Open-Source AI

AI Solved 370 Math Problems, Mathematicians Panic, Musk Seeks $40 Billion: This Week's Highlights

Oct 12, 2026·4 min

Robotics

SAMF Teaches Drones to Locate Themselves on a Satellite Map Using Thermal Imaging, No GNSS Required

APEX Compresses Diffusion Latents 128x and Beats a 256-Token Model on ImageNet-512

Research·arXiv cs.LG

Wasserstein BaryBooster Turns Diffusion Realism into Event Forecast Accuracy

The authors recast event-sequence forecasting as conditional generation and show that no method wins on both accuracy and realism. BaryBooster offers a way out.

LLM·arXiv cs.LG

Kairos: a byte-level model for all modalities converges on 440 bytes

Kairos' author built a byte-level multimodal architecture with MoE and linear attention and checked that it trains at all: loss of about 10⁻³ on a single 440-byte…

LLM·arXiv cs.CL

U4RASD Takes 1st Place at KnowledgeGraphEval-2026: 94.06 micro-F1 on Arabic

An ensemble of Yehia, Qwen3.5 and an encoder with typed markers won the Arabic relation extraction subtask: micro-F1 of 94.06 across 40 relation types.

LLM·arXiv cs.CL

GeopoliticalQA: bge-m3 Beats Commercial Embeddings on 5 525 Report-Based Questions

A new benchmark tests RAG on 415 geopolitical reports and 5 525 questions: compact bge-m3 outperforms commercial APIs, while generation struggles with inference and…

01Qwen Picks What to Remember: Replay From a Frozen Model Reduces ForgettingOct 12, 2026·3 min 02CLUE Builds an Event Ontology Without Experts in Three Steps: Embeddings, Clusters, NamesOct 12, 2026·2 min 03AutoIndexer Teaches a Language Model to Edit Its Own Output with Three Types of EditsOct 12, 2026·2 min 04Seq-Drifting Generates Text in a Single Pass: Perplexity of 36.01 on LM1BOct 12, 2026·2 min 05GeoSpatialBench Tests MLLMs on 28 Types of Spatial Relations Across Six CitiesOct 12, 2026·2 min

Top themes

LLM·arXiv cs.AI

RGR Protects LLMs from Skill Loss in Self-Distillation by Preserving Representation Geometry

LLM·arXiv cs.AI

FrontierCode: Top AI Models Fail Maintainer Review in 42% of Attempts

Cognition and Artificial Analysis launched FrontierCode 1.1: Claude Opus 5.5 leads with 52% and a 58% pass rate, at $10.15 per task. GPT-6.1 Sol scores 49% for…

Oct 12, 2026·3 min

LLM·Habr AI

Why Repeated Fine-Tuning Hits a Ceiling: The LANN-4 Study

The LANN-4 author ran experiments with repeated fine-tuning of a small model. The first round produced a jump from 53% to 59% correct answers on 200 tasks — the…

LLM·YouTube: Two Minute Papers

DeepSeek-OCR Compresses Text Context 10x at 97% Accuracy

DeepSeek proposes storing long text as an image: at 10x compression the model recovers 97% of the content, and at 20x accuracy drops to 60%.

LLM·Habr AI

Agentic search on open-source models: who it found, what it cost, and where it all broke

LLM·Хабр: Машинное обучение

How to Build a RAG Bot on a T4 for Civics Textbooks: Qwen Instead of Web Search

LLM·The Verge

The Verge Launches a Local AI Diary: Hermes, Qwen and a Mac Studio

LLM·Selectel

Selectel Walks Through Installing Jenkins on a Cloud Server: From Resources to a Pipeline

LLM·RunPod Blog

Runpod's roundup: Llama-2 with a 32k-token context and SDXL 1.0 now publicly available

Runpod: How to Set Up Full SSH for SCP and IDEs on GPU Pods

Runpod Sponsors Project Odyssey, CivitAI's Open-Source AI Film Contest

Deep Cogito trained its Cogito v1 family on Runpod infrastructure on the road to ASI

LLM·RunPod Blog

Runpod shows how to deploy Llama 3.1 8B on vLLM without your own servers

Runpod Blog has published a guide to running Llama 3.1 8B Instruct on vLLM via Serverless: GPU time is billed per second, and workers scale up from zero under load.

LLM·RunPod Blog

RunPod Publishes Guide to Moving from the OpenAI API to a Self-Hosted Open-Weight LLM

RunPod has published a guide to migrating from the OpenAI API to a self-hosted open-weight LLM, aimed at teams hitting rate limits and growing token bills.