Engineering
Making Knowledge Distillation Cheap Enough to Run at Scale
Researchers at Multiverse Computing introduce two systems-level optimizations for LLM knowledge distillation: caching the teacher model’s top-100 logits offline so…
Engineering
Thinking of ACE? We Can Do It with Fewer Tokens
IBM Research compares its ALTK-Evolve agentic memory approach to the ACE method, which injects one large, comprehensive evolving playbook into…
Opinions
An FT column was condensed with AI. It crossed a novel ethical line.
Poynter’s Ren LaForme examines a Financial Times editorial ethics incident in which Harvard economist Ricardo Hausmann used an AI tool…
Opinions
‘Rogue AI Agents’ Aren’t Rogue, They’re Fulfilling Their Functional Goal: Automating Sociopathology
Blogger Charles Hugh Smith argues that so-called ‘rogue’ AI agents are functioning exactly as designed — pursuing reward maximization without…
Opinions
We’re Having the Wrong Debate About AI Safety
Bloomberg Opinion columnist Catherine Thorbecke argues that public and policy attention on AI safety is misdirected toward speculative frontier risks,…
Engineering
Hierarchical Self-Improvement: A Framework for Task-Specific Evolvable Agent Harnesses
Researchers proposed Hierarchical Self-Improvement (HSI), a framework that lets frozen language models improve their performance by evolving task-specific execution harnesses…
Engineering
FlashPrefill V2: Block-Sparse Prefill Attention for Long-Context LLM Serving
Researchers presented FlashPrefill V2, an advancement to block-sparse attention that targets the computational bottleneck of processing long text sequences during…
Engineering
Base Optimization Stack: From Open Weights to Frontier On-Device Inference Speed
Base Compute described an automated optimization pipeline that converts open-source model weights into hardware-specific inference implementations, using AI research agents…
Engineering
VLM Run Gateway: Run GLM-OCR, DeepSeek-OCR-2, and dots.mocr with an OpenAI-Compatible API
VLM Run Gateway is a unified platform offering a single OpenAI-compatible endpoint for open-weight OCR and vision-language models, including DeepSeek-OCR-2,…
Engineering
mLateOn: A New SoTA for Multilingual ColBERT-Style Retrieval, Evaluated on HAKARI-Bench
mLateOn is a multilingual retrieval model that combines a multilingual mmBERT encoder with ColBERT-style late-interaction token matching. Evaluated on the…
Engineering
Who Tells You Whether the Molecule Your AI Just Designed Is Any Good?
The Open Discovery Challenge introduces a computational leaderboard for scoring AI-designed malaria drug candidates across six dimensions: whole-cell activity, target…
Engineering
LFM2.5-VL-3B for Better and Faster Vision Capabilities for the Edge
Liquid AI introduced LFM2.5-VL-3B, a vision-language model that pairs a SigLIP2 400M NaFlex vision encoder with the LFM2.5-2.6B text backbone.…