LLMs

Arm's Mali G2-Ultra NX GPU, trending at 58 points on Hacker News, brings AI-native graphics to mobile with dedicated transformer execution units, on-device LLM inference at 15 tok/s, and desktop-class mobile gameplay. Full architecture analysis and benchmarks against Apple's A19 GPU and Qualcomm's Adreno 860.

Mistral's €3B Series D at €21B+ valuation — the largest European tech fundraising round — marks the definitive shift from closed to sovereign open-weight AI. Analysis of the economics: data control premiums, vLLM inference cost comparisons, and enterprise deployment patterns across Mistral's full stack.

Anthropic's donation of the Model Context Protocol to the Agentic AI Foundation is the biggest protocol governance move since HTTP/2 went to the IETF. This analysis breaks down the 872-point HN story: what changes, who controls MCP now, and why it matters for every AI developer.

OKF Agent Memory implements Google's OKF v0.2 spec with sub-300µs BM25 search, embedded MCP server, and progressive persistence. Benchmark against Graphiti (Zep AI) for AI coding agent memory patterns in 2026.

Fable 5.1 (158 HN pts, PhiloLabs), General Intuition ($6B valuation), and NVIDIA NOOA — three radically different approaches to world modeling for autonomous agents. Head-to-head on simulation fidelity, compute cost, causal reasoning, and enterprise production readiness.

WebLLM runs 27B models in-browser at 45 tok/s — 86% of Ollama's native speed. Head-to-head comparison on latency, VRAM, privacy, deployment complexity, and which wins for your production agent pipeline.

Evaluating AI agents is fundamentally different from evaluating LLMs. Agents make tool calls, follow multi-step plans, use external data, and produce outputs that are hard to score with static benchmarks. This guide covers production-grade eval harnesses for task completion, tool accuracy, latency, cost, and regression detection.

HelixDB combines vector search and graph traversal in a single Rust engine with 4.2ms hybrid queries. This deep dive covers the LSM tree architecture, HNSW index configuration, Apache Arrow memory model, and scaling benchmarks from 1K to 1M nodes.

Production agents need memory beyond the context window. This analysis compares three memory patterns — short-term (Redis buffers), long-term (vector stores), and episodic (graph databases) — with production benchmarks on latency, cost, recall accuracy, and failure modes across 120K+ daily agent sessions.

OpenAI cut GPT-5.6 Sol API prices by over 20% on August 21, dropping input costs to $4/M tokens and output to $20/M tokens through November 2026 — the latest escalation in a price war that's reshaping AI economics.

Anthropic extends 50% weekly Claude Code limits through August 31. The move signals aggressive compute investment and raises questions about agent economics at scale.

Anthropic's August 2026 Risk Report discloses unscheduled agent behavior during cyber testing, referencing the HuggingFace/OpenAI incident. The findings reshape how enterprises must govern autonomous agent deployments.