The archive · Page 2 of 7
The serving stack, network fabric, and cloud-account substrate beneath production AI, where every throughput claim collides with rebuild windows, egress invoices, and control-plane risk.
25–48 of 167 articles · Newest first
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
A preprint shows idle Macs can pool for batch embeddings and Ollama completions, offering private inference that complements rented GPUs for latency-tolerant work.
Ai2 reports that replacing priority queues with GPU budgets and fair-share allocation delivered 98% of owed hours, though evidence is self-reported and limited to one 30-day.
Cloudflare Artifacts offers Git-native storage for Workers, but lacks documented consistency models and pack generation benchmarks, leaving critical protocol costs unverified.
A 2026 survey and primary studies show hybrid retrieval with reranking and contextual chunks outperforms complex agentic stacks for most queries, while deferral policies cut L
KyGround preprint shows vector RAG outperforms tool-calling retrieval on small Greek corpora, while whole-KB prompting achieves highest accuracy.
Cloudflare's Web Search API routes queries to partners at list price without markup. It consolidates billing but does not guarantee lower cost or better quality than direct Ex
A preprint shows CPU-based NLI and similarity ensembles detect QA hallucinations well, but summarization detection remains weak without chunking.
Preprint benchmarks show LLM-generated GPU kernels rarely clear production gates, with end-to-end gains shrinking and memory regressions persisting in real serving stacks.
Cloudflare logs TLS key-exchange groups to measure hybrid post-quantum adoption. The metric tracks encryption only, not authentication, and classical shares often reflect non-
A 2026 preprint shows fixed-configuration kernels achieve cross-GPU bitwise determinism for linear layers, challenging the assumption that reproducibility requires a major end
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.
DuckDB research extension FlockMTL integrates LLMs and RAG into SQL via schema objects, moving governance to DDL, though performance claims remain author-reported.
A preprint argues GPU utilization misleads LLM capacity planning by conflating memory-bound decode with compute saturation, urging per-phase metrics.
FluxMoE streams MoE weights from host DRAM to save VRAM, trading capacity for bandwidth. Author-reported gains on multi-GPU setups lack independent replication.
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.
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
A new NCCL shim recovers 13-38% bandwidth on shared GPU clusters by tuning collective patterns. Test for cross-tenant interference before buying more fabric.
Self-hosting Nitter in October 2026 means rotating X account sessions, tracking a twice-archived repo, and weighing X Corp's August 2026 cease-and-desist letters.
A claimed $60 AMD BC-250 with 16GB VRAM shifts the local LLM bottleneck from capacity to software support. Verify ROCm and llama.cpp backend compatibility before buying, as un
Spruce claims 0.21-2.97s private retrieval via MPC, challenging TEE defaults. Compare self-hosted Qdrant, TEEs, and cryptographic outsourcing for privacy-sensitive RAG.
Browser LLM inference is local but not private. The host page controls the GPU, model bytes, and network egress, making 'data never leaves device' a policy promise, not a ver
Cloudflare's bundled bot defense changes procurement math, but efficacy claims remain self-reported. Audit standalone spend and measure false positives before migrating.