📖 AI-Radar · 2026-W43

DeepSeek Flash GGUF Without Documentation: The Hidden Cost of On-Premise

A GGUF repository with no notes on Hugging Face shifts verification costs to deployers. AI-Radar analyzes quantization, context, VRAM, and provenance, showing why missing documentation blocks self-hosted pipelines, TCO calculation, and sovereignty.

19 Oct

Forgetting Is Not Erasure: Relearning Savings in LLMs

A Qwen2.5-1.5B study shows that a previously learned fact is reacquired faster after partial forgetting, even when current accuracy fails to detect it. The advantage is binding-specific, not just familiarity with entities or format. In a secondary scenario, relearning requires about 18% fewer optimizer updates. For teams updating local LLMs, behavioral accuracy may underestimate the persistent effect of fine-tuning.

19 Oct #LLM On-Premise #Fine-Tuning #DevOps

An LLM as calibration agent: early edge, then hybrid takes over

A study on wake simulations and damped oscillators shows that physics-labeled LLM prompts improve initial calibration in 23 of 25 trials, but the advantage fades after 15 evaluations. Hybridizing with a classical optimizer cuts runs to 9, versus 17 for Nelder-Mead and over 60 for Bayesian optimization.

19 Oct #LLM On-Premise #DevOps

Persistent Human Control for LLM Agents with Typed MCP Overrides

Researchers propose Structured Authority Injection, a protocol that routes human corrections through typed, schema-validated MCP tool calls and maintains a ranked authority context prepended to each LLM prompt. It defines three override classes with distinct persistence semantics and requires no fine-tuning or weight access. Tests across nine LLMs and eight scenarios show reduced ambiguous interpretation and less authority decay, with prompt overhead reduced up to 38% in the strongest models.

19 Oct #LLM On-Premise #Fine-Tuning #DevOps