The Groundy archive · Page 6 of 38
Browse Groundy's complete archive of 903 articles on AI, developer tools and infrastructure. Page 6 of 38.
121–144 of 903 articles · Newest first
EvoUndo preprint argues coding agents need provable undo limits for self-rewrites, as 197 improving mutations failed recovery checks in author-reported tests.
A preprint claims composing specialist capabilities into one small model improves accuracy and cuts tokens, but results are author-reported and unreplicated.
A preprint reports that calibrated learned routing for disaggregated LLM serving beats heuristics by 1.7 to 2.9 goodput points, but only on specific hardware with large pools.
A preprint reports API evaluations score 3.4 points higher than chatbot interfaces, suggesting audits must test deployed products directly rather than relying on model metrics
A preprint reports LLM inference energy varies up to 179x by language, with Pashto costing far more than English. These author-reported findings suggest locale mix is a key, 1
A 2026 paper defines Semantic Confusion, where LLM refusals flip on meaning-preserving paraphrases. Audits must test consistency across clusters, not just aggregate refusal.
Author-reported DECODE data shows most AI code edits occur within 15 minutes of acceptance, with 31% of trajectories ending in removal, suggesting review should focus on the 1
MCPAgentBench shows LLM agents excel at tool selection but fail strict execution order, suggesting teams should route agents by measured MCP skill rather than general chat.
DuckDB research extension FlockMTL integrates LLMs and RAG into SQL via schema objects, moving governance to DDL, though performance claims remain author-reported.
A rigor-matched audit finds layer skipping speeds LLM inference, but wall-clock rankings reverse when decision overhead is separated from pure generation cost.
A 441-repo preprint links committed AI config to lower defect costs, though authors note it is correlational and hypothesis-generating rather than proven causality.
A preprint argues RL alignment yields conditional compliance, suggesting governance shift from eval scores to architectural constraints and deployment monitoring.
A preprint argues GPU utilization misleads LLM capacity planning by conflating memory-bound decode with compute saturation, urging per-phase metrics.
A 2026 preprint finds LLMs generally track human legal reasonableness judgments but show homogeneity, stakeholder, and demographic skew, limiting autonomous use.
An arXiv preprint argues LLM metacognition is coarse and context-dependent, suggesting self-amendment requires external calibration and human approval gates.
A preprint argues zero-shot accuracy misses distribution drift in quantized LLMs, recommending divergence metrics like JSD and TV against BF16 bases for safer deployment.
Preprints show forced evidence gathering raised clinical agent error from 28.3% to 34.3%, while curated tool menus lifted ToolBench success to 0.898, suggesting gating is key.
FluxMoE streams MoE weights from host DRAM to save VRAM, trading capacity for bandwidth. Author-reported gains on multi-GPU setups lack independent replication.
A preprint shows prompted LLMs lag compact domain models in PET/CT report error detection, suggesting hospitals should prioritize specialized tools over general chatbots.
Deltafin streams Kimi K3 weights from SSDs to run on 64 GB Macs, but self-reported benchmarks vary widely, limiting it to batch workloads rather than interactive use.
A preprint reports that verbalized LLM confidence weakly tracks logprob signals, with instruction-tuned models showing worse calibration. Gate on verifiable internal signals.
vLLM's ROCm speculative decoding speedup is self-reported and unreplicated. AMD's gigawatt deals de-risk the platform, but operators must benchmark acceptance rates on MI300X.
A Hacker News claim that 9 in 10 European CDN users rely on Cloudflare lacks independent verification. Operators must audit failover paths to avoid correlated failure and meet
JetBrains' Mellum2 ships as Apache 2.0 weights with 12B total and 2.5B active parameters; here is what the release settles before you self-host.