LLMS

llama.cpp now caches hot MoE experts on the GPU. In the PR author's RTX 5090 run, --cpu-moe plus --moe-cache-mib decoded 2.2x faster than --fit. When to use each flag.

LLM distillation on one GPU: cache the teacher's top-100 logits offline and use a fused chunked KL loss to cut a GPT-OSS 20B run from four nodes to one.

TutorMoments, AI2's replay-based benchmark for LLM tutors, finds models over-help by default. How it scores scaffolding and rigor in real tutoring transcripts.

Blind resampling beats self-repair in small code models: a placebo-controlled MBPP+ study finds retrying from scratch beats feeding failed code back.

CAPA benchmark: coding agents that read a user's past sessions resolve ambiguous requests with far fewer clarifying questions, tested across 12 models.

Inkling-Small is a 12B-active, Apache 2.0 open-weights MoE that beats the 975B Inkling on agentic coding and tool use at a quarter of the size.

Fable 5 effort levels vs switching models: effort spends more of one model's ceiling, a model switch buys a new one. When to move each dial.

DeepSeek paused its fundraising after a leaked investor transcript exposed its compute gap with US labs. What Liang Wenfeng said and how HN read it.

Mozilla's State of Open Source AI report puts open models within 3.3% of closed ones, yet only 51% reach production. The deployment barriers explained.

Detecting LLM text with classical ML: how a TF-IDF and linear SVM detector reached about 85% accuracy without neural networks, and the HN false positive debate.

A new American open-weights frontier model with multimodal capabilities, 1M token context, and competitive benchmarks. Here's what the HN community thinks.

A Hacker News discussion blows up over LLM vocabulary quirks, with developers sharing hooks, filters, and coping mechanisms for repetitive Claude-isms.

George Hotz on LLMs: why he loves the tech but rejects doom narratives and AGI inevitability hype, plus how the Hacker News thread split on his argument.

A solo developer built a 1,300-line C inference engine that runs the 744B GLM 5.2 model on consumer hardware by streaming routed experts from disk.

A Dartmouth study of an AI tutor in intro statistics links full engagement to a 0.71-1.30 SD exam gain. What it measured, and why selection bias still matters.

Anthropic's new research reveals LLMs have an internal 'workspace' for silent reasoning - and it could change how we build safer AI.

Comparing LLMs by token pricing alone can lead you to choose worse, more expensive models. Cost per task tells the real story.

Filippo Valsorda argues LLMs have made vulnerability reports less special. His case against old coordinated disclosure norms, plus the HN pushback.

Run GLM-5.2 locally with Unsloth dynamic quantization: hardware math for a 256GB Mac or 24GB GPU, quantization tradeoffs and lessons from the HN thread.

VibeThinker-3B is a 3B parameter reasoning model reporting 94.3% on AIME 2026. What the paper claims, how to run it locally, and what it trades away.