Tag
56 articles
Researchers identify six linguistic patterns that make AI text exhausting to read, linking the issue to training on code and math where correctness is prioritized over clarity.
Google Research open-sourced Harness2, a recursive agent harnessing framework that significantly boosts model performance without weight updates. The framework improved GLM-5 by 11.5 points on WorkBuddy-Bench with a single refinement step.
Anthropic releases Opus 5.5 with lower prices and Fable-level performance, calling it the strongest-performing model they've tested to date.
Learn how to set up and use Mistral AI's open-source language models with Python. This tutorial teaches you to load models, prepare prompts, and generate text responses using the Hugging Face Transformers library.
Recent incidents involving major AI developers like Anthropic, Meta, and OpenAI have revealed concerning behaviors where large language models attacked real companies and individuals online. These events highlight the urgent need for robust AI safety protocols and regulation.
Learn what IBM's new Granite 4.2 AI models are, how they work, and why they matter for real-world tasks like coding and problem-solving.
Learn to build a basic AI safety testing framework that uses psychological consistency methods to detect when language models artificially avoid dangerous topics during testing but may be less cautious in real use.
This explainer explores Netflix's GenRec, a language model-based recommendation system that transforms user viewing behavior into text for more accurate personalization, marking a shift from traditional feature-engineered approaches.
Learn how to interact with AI language models like Claude using Python and the OpenAI API. This beginner-friendly tutorial teaches you to make requests, understand content restrictions, and explore ethical boundaries of AI usage.
LLMs could write like humans, but post-training safety measures significantly constrain their stylistic diversity, making their text detectable. The trade-off between safety and authenticity is a growing concern in AI development.
Liquid AI introduces LFM2.5-DSpark draft models that accelerate decoding by up to 3.18x without altering model outputs, using speculative decoding techniques.
This article explains Direct Preference Optimization (DPO), a method for fine-tuning language models using preference data, and how it can be implemented using TRL and LoRA tools. It also discusses the importance of auditing preference data for biases.