The Groundy archive · Page 10 of 34
Browse Groundy's complete archive of 797 articles on AI, developer tools and infrastructure. Page 10 of 34.
217–240 of 797 articles · Newest first
Spectral Compute proposes CUDA binary translation for AMD and Intel GPUs, but vLLM already supports native HIP. The real question is whether translation overhead beats the.
MiniCPM-V-4.6 runs on a 2011 Fermi GPU with 6 GB VRAM. The study shows software staging recovers multimodal inference, shifting the constraint from hardware procurement to.
AGENTS.md and .cursorrules are a write-once, compromise-many attack surface. arXiv 2607.15143 shows a deterministic pre-install check, not sharper models, stops it.
ATSInfer schedules LLM tensors across CPU and integrated GPU at tensor granularity, reporting up to 3.29× decode speedup where pure-CPU inference is memory-bandwidth-bound.
A Korean alignment paper argues LLM cultural work leans on suppression lists and needs positive specs. Whoever writes the spec owns the model's definition of a culture.
A July 2026 arXiv study of 25,264 agentic PRs finds most GitHub repos generate only one to two agent PRs per quarter, contradicting vendor hype about intensive adoption.
CUDA-L2 used reinforcement learning over 1,000 kernel configurations to beat cuBLAS by 19.2% on HGEMM, shifting the scarce kernel skill from hand-tuning to reward design.
A July 2026 risk-field digital twin replays rare AV hazards no road fleet can accumulate, but no regulator defines when simulated miles count as certification evidence.
A NISCAN audit of 461 Linux Foundation repos finds non-inclusive terms down 47% since 2020, yet 62.7% still carry Tier-1 terms, and LLMs keep resurrecting the old names.
Hair-Trigger Alignment shows an LLM can pass every black-box probe, then flip after one benign update, so EU AI Act periodic monitoring certifies less than it assumes.
Two Apache-2.0 TypeScript ORMs with comparable npm volume. The split is abstraction and lock-in: Prisma's generated client versus Drizzle's zero-dependency SQL builder.
pgvector, Pinecone, and Qdrant split on deployment, not on which ANN index is fastest. The call is where vectors live, who runs the cluster, and what filtered search costs.
TALRanker folds the tool-call decision into the reranker's scoring policy, turning tool latency from a fixed per-query tax into a budget the model spends only when uncertain.
An 82.5% baseline unsafe-action rate against LLM agents with network tools shifts prompt injection from a data leak to a production availability and integrity problem.
Both wrap llama.cpp, so Ollama vs LM Studio is a headless MIT daemon versus a proprietary desktop GUI, with one ceiling: neither is a production serving engine.
A new survey reframes LLM efficiency as a memory-bandwidth co-design problem, pushing 2027 fleet sizing toward HBM capacity over peak GPU FLOPS.
Progressive Tree Drafting grows draft tokens as a pruned tree to claim a 2× speedup, but the gain hinges on verifier acceptance and shrinks on code and reasoning.
CLI coding agents fail when stale early-turn assumptions compound across a trajectory, not on a single bad tool call, so per-turn metrics miss where runs actually break.
CoreWeave and Nebius finance their Nvidia fleets with hardware-collateralized debt, so the inference rate you sign embeds a credit risk premium that can move mid-contract.
BERTopic's Hugging Face Hub integration pulls transformer embeddings over HTTP without a local GPU, which collapses Gensim LDA's CPU-only cost advantage for most corpuses.
When a feature flag vendor in your critical path gets acquired by an AI platform company, a routine infrastructure choice becomes a strategic lock-in question for DevOps.
A three-layer sparse matmul kernel achieves 1.64x kernel-level and 1.41x end-to-end speedups, making moderate unstructured sparsity a viable third optimization lever.
LLMbda Calculus proves agent confidentiality via labeled reduction semantics, exposing a gap between vendor sandboxing claims and verifiable information-flow control.
Vercel SAMLStorm protection blocks signature-wrapping attacks at the edge, but self-hosted SAML deployments get no mitigation. Patched libraries can still authenticate forged.