AI BriefingPyTorchKR news
The latest AI and developer news from PyTorchKR, with short summaries.
Visit website →arXiv Paper Structure-Preserving Translation Skill Released
A tool that performs parallel translation based on LaTeX source to prevent formulas and tables from breaking
Tencent Releases A.I.G, an Open-Source Red-Team Platform for Integrated AI Infrastructure, MCP, and Agent Auditing
An Apache-2.0 licensed tool that automates vulnerability checks across layers, from infrastructure to model jailbreaking.
AI Agent Harness Enhancements and Visual Latent Space Utilization: A Collection of Papers
Includes research on improving agent performance and safety, such as ScholarEvolve and VeriHarness, along with practical application papers like LIME and FlowAgent.
Humanfia Releases 'humanize', a Coding Agent Flow Runner
Supports 12 backends including Claude Code and offers 14 official flows
NVIDIA Unveils Technology to Cut Video AI Development and Operational Costs with VSS 3.3
Adaptive EVS reduces VLM input tokens by 80% and improves concurrent stream throughput by 46%
Open-source 'dots' attempts to bypass bot detection with Firefox patch
The open-source agent 'dots', unrelated to OpenAI, claims to bypass bot detection systems by using a Firefox with fingerprints defined at the C++ stage.
Tracing Agent Harness Call Paths with NVIDIA NeMo Relay
Using NVIDIA NeMo Relay to trace Hermes Agent before and after harness changes revealed that the success rate increase for Qwen3 Coder 30B was partly due to eliminating provider errors, while average LLM calls and duration increased.
NVIDIA Releases OpenShell for AI Agent Security
Rust-based OpenShell controls agent access through kernel-level isolation and Z3-based policy proofs.
Bilibili Releases Index-Translate, a Translation Model Supporting 150 Languages
Released five models and a technical report under the Apache 2.0 license, covering text, speech, and long-form translation.
Liquid AI Releases Decision Model d1
A Jev-compatible API-based model offering low latency and calibrated probabilities with 0 output tokens
PickOpenAI Announces Three-Tier Overhaul of ChatGPT Pro Subscription
OpenAI has restructured the ChatGPT Pro subscription into three tiers—Pro 100, 200, and 500—including the Ultrafast mode for GPT-6 Astra in the Pro 500 tier.
x402 v2 and MCP 2026-07-28: Separating Budget Authorization from Payment Visibility in AI Agents
The new specifications move payment metadata to HTTP headers for gateway visibility but require a separate, isolated authorization service to prevent agents from self-approving expenditures.
Enterprise MCP Server Obligations for x402 v2 and MCP 2026-07-28 Payment Integration
The article defines five critical server obligations, including price re-verification and replay rejection, to secure agent payments under the new x402 v2 and MCP 2026-07-28 specifications.
RelateAnything Supports 17.2ms Inference on RTX 3080 Laptop
RelateAnything supports an inference speed of 17.2ms per image on the RTX 3080 Laptop GPU using TensorRT FP32 and has released the RA-4M training corpus, which includes over 474,000 images and 4.3 million relationships.
Fireworks AI Releases 'Ember-1' Based on Kimi K3: Reduces Tokens by 40% While Maintaining Quality
Fireworks AI has released 'Ember-1' as a research preview, an inference-optimized model based on Kimi K3 that reduces token usage by approximately 35–40% without quality loss in benchmarks and real-traffic tests.
OpenRig Released, Supports Saving and Restoring Claude Code and Codex Sessions
A local-based multi-agent management tool that solves the issue of restoring agent configurations taking 30 minutes upon reboot.
PickOpenProse Released: Markdown-Based AI Agent Task Definition
A Markdown-based contract language that declares states and conditions instead of execution order, re-running only upon changes to reduce costs.
VIDRAFT Releases Darwin-180B-RSI… Achieves Five First-Place Rankings Including Perfect Score on AIME2026
By modifying only 0.02% of the parameters of an open model, it achieved first place in five categories, including a perfect score on AIME2026, and released the weights.
Datadog Releases Fine-Tuning Technique for Small LLMs for Root Cause Analysis
Datadog fine-tuned Qwen3.5-9B using GLM-5.3 investigation traces, reducing investigation costs by 1/20 while maintaining 87% performance.
Hugging Face Adds Native GGUF Execution to Transformers
Hugging Face Transformers has released a feature that allows direct inference of GGUF models in a compressed state without dequantization on Apple Silicon Macs, utilizing llama.cpp's ggml Metal kernels.