AI
Artificial intelligence: LLMs, agents, diffusion, vision, NLP and the latest from top labs.
65 published articles
AI Agents
AI agents can't finish a Java migration without the build server telling them to stop
IBM Research introduces ScarfBench, an open benchmark for evaluating AI agents on enterprise Java framework migration. Early tests reveal that frontier agents are systematically overconfident about their own results, that configuration layers dominate effort, and that environment issues like Docker caches regularly derail migrations even when code transformations succeed.
2026-07-06
Artificial Intelligence
MiniMax's product blitz: new models for code, music, and video in one broad refresh
Chinese AI startup MiniMax launches MiniMax M3, Hailuo 2.3, MiniMax Code, and new speech/music models, broadening its product lineup in a competitive landscape.
2026-07-06
Model Release
Claude Mythos 5 already found 10,000 critical bugs, and Anthropic is terrified of what it can do
Anthropic's Claude Mythos 5, a powerful new model for cybersecurity and biology, has already found over 10,000 critical software vulnerabilities through Project Glasswing. Access is limited to vetted partners, with pricing at $10 per million input tokens.
2026-07-06
Open AI Infrastructure
NSF OMAI goes live: Ai2's fully open AI cluster is a bet against closed research
Ai2 just turned on a new AI cluster under the NSF OMAI project, one that makes models, tools, and processes fully open. Every GPU hour counts more when it fuels research across language, multimodal, and scientific domains.
2026-07-06
Artificial Intelligence
MiniMax M3 broke the open-weight ceiling with a 9.4x CUDA speedup and zero human help
MiniMax M3 delivers a 9.4x CUDA kernel speedup, beats Opus 4.7 on BrowseComp, and autonomously replicated an ICLR paper. All in an open-weight package.
2026-07-06
Model Evaluation
Ai2's olmo-eval gives LLM developers a microscope for every checkpoint
Ai2's olmo-eval brings per-question diffs and modular benchmarks to active LLM development, helping researchers tell real progress from statistical noise.
2026-07-06
Artificial Intelligence
Kog's Laneformer 2B hits 3,000 tokens/s by making latency the whole point
Paris-based startup Kog releases Laneformer 2B, a latency-first Transformer that hits 3,000 tokens/s on AMD MI300X. The model uses Delayed Tensor Parallelism (DTP) to mask communication overhead and was trained from scratch on 6 trillion tokens. Open-source weights are on Hugging Face.
2026-07-06
AI Research
Your AI agent passed the test by accident. Now there's a rubric for that.
SkillCoach is a self-evolving rubric framework that evaluates and improves agentic skill-use by analyzing skill selection, following, composition, and reflection processes, providing better supervision than outcome-only metrics.
2026-07-06
AI Research
The real bottleneck in desktop AI agents isn't the model. It's the skill library.
A matched execution-layer benchmark of 440 desktop tasks shows GUI agents reaching a 59.1% pass rate versus 48.2% for CLI agents. Skill augmentation raises CLI success to 69.3%, indicating that the main bottleneck is skill coverage, not raw capability.
2026-07-06
Reinforcement Learning Research
OPID gives language agents a reward signal dense enough to ditch external memory
OPID extracts hierarchical skill supervision from completed on-policy trajectories, providing dense token-level guidance for language agent training without external memory. Experiments on ALFWorld, WebShop, and Search-based QA show improved performance and sample efficiency over outcome-only RL and existing skill-distillation methods.
2026-07-06
Nature Neuroscience paper
Microsoft's new method turns black-box brain AI into readable theories
GCT translates uninterpretable LLM-based brain models into short phrases like 'food preparation' or 'location names,' then uses an LLM to write stories that causally test those explanations in real subjects. The method promises to bridge predictive AI and human-readable scientific theory.
2026-07-05
Deep Learning
M3D and Real-Guidance Bring Dataset Distillation to High-Resolution Realms
Dataset distillation has long been stuck on low-res benchmarks, but a new approach called M3D changes that. By combining multi-scale matching, a data manifold prior, and a Real-guidance strategy, it scales to ImageNet-1K at 128×128 resolution, achieving 68.5% top-1 accuracy with just one image per class and cutting memory usage by ten times.
2026-07-05