Plain-English explainers: tokens, context windows, hallucinations, fine-tuning, RAG and agents.
MIT Schwarzman's inaugural AI Educators Pilot trained 19 faculty to adapt machine learning teaching across disciplines using shared course materials.
vec2text shows embeddings can recover original text, with 92% exact 32-token recovery, challenging privacy assumptions in RAG vector databases.
AI gender bias appears in embeddings, face classifiers, language models, and image systems; 2016-2023 studies show it can be measured and mitigated.
LLM chatbots need purposeful dialogue; instruction stability can drop in long chats, so assess goal-directed UX beyond static benchmarks.
This article introduces reward misspecification in reinforcement learning: reward functions are only proxies for our true goals, and powerful optimizers exploit gaps between what we write and what we …
Since 2012, the computing power used to train the largest AI systems has grown more than 300,000-fold, doubling every 3.4 months—a pace far exceeding Moore's Law's two-year doubling period. Because gr…
This article explains OpenAI's GPT-2, a large-scale unsupervised language model trained only to predict the next word. Despite no task-specific training, it generated coherent paragraphs, set language…
This article reviews OpenAI's early experiment fine-tuning a 774M-parameter GPT-2 model using human feedback. Labelers taught the model to match their preferences, but narrow instructions led to unint…
CLIP is a neural network that learns visual concepts through natural language supervision rather than fixed-label datasets. Like GPT-2 and GPT-3's zero-shot capabilities in language, CLIP accepts cate…
DALL·E is a neural network that creates images directly from text descriptions. It can render a broad range of concepts expressed in natural language, combining distinct ideas to produce novel, realis…
OpenAI released new versions of GPT-3 and Codex that go beyond forward-only text completion. The models can now edit existing text and insert new content at specific points within a document. This ena…
Researchers trained AI critique models to identify flaws in AI-generated summaries, such as missing details and unsupported claims. Human evaluators shown these critiques caught far more errors than w…
OpenAI is introducing a new approach to mitigate bias and enhance safety in DALL·E 2, enabling the image generation model to create images of people that more accurately reflect the true diversity of …
Whisper is OpenAI's newly released open-source neural network for automatic speech recognition (ASR). It stands out by combining accuracy with robustness, achieving near human-level performance on Eng…
DALL·E has removed its waitlist, allowing anyone to start creating with AI immediately. Enabled by improved safety systems and lessons from real-world deployment, new users get quick sign-up, guided o…
This article introduces ChatGPT, a language model designed for natural dialogue rather than single-shot commands. Its conversational format lets it answer follow-up questions, admit mistakes, challeng…
OpenAI explains how ChatGPT's behavior is shaped through training, data, and guidelines, and outlines plans to improve it over time, offer users more customization, and invite public input on decision…
OpenAI and other leading AI labs have made voluntary commitments to safer AI development, covering safety testing, security measures, and transparency. These pledges create public accountability, shap…
A comprehensive guide helps teachers at all tech levels use ChatGPT in the classroom. It offers ready-made prompts for lesson planning, quizzes, and differentiated materials, explains how the AI gener…
OpenAI's GPTs let anyone build custom versions of ChatGPT by combining instructions, uploaded knowledge files, and capabilities like web browsing, DALL·E image generation, and code interpretation—no c…
