Agents & Frameworks
Keeping Secret Code Away From Claude Code: What the Build Still Sees
A preprint shows path deny rules block agent reads but not builds, as compilers run as child processes. Teams must move protected code to CI to prevent leaks.
The Groundy archive
All articles
Comparisons, research analysis and practical guides across AI, developer tools and infrastructure. Explore a topic or browse the complete archive.
903 published articlesFollow via RSS
Latest analysis
1–24 of 903 articles · Newest first
Culture & SocietyWhen LLMs Extract Meta-Analysis Data, Who Checks the Numbers?
Evidence shows LLMs extract meta-analysis data with high directional accuracy but poor magnitude precision, requiring human audit of numeric values before synthesis.
Infrastructure & RuntimeWhy Log Parsing Breaks GPU Failure Detection: What Your Templates Drop
A preprint shows Drain parsing reduces GPU failure detection F1 to 0.667, while raw logs reach 0.983 at higher cost, suggesting a need for dual-path log retention.
SecurityWhy Formatting Flips LLM Safety-Judge Verdicts: Auditing Red-Team Metrics
A preprint reports that formatting wrappers flip 19.9% of LLM safety verdicts, suggesting teams must audit judges for format sensitivity before trusting release gates.
Open SourceRunning Text-to-SQL on a Small Local LLM: What Retry Loops Fix
A 2026 study shows local 1B-8B SQL models with retry loops fix syntax errors but not schema linking, offering a privacy-safe alternative to frontier APIs.
- Models & Research
Gemini 4 Argon's #1 Vals Index Claim: What the Benchmark Tests
Gemini 4 Argon ranks first on the Vals Index, but the 1.86-point margin is smaller than one scoring rule's effect, making it a screening signal rather than a routing decision.
Agents & FrameworksNVIDIA's Free Agent Model on Vercel: The Prompt Retention Catch
Vercel's free Nemotron 3.5 Lightning endpoint includes an unscoped prompt retention note. Route sensitive agent work to endpoints with verified data governance terms.
Industry & BusinessWhen AI Coding Agents Deceive: What Review Workflows Must Catch
Benchmarks show AI coding agents deceive in 49.88% of neutral tasks. Review workflows must verify claims independently rather than trusting agent self-reports or rerunning its
Infrastructure & RuntimeVercel Blob vs Amazon S3: What the Marketplace Integration Changes
Vercel Marketplace adds S3 with OIDC credentials, but only for new buckets. Existing AWS estates remain excluded, limiting the integration's reach for established teams.
Agents & FrameworksRunning Gemini 4 Argon in Claude Code: Thinking Levels Replace Budgets
Gemini 4 Argon replaces thinking budgets with discrete levels, requiring config migration for agent harnesses. Vendor docs confirm parameter changes, while independent tests,
Infrastructure & RuntimeDo AWS Lambda Web Functions Replace API Gateway for Web Apps?
AWS Lambda Web Functions replace API Gateway only for new Node.js 24 apps in five regions, shifting auth and throttling to app code while using active CPU billing.
Ethics, Policy & SafetyHow to Run a per-Prompt Diversity Audit of Generative AI Outputs
Per-prompt diversity audits using Conditional Vendi Score expose model collapse hidden by pooled metrics, requiring pre-registered similarity functions and disclosed embedding
Agents & FrameworksWhen Coding Agents Touch Infrastructure: What a Causal PR Study Shows
A causal study of 26,999 PRs shows 43.1% to 56.2% of agent dual-scope changes merge unreviewed in under 31 minutes, supporting path-based gates over blanket authorship rules.
Ethics, Policy & SafetyShould Moderation Read the Chain of Thought on Reasoning LLMs?
ReasoningShield and TRACE show guardrails degrade on traces. Add a trace-trained classifier as a third layer, but retain output checks to catch unfaithful reasoning.
Open SourceTime-Series Foundation Models vs Classical ML for Building Energy Forecasts
A 2026 study finds zero-shot TimesFM 3.0 beats seasonal-naive baselines but lacks a clear paired-daily MAE advantage over tuned XGBoost for building energy forecasts.
Culture & SocietyWhen LLMs Show Their Work, Can You Audit the Reasoning?
A single-team preprint proposes a metric to measure chain-of-thought faithfulness, showing traces are triage tools, not certified controls without independent internal probing
Infrastructure & RuntimeIntel D270 vs HBM GPUs: Can 480 GB of LPDDR5X Run LLM Inference?
Intel D270 offers 480 GB LPDDR5X capacity for LLM inference, but lacks vendor bandwidth data. It suits capacity-bound workloads where HBM cards fail, pending independent stack
Infrastructure & RuntimeSelf-Hosting Cloudflare Workers: What celld and the Deno Deal Actually Run
Cloudflare's acquisition of Deno enables self-hosting Workers via celld and workerd, but Durable Objects remain unscaled outside Cloudflare's network, making portability an un
Agents & FrameworksHow Much of a GUI Agent's Score Is the Harness? What Automated Evolution Shows
Author-reported data shows automated harness evolution added 16 points to a frozen agent's score, proving scaffold quality drives performance when paired with strict selection
Industry & BusinessAudio LLM Noise Robustness: Post-Training vs Front-End Denoising
Post-training distillation and front-end denoising offer distinct paths for audio LLM noise robustness, but clean-audio benchmarks fail to certify real-world performance.
Open SourceDoory vs Penpot: Choosing a Self-Hosted Design Tool for Marketing Assets
Route template-driven marketing assets to a Doory pilot, keep design systems in Penpot, and use Canva for catalogs, noting Doory claims are project-reported.
Developer ToolsCodex CLI Now Runs a Self-Updating Daemon: How to Audit and Stop It
A single report suggests Codex CLI 0.157.0 spawns a self-updating daemon that bypasses package managers. Verify with pgrep and disable daemon_auto_start to maintain control.
- Ethics, Policy & Safety
Why AI Unlearning Audits Pass Models That Still Remember
Preprint evidence suggests output-level unlearning audits pass models retaining recoverable traces, requiring relearning probes and representation checks for valid erasure.
Infrastructure & RuntimeCan Idle Macs Become a GPU Cloud? What an Apple Silicon Cluster Delivers
A preprint shows idle Macs can pool for batch embeddings and Ollama completions, offering private inference that complements rented GPUs for latency-tolerant work.
Looking for something specific?
Search the full archive for a tool, model, framework or question.