Improving 15 LLMs at Coding in One Afternoon. Only the Harness Changed
Changing the edit tool in a coding harness improved the performance of 15 LLMs significantly, demonstrating that the harness itself is a critical factor in coding success. This shift highlights that the harness is often overlooked, yet it plays a pivotal role in how models interact with coding tasks.
Gemini 3 Deep Think
Gemini 3 Deep Think has undergone a significant upgrade, enhancing its ability to tackle complex challenges in science, research, and engineering, now accessible to Google AI Ultra subscribers and select researchers via the Gemini API.
MiniMax M2.5 released: 80.2% in SWE-bench Verified
The MiniMax M2.5 model excels in real-world productivity, achieving SOTA performance in coding and agentic tasks, with notable scores of 80.2% in SWE-Bench Verified and completing tasks 37% faster than its predecessor, M2.1.
GLM-OCR – A multimodal OCR model for complex document understanding
GLM-OCR is a multimodal OCR model that enhances document understanding through innovations like Multi-Token Prediction (MTP) loss and a two-stage pipeline for layout analysis, achieving a top score of 94.62 on OmniDocBench V1.5.
Text classification with Python 3.14's ZSTD module
Python 3.14's zstd module introduces a powerful tool for text classification, leveraging incremental compression to enhance efficiency and accuracy in machine learning tasks, particularly for small data sets.
Anthropic raises $30B in Series G funding at $380B post-money valuation
Anthropic has secured $30 billion in Series G funding, elevating its valuation to $380 billion, with plans to enhance enterprise AI and coding capabilities, driven by significant investor interest from firms like GIC and Coatue.
Beginning autonomous operations with the 6th-generation Waymo Driver
Waymo's 6th-generation Driver marks a pivotal advancement in fully autonomous operations, leveraging a custom, multi-modal sensing suite that integrates high-resolution cameras, lidar, and radar to enhance safety and efficiency across diverse environments, including extreme weather conditions.
Step 3.5 Flash: Open Frontier-Level Intelligence with 11B Active Parameters
Step 3.5 Flash is a sparse Mixture-of-Experts (MoE) model that combines a 196B-parameter foundation with 11B active parameters, optimizing for sharp reasoning and fast execution in agentic intelligence.
Show HN: Agent framework that generates its own topology and evolves at runtime
Hive enables the creation of autonomous, self-improving AI agents through a framework that dynamically generates workflows based on user-defined goals, eliminating the need for hardcoding.
The Problem with LLMs
LLMs are fundamentally flawed due to their reliance on plagiarism, as they generate content by appropriating existing works without proper attribution, raising significant ethical concerns about copyright violations and the integrity of creative output.
From specification to stress test: a weekend with Claude
Claude successfully built a distributed system with Byzantine fault tolerance, achieving thousands of requests per second and sub-100ms latency, all through conversational specifications without writing implementation code.
Lance table format explained with simple animations
Lance is a new file format that enhances random read optimization and allows for ad-hoc column additions without data duplication, positioning it as a successor to Iceberg and Delta Lake.
Show HN: 20+ Claude Code agents coordinating on real work (open source)
Lean 4 and Ensue Memory Network enable multi-agent collaborative theorem proving, enhancing efficiency in complex problem-solving by leveraging parallel processing capabilities.
Show HN: Pgclaw – A "Clawdbot" in every row with 400 lines of Postgres SQL
pgclaw is an open-source Postgres extension that allows users to integrate AI agents directly into their database tables, enabling both simple LLMs and more complex "OpenClaw" agents to operate as a column data type.
[R] I am looking for good research papers on compute optimization during model training, ways to reduce FLOPs, memory usage, and training time without hurting convergence.
Key areas of interest include mixed precision, gradient checkpointing, and optimizer efficiency, which are essential for optimizing compute during model training without compromising convergence.