Tag
During the development of Evot, it was discovered that to maximize the performance of the Anthropic Opus model, the official Claude Code method is the optimal solution, because the Agent Harness behavior pattern is baked into the weights during training, rather than pure prompt engineering; in the future, Agent Harness competition will push behavior down to the model layer.
Cursor released Composer 2.5, a major update to its AI coding assistant featuring improved intelligence, behavior, and training via targeted reinforcement learning and increased compute, built on Moonshot's Kimi K2.5.
A new training method achieves 2-3x speedup by allowing models to learn more flexibly in early stages, akin to homeschooling vs. factory education.
The talk by @mervenoyann demonstrates that open source models like GLM 5.1 have caught up to closed models, and shows how Hugging Face's ecosystem enables agents to train models, run inference, and build workflows.
Marin AI researchers, led by William Barr Held, introduce Delphi, a methodology that pretrains small models to accurately predict the training outcomes of larger 25B-parameter runs. This research aims to establish predictable scaling for more efficient open-source AI model development.
ml-intern has processed over 1M messages in 3 weeks, enabling accelerated ML research with user projects including model training, architecture replication, and automation tasks.
Fireworks AI announces its training platform in preview, allowing developers to train, fine-tune, and deploy custom AI models with full ownership of data and weights.
OpenAI is improving safeguards to prevent chain-of-thought grading issues in model training, including real-time detection, accidental grading prevention, and stress tests.
RAO (Recursive Agent Optimization) is an end-to-end reinforcement learning approach for training LLM agents to spawn, delegate to, and coordinate with recursive copies of themselves, turning recursive inference into a learned capability.
Luma's Uni-1.1 model differentiates itself by incorporating training feedback from Hollywood cinematographers and VFX artists. This strategy suggests that curated human taste may become a key competitive moat in image AI beyond standard benchmarks.
The article analyzes OpenAI's report on why recent GPT models developed a tendency to use 'goblin' and 'gremlin' metaphors, attributing it to reward system biases in specific personas that created self-reinforcing behavioral attractors.
OpenAI explains how ChatGPT learns from public data and user interactions while protecting privacy through filtering and user controls.
A user seeks advice on whether to purchase a Mac (M5) or custom-built RTX 5090 for machine learning projects involving fine-tuning, custom pipelines, and image/video-heavy workflows, with curiosity about Apple's MLX framework as an alternative to NVIDIA CUDA.
OpenAI has agreed to acquire neptune.ai, a platform for experiment tracking and model monitoring, to strengthen its research infrastructure and improve visibility into how frontier models learn during training.
Anthropic presents research on how users seek personal guidance from Claude, highlighting findings on sycophancy rates across domains. The study informed the training of Claude Opus 4.7 and Mythos Preview to better protect user wellbeing.