AI Briefing
TLDR AI news
The latest AI and developer news from TLDR AI, with short summaries.
Visit website →Learning to Replicate Expert Judgment in Financial Work
High-quality human annotations were leveraged to train LLMs to perform information filtering and judgment at the level of financial experts.
ZCode
ZCode, the official development environment for GLM-5.2, has been released for macOS, Windows, and Linux.
PorTAL: A Portable Task Adapter for LLMs
PorTAL is a base-agnostic technology that allows task adaptation learned once to be easily transferred to a new LLM.
Continual Harness: An Efficient Self-Improving Agent for ARC-AGI-3
Continual Harness demonstrated extremely high efficiency on the ARC-AGI-3 benchmark through skill reuse and self-improvement.
Autoresearch: The Core of Self-Improving Agents Is the Feedback Loop (11 min read)
It explains the 'outer loop' structure through which AI agents improve their own performance using feedback signals and human input.
How OpenAI Delivers Low-Latency Voice AI to 900 Million Users
OpenAI leverages WebRTC technology to deliver a near-real-time, low-latency Voice AI experience to 900 million users.
Meta Reportedly Planning Cloud Business to Sell AI Computing Power
Meta is preparing a cloud business that sells its surplus AI computing resources and models to external developers, challenging existing hyperscalers.
Recent OpenAI Research Demonstrates LLMs' Ability to Solve Difficult Mathematical Problems
OpenAI researchers used GPT 5.5 Pro and Claude Opus 4.8 to solve 9 difficult open problems in mathematics and theoretical computer science.
Miles: A PyTorch-Native Stack for Large-Scale LLM RL Post-Training
RadixArk has released Miles, an open-source framework that integrates SGLang, Megatron-LM, and Ray for large-scale LLM reinforcement learning (RL) post-training.
GeneBench-Pro: Evaluating AI Agents' Scientific Judgment Capabilities
OpenAI has released GeneBench-Pro, which evaluates AI agents' ability to conduct biological research.
Meituan Launches LongCat-2.0 API with 1.6T Parameters
Meituan has released LongCat-2.0, a 1.6-trillion-parameter MoE model, in API form, strengthening agentic coding and long-context processing capabilities.
Nvidia Rival Etched Reaches $5 Billion Valuation
Etched, gaining attention as a rival to Nvidia, is showing strong growth after reaching a valuation of $5 billion.
AI and the Future of Mathematics
The pattern of AI's mathematical progress will serve as a milestone showing the changes to come across the broader economy.
Inside Thinking Machines' Interaction Model
Thinking Machines is developing a new interaction model that supports continuous collaboration between humans and AI.
Why Specialization Is Inevitable
Under resource constraints, specialized AI models focused on specific domains show higher efficiency than general-purpose versatility.
Building a Document Processing Workflow in 30 Minutes
Explains how to build a medical document processing pipeline using Mistral's OCR, Agents, and Workflows.
DeepSeek Open-Sources New Framework 'DSpark' That Boosts LLM Inference Speed by Up to 85%
DeepSeek has released DSpark, a speculative decoding framework that dramatically improves LLM response speed, under the MIT license.
Devin Fusion (8-minute read)
Cognition unveiled Devin Fusion, a multi-model harness that cuts costs by 35% while maintaining frontier-level performance.
Cursor Launches iOS App: Control Coding Agents from Anywhere
Cursor has launched an iOS app, enabling users to control cloud and local agents and manage PRs on mobile.
The Economics of Tokens (10-minute read)
The AI industry is shifting from a vertically integrated structure to a modular ecosystem centered on Transformer and APIs, accelerating innovation.