A must‑follow public account sharing practical AI insights. Follow now. internet + machine learning + big data + architecture = IMBA
vLLM 0.32.0 introduces ROCm Segmented Attention backend and W4A16 kernel optimizations for AMD Ryzen AI MAX+ 395 (gfx1151), delivering up to 2.5x attention throughput and 22% TTFT reduction, though FP8 support remains limited without native hardware acceleration.
Shanghai AI Lab's Harness-of-Harness (HoH) framework organizes AI coding agents into planning, development, and QA roles with persistent state across iterations, achieving 52.25% average improvement over vanilla agents on three benchmarks and demonstrating 70-round autonomous FPS game development with 81 tracked issues.
DeepSeek Harness 0.2 shifts focus from Desktop to core runtime improvements: Schedule becomes an optional Bundle, Windows Sandbox adds permission diagnosis with backup/restore, an experimental Async Question Mode decouples user prompts from agent execution, and DeepSeek account web search no longer requires a separate API key.
This article explains how memory poisoning attacks persist across AI agent sessions, details the MINJA attack technique, and proposes auditable memory storage formats, secure write paths with trust scoring, and detection methods including integrity checks and activation monitoring to defend against persistent memory corruption.
This article details the major data structure changes in DeepSeek Harness Session V4, including tool/result message restructuring, source.kind identity model shift, new developer/message surface type, dynamic tool history via headerSeq, deferLoading schema support, and migration rules affecting plugin authors and session parsers.
DeepSeek Harness 0.1.7 moves beyond adding tools (0.1.6) to solving long-term agent stability with Session V4, background job scheduling, improved file review, Messages API-only adoption, plugin version compatibility, unified configuration, and read-only agent team task boards.
Google's TimesFM-3 introduces native multivariate time series forecasting with non-autoregressive decoding, enabling joint predictions across multiple targets and covariates in a single forward pass, outperforming competitors on benchmarks while offering zero-shot generalization and probabilistic outputs.
FlashSpec implements speculative decoding using GPU-native Triton kernel verification and online bandit-based draft model selection, sharing six practical lessons on specification-first development, hidden temperature bugs, cross-platform packaging pitfalls, kernel performance trade-offs, property-based testing value, and adaptive algorithm prerequisites.
The article distinguishes Prompt Engineering (single human-verified interactions) from Loop Engineering (automated iterative loops with testable success conditions), explains their hierarchical relationship, compares use cases, risks, and argues that Loop Engineering builds on Prompt Engineering to automate repetitive, verifiable tasks.
The article details why single-step RAG fails for complex queries, presents a production-grade multi-step agentic RAG architecture using a DAG state machine, demonstrates a three-iteration Rivian supply chain example, and outlines key engineering principles including explicit state serialization, async execution boundaries, and deterministic evaluation guards.
