All articles
39569 articles
DSH + Ollama Local LLM Deployment in Practice: Free Chat with Da Fei Yu
Bilibili creator tests DSH + Ollama local LLM deployment end to end: Ollama install, Xunlei-accelerated downloads, ModelScope GGUF model import, DSH custom API address setup, and a comparison of Qwen 3.5 4B, 8B, Gemma, and other models on 8GB VRAM, with advice on choosing local deployment vs. API calls.
Tested: A $899 Mac Mini vs. a $200/Month AI Subscription
Real-world test of whether an M4 Mac mini can replace a $200 AI subscription: 16GB can't fit large models, and both speed and quality fall short—but privacy, no caps, and near-zero running costs make it an ideal host for an always-on agent. Includes buying advice per configuration and money-saving strategies.
RSTGameTranslation: Real-Time OCR Localization + Offline Translation with Local LLMs
RSTGameTranslation is an open-source real-time Windows game translation tool: OCR for English, Japanese, and Korean with Chinese overlays in place, plus Ollama or LM Studio local LLM offline translation and Whisper speech transcription for games, manga, and visual novels.
OpenAI Agents Uncover 948 Linux Kernel Vulnerabilities at a Staggering Cost Per Run
OpenAI's AI agent pipeline found 948 Linux kernel vulnerabilities in one run, auto-generating exploits and patches at $1.2M cost. GPT-6 Astra shows odd drop in security research tasks.
Mistral Opens Up Trillion-Parameter Model, Microsoft Discloses $10B OpenAI Revenue
AI news roundup: Mistral opens trillion-parameter Large 4 weights, Tencent previews i4 with vision, Microsoft discloses ~$10B quarterly OpenAI revenue, Claude Opus 4.8 fast mode removal, OpenAI Model Spec update, NVIDIA open speech model.
Running Qwen 3.8 27B Locally: Identifying Games from Model Weights Alone
Engineer Simon Willison ran Qwen 3.8 27B locally and identified Breakout (88%) and Pong (75%) from game engine weights alone, showcasing weight space programming.
The Robustness Dilemma of LLM Social Agents: Context Enrichment and a Selective Robustness Framework
LLM social agents drift under semantically equivalent prompt changes, threatening simulation credibility. This article analyzes a new study's controlled context enrichment method and selective robustness framework, exploring how four types of context enrichment affect agent behavior and model differences in Prisoner's Dilemma experiments.
How Do Hierarchical Embeddings and Quantization Trade Off? MTEB Benchmark Reveals the Balance Between Efficiency and Effectiveness
MTEB benchmark study evaluates the interaction between hierarchical representation and quantization in text embeddings: int8 quantization is nearly lossless, binary quantization is significantly worse; 8x dimension reduction plus int8 achieves 32x storage compression while retaining 94.4% of performance; retrieval tasks degrade fastest.
SpecCompact: Slot-Parallel Speculation Accelerates LLM Agent Context Compaction
SpecCompact accelerates LLM agent context compaction via slot-parallel semantic speculation, achieving 2.04–5.63x speedup on AppWorld and τ³, with 8.68x median in user studies, while revealing model- and task-dependent trade-offs.
LongBanana: An Expert-Validated Benchmark for Long-Context Multi-Reference Image Synthesis
LongBanana is the first expert-designed, validated benchmark for long-context multi-reference image synthesis—90 tasks, 4,670 atomic checks, and the LongBanana-Gen agent framework.
Persona Perspectives Reshape LLM Internal Representations: Challenges for Safety Monitoring
New research finds that when LLMs adopt different personas, the geometric structure of their internal representations changes systematically, causing monitoring tools like linear probes and SAEs to produce politically biased evaluations of truth and morality—posing challenges for safety monitoring.
The Temporal Safety Blind Spot in LLM Assistants: Computing Time ≠ Applying It — TASC Benchmark Reveals Only 30.5% Compliance Rate
TASC benchmark evaluates 13 LLM assistants' temporal constraint application: only 30.5% of violating recommendations considered time limits; 62.1% of failures stem from computing time without applying it; explicit temporal info fixes only about half.
SAUCE: A New Reference-Free, Explainable Method for Dialogue Summarization Evaluation
SAUCE is a reference-free, fact-based evaluation method for cross-lingual dialogue speech summarization that measures accuracy and factual coverage via entailment between conversation and summary, localizes hallucinations and omissions, and shows high consistency with human judgment in preliminary studies.
LooM: Loop-Invariant Memory Makes Loop Transformer Inference Budget Adjustable
LooM proposes loop-invariant memory to solve three loop Transformer issues: memory growing with loops, coarse computation granularity, and hard-to-change loop counts. Looping only in CED decoder layers keeps the KV cache from growing with loops, supporting a continuous budget of 3 to 21 layer executions per token with a single model covering all budgets and no router.
LLMs Don't Know What They Know: The Temporal Boundary Dilemma in Financial Backtesting
FinDendro benchmark reveals LLM self-knowledge reports severely underestimate true knowledge: GPT-5 answers 76% correctly but claims to know only 38%. Vendor cutoff dates, model self-reports, and date instructions all fail to guarantee leak-free financial backtesting—behavioral measurement offers a new path to defining credible backtest windows.
FlowReconBench: Sparse-Sensor Flow Field Reconstruction — Are Learned Models or Classical Methods Stronger?
FlowReconBench benchmark evaluates ten learned and five classical methods for sparse-sensor flow reconstruction. Classical methods lead with low-order, dense sensing; learned methods win with sparse sensing and low observability. Calibration and cost metrics yield different rankings.
Measuring Risk, Not Harm: A New Framework for Detecting Psychological Harm Risk Factors in LLM Conversations
Researchers propose "measuring risk, not harm": operationalizing LLM psychological harm risk as eight observable AI behaviors, building a synthetic multi-label corpus of 47K+ responses, and comparing embedding classifiers, LLM judging, and a cascade achieving F1 0.79 with 97.1% fewer judge calls.
Reinforcement Learning Unlocks Multi-Task Self-Evolution for Generalist Robot Policies
New research shows reinforcement learning enables multi-task self-improvement of generalist robot policies during deployment, with value-based RL and training task distribution as key factors, even generalizing to untrained tasks.
Where Is Mbed TLS Maintenance Heading? Community Discussion Draws Attention
Hacker News discusses the future maintenance of Mbed TLS. This article analyzes the potential impact of changes to this lightweight TLS library's maintenance arrangements on embedded and IoT projects, and offers advice for dependents on tracking and response.
Why Do Faithfulness Evaluation Criteria Disagree? A New Framework Offers a Navigation Path
The paper proposes a faithfulness functional framework that locates disagreement among faithfulness evaluation criteria in three dimensions—claims, comparisons, and aggregation—and uses adverse action notices to show how institutional context narrows operationalization choices, offering auditors a path to navigate conflicting evaluations.