GPT Updates Cover Fine-Tuning, Editing, New Models and Text-Image Links

A batch of GPT-focused updates describes model releases, customization options and applications. GPT-3 can be customized for an application with a single command, and fine-tuning GPT-3 is said to power and scale done-for-you video creation. New versions of GPT-3 and Codex can edit or insert content into existing text rather than only completing it. The updates also state that more than 300 applications deliver GPT-3-powered search, conversation, text completion and other advanced AI features through an API.

The GPT-2 material is presented as a staged release. A 774 million parameter GPT-2 model was released after a small 124M model in February and a medium 355M model in May, following research with partners and the AI community into the model's potential for misuse and societal benefit. The 774M model was also fine-tuned using human feedback for various tasks, matching the preferences of external human labelers, though those preferences did not always match the researchers' own. The largest GPT-2 version, with 1.5B parameters, is described as the final model release of the staged process and was accompanied by code and model weights to facilitate detection of outputs from GPT-2 models.

GPT-6 entries introduce Sol and Luna, two models described as bringing frontier intelligence to everyday work with different balances of capability and cost. Harvey turns legal context into stronger drafts with GPT-6 Astra, which produces more structured, context-aware legal documents, freeing lawyers to focus on strategy. With GPT-6 Astra, invideo plans edits with greater precision, improves color correction and grading threefold and produces 50 custom effects in one day. GPT-6 Astra also allowed Parallel's agents to research and synthesize labor-market data in half the time and at half the cost versus prior models.

Other updates link language and vision. Image GPT reports that the same transformer model can generate coherent image completions and samples when trained on pixel sequences, and that sample quality is correlated with image classification accuracy. CLIP is introduced as a neural network that efficiently learns visual concepts from natural language supervision and can be applied to any visual classification benchmark by providing the names of visual categories, similar to the zero-shot capabilities of GPT. A further update describes language models trained to follow user intentions better than GPT-3 while also making them more truthful.