📖 AI-Radar · 2026-W38

R2VC: Modular Fact-Checking Splits Retrieval, Verification, and Confidence

R2VC introduces a modular fact-checking pipeline that separates Wikipedia retrieval, candidate generation, NLI verification, and confidence calibration. With an 8B backbone, it improves FEVER accuracy by 13.74% over baseline. Ablations show verifier-based selection and calibration contribute most; retrieval failures, especially wrong-entity evidence, remain the main bottleneck.

14 Sep #Hardware #LLM On-Premise #Fine-Tuning

Moonshot’s Kimi K3 tests the business case for open-weight AI

Moonshot AI reports more than US$1 billion in annual recurring revenue as of August and targets US$2 billion by end-2026. The gains come from Kimi subscriptions and services built on the open-weight K3 model. The licence imposes commercial agreements on large API providers, while inference costs remain opaque. The growth tests the economics of open-weight AI.

14 Sep #Hardware #LLM On-Premise #DevOps

Google's TPU pays back in one year, boosting MediaTek and Broadcom ASIC revenue

According to DIGITIMES, Google's TPU recoups its investment in about a year. The figure opens upside for MediaTek and Broadcom ASIC revenues and signals that custom silicon for training and inference is now a concrete economic option. The fast payback shifts demand toward dedicated accelerators, with structural effects on the supply chain and deployment choices.

14 Sep #Hardware #LLM On-Premise #DevOps

The Right to Run Local AI: When Safety Becomes a Pretext for Centralization

A Reddit post raises a thorny issue: the AI safety debate risks hitting open source, limiting the right to run models locally. The analysis explores implications for data sovereignty, hardware, and market incentives, showing how poorly calibrated restrictions could favor closed cloud services and weaken those who need direct control.

14 Sep #Hardware #LLM On-Premise #DevOps