The Groundy archive · Page 2 of 43
Browse Groundy's complete archive of 1027 articles on AI, developer tools and infrastructure. Page 2 of 43.
25–48 of 1027 articles · Newest first
DualGauge-Bench shows LLM code often passes functional tests but fails security checks. A layered gate using static analysis, sanitizers, and targeted review is required for L
A repo-stated run shows five Qwen3.8-27B agents matching a harness on ten problems, but independent verification and compute-matched baselines are required before routing.
Scale-conditioned evaluation shows agent memory failure modes vary by retrieval topology and model size, requiring specific measurement before adding more storage.
LLM decode speed is limited by memory bandwidth, not FLOPS. Preprint evidence shows weight streaming dominates latency, making bandwidth the key sizing metric for inference.
Local LLM token generation is limited by memory bandwidth, not compute. Prioritize GB/s over TFLOPS for single-user decode speed and model fit.
Estimate LLM decode speed by dividing effective memory bandwidth by active bytes per token. Rank GPUs for local chat on bandwidth and VRAM, not FLOPS, as single-stream decode
A preprint finds GPU power caps rarely save energy in LLM decode, as memory-bound workloads stay under limits. Savings appear only in large-batch prefill, trading throughput.
A study of 3,997 attempts shows Pydantic errors often hide length violations behind markdown fences. Stripping fences repairs 48% of failures, revealing the true binding.
Mixed text and vision requests stall GPUs due to heavy prefills. Research suggests modality-aware scheduling and routing can reduce latency without new hardware.
The GUEST preprint proposes 51 recommendations for GenAI-assisted systematic reviews, concluding that human oversight is required for all stages except supervised screening.
A July 2026 study shows training-free AI image attribution works best using intermediate CLIP or DINOv2 layers and semantically curated references, though robustness to neural
eMudhra's weak intermediate keys allow public forgery, leaving non-browser clients exposed until trust stores are updated with SPKI-based distrust entries.
nesformer compiles a 6502 CPU into Qwen3.5 weights without training, executing NES games via forward pass. Author-reported tests show frame-exact accuracy but 150-290x slower.
JetBrains ThinkRail adds worktree isolation and review surfaces to the pi agent, but vendor-documented MCP gating remains unverified and telemetry cannot be fully disabled.
An agentic workflow ports Fortran kernels to Kokkos with bounded autonomy, but verification costs and MPI entanglement require human review for production reliability.
A protocol for verifying Particle-1 token savings in Claude Code and Cursor by measuring tokens per completed task against a quality gate, as vendor claims lack independent.
Cloudflare's GA private origins service allows proxied DNS records to reach internal APIs via Cloudflare One, eliminating per-host daemons but requiring return routes and forf
A preprint suggests AR jailbreaks transfer to diffusion LLMs via response templates, reaching 61.3-76.5% success. This supports rerunning corpora and funding new masking tests
Vercel AI Gateway swapped Mistral Large 4 weights under the same ID. Teams must re-baseline coding, cost, and injection metrics before trusting old evals.
Surge AI's 10.8% DAYJOB Tax score highlights omission risks in AI agents. Firms must build independent review controls to catch missed deadlines and legal issues.
Cursor's Agent SDK offers clearer self-hosted execution, while Claude's side has better-documented permissions. This asymmetry requires verification before standardizing on a
A single-author demo embeds a RISC-V CPU in Qwen 3.5 weights, showing GDN state holds exact computation invisible to CoT monitors, a structural gap for hybrid models.
Claude Code token overhead splits into background agent summaries and in-context MCP schemas. Measure both via OpenTelemetry to isolate costs and optimize cache usage.
Qwenputer embeds a RISC-V CPU in Qwen 3.5 weights, using GDN state as exact RAM. This author-reported demo shows CoT logs miss execution, urging state-level inspection for aud