defects shows 90% of prompt caching failures are silent. Auditing usage fields and prefix stability is required to verify discount eligibility.
Vercel's AI SDK HarnessAgent now accepts image inputs across ten adapters, but independent evidence on vision quality is absent, requiring per-adapter validation.
Preprint evidence shows pooled CoT monitor accuracy hides reasoning-dependent fragility, where rewrites collapse catch rates while aggregate scores remain stable.
A preprint shows folding MCP usage examples into tool indices lowers recall for unused tools. Measure held-out Hit@5 before adopting this index change.
OpenAI's GPT-6.1 Terra targets high-volume document and desktop tasks, but vendor-reported claims lack benchmarks, requiring per-document-type validation before routing.
A preprint reports that renaming variables in third-party Python files steered six AI coding agents toward four CWE classes, with self-reported success rates of 22.5% to 56.0%
Claude Code, Cursor, and Codex persist instructions in files that agents follow at high rates. Stale entries survive weeks, so developers must audit these files to prevent mis
Grossman's reanalysis shows Relativity's GenAI review is comparable to, not superior to, TAR. Defensible eDiscovery requires validated CAL workflows and strict control-sampled
A registry census found 2.3% of Ollama GGUF tags are broken. Use quantcheck to screen low-bit quants before deployment to avoid serving non-functional models.
Cloudflare reports a GLM-5.3 KV cache leak caused by pool misconfiguration and an SGLang bug, affecting 155 responses. Operators must audit pool sizing against context length.
A 2026 census of 22,854 repos shows AGENTS.md adoption is high, but evaluation studies find overviews unhelpful and stale files override code, favoring imperative commands.
php2shell parser bugs require runtime patching; WAF virtual patches offer only temporary containment. Cloudflare and ELLIO confirm edge rules cannot fix the runtime bug.
TouchSafeBench shows VLMs score 56% on collision grounding, 18 points below human experts. Keep certification in geometric layers and treat VLMs as advisory context only.
Cuffless PPG blood pressure claims require prospective validation with repeated references, subject-level splits, and calibration stability to meet AAMI standards.
MCPHunt preprint shows multi-server MCP agents leak data across boundaries via benign tasks. Seed markers and trace write calls to detect propagation that per-server reviews.
PaperPilot offers a local Rust PDF engine via MCP for merge, split, OCR, and compress. It keeps data on-device but requires manual OCR fidelity testing for scanned documents.