Boring AI News

OpenAI introduced GPT-Rosalind, a specialized model for drug discovery and biological research with improved reasoning, tool use, and integration with scientific data sources.

Qwen3.6-35B-A3B is a 35B-parameter MoE model with 3B active parameters, combining strong agentic coding, multimodal reasoning, and open weights. It outperforms its predecessor on key coding benchmarks and is available on Qwen Studio, via API, and for self-hosting.

Technical guide to fine-tuning Google’s Gemma 4 models, covering VRAM requirements, bug fixes for KV-sharing and gradient accumulation, and training recipes for SFT, vision, audio, and GRPO.

IBM Research evaluates multi-step agent reasoning and tool use with an executable benchmark spanning thousands of APIs and documents, highlighting consistent performance gaps and common failure modes.

Canon structures its search pipeline as a DAG to run classification, localization, retrieval, and ranking concurrently. The design supports independent subsystem changes, parallel execution, durable retries, and clearer pipeline introspection.

HY-World 2.0 is an open-source multi-modal world model framework for generating and reconstructing 3D worlds from text, images, multi-view images, and video. It produces editable 3D representations such as meshes and Gaussian splattings, and currently includes WorldMirror 2.0 inference code and model weights for feed-forward 3D reconstruction.

This work studies how to split LLM prefill and decode across different datacenters by sending KV cache over normal Ethernet. It uses selective offloading for long-context requests, with bandwidth-aware scheduling and cache-aware placement, to handle bursty traffic and uneven workloads. In experiments with a 1T-parameter hybrid model, this design improves serving throughput while using only modest cross-datacenter bandwidth.

OpenAI updated its Agents SDK with a model-native harness for cross-file and tool workflows, plus sandboxed execution for safer task handling.

Building timelines in documents usually means wrestling with HTML or settling for static images. This Quarto extension lets you create customizable timelines in HTML docs and RevealJS slides using simple div syntax. It supports vertical layouts, grouping, CSS styling, and animated reveals for presentations.

A research post on CowCorpus and PlowPilot, a dataset and intervention-aware web agent system that predicts when users want to take over, improving user-rated usefulness by 26.5% over a fully autonomous baseline.

PrismML’s 1.58-bit Ternary Bonsai models come in 1.7B, 4B, and 8B sizes, using ternary weights to cut memory use by about 9x versus 16-bit models while posting strong benchmark results. The 8B model reaches a 75.5 average benchmark score at 1.75 GB, with fast on-device throughput and improved energy efficiency on Apple hardware.

LLM agents get instructions from many places, but these instructions can have different authority and can conflict. This work proposes a many-level hierarchy for deciding which instruction to follow, and introduces a benchmark with up to 12 conflict levels across 853 agent tasks.

Results show current top models handle this badly, with about 40% accuracy when conflict becomes more complex. The study highlights need for better fine-grained instruction priority in agent systems.

Microsoft releases MAI-Image-2-Efficient, a version of MAI-Image-2 designed for high-volume and real-time image generation with faster performance and greater GPU efficiency.

Claude Opus 4.7 is now generally available with improved performance on complex software engineering, higher-resolution vision, and more reliable long-running work. The release also adds cyber safeguards, new effort controls, and Claude Code and API updates while keeping Opus 4.6 pricing unchanged.

Cursor describes a multi-agent system that optimized 235 CUDA kernels for NVIDIA Blackwell GPUs, delivering an average 38% speedup and more than 2x improvement in some cases.

This work generates long camera-controlled videos for scene walkthroughs, then turns them into 3D worlds for real-time rendering and simulation. It reduces two main problems in long generation: forgetting places seen before and gradual visual drift over time. The method keeps 3D geometry to retrieve relevant past views and trains on its own degraded outputs so it can correct errors during exploration.

AI agents represent a real shift beyond chatbots, with the potential to change how people work, build and create. A TED2026 talk traces the release of OpenClaw and the rapid growth of an open-source project.