GPT-4 in ChatGPT is gaining a more flexible way to handle requests: it can choose among several available tools, while users can still select models themselves. The update also makes it possible to upload PDFs and other file types for discussion, extraction, or explanation.
File uploads move into the broader ChatGPT experience
Users can provide a document and ask ChatGPT to work with its contents. That could mean discussing information in a PDF, pulling out particular aspects, or asking for an explanation. The change is useful because it brings document-based tasks into a more direct interaction with GPT-4.
The upload capability itself is not new. It was previously available through Advanced Data Analysis, which users had to choose in the model switcher to access the file upload button. The update described by The Decoder makes this capability part of a wider GPT-4 experience.
That shift changes the steps a user needs to take. Instead of first identifying a specific model for a file task, users can upload the material and describe what they want to do with it. The model can then determine which available capability fits the request.
GPT-4 can route requests to different tools
The update pairs file uploads with automatic tool selection. In this context, tools are alternative models or capabilities: GPT-4 with Browsing, Advanced Data Analysis, and DALL-E 3. ChatGPT can choose among them according to the task it understands from the user’s request.
For example, the system is expected to recognize when a request involves generating an image or creating a chart, and switch to DALL-E 3 or Advanced Data Analysis as appropriate. Users retain the option to choose models manually, but they may not need to make that choice for every task.
This approach treats GPT-4 as a coordinator for different capabilities. A single conversation can start with a request, then rely on a suitable tool when the task calls for something beyond the model’s ordinary response. The practical promise is less model-switching for users; how well the system makes those choices is a separate question that the source does not answer.
A step toward models that use tools
The article frames this tool selection as part of OpenAI’s broader aim to make language models useful beyond their trained capabilities. A model can call on tools in areas where it would otherwise be less capable, rather than handling every request through its own generated text.
The source also points to a study conducted this spring by OpenAI, OpenResearch, and the University of Pennsylvania. It predicts that, with access to tools, large language models like GPT-4 could affect at least 10 percent of the work of about 80 percent of U.S. workers. For 19 percent of workers, the study predicts that at least 50 percent of tasks could be affected.
Those estimates concern the potential reach of language models with tools. They help explain why tool use matters beyond a convenience feature in ChatGPT: if models can draw on specialized capabilities, more kinds of work may be within reach. The figures describe predicted impact, not a report that those changes have already occurred.
Reports of newer information remain uncertain
Some ChatGPT users have reported GPT-4 access to current information through April 2023 or September 2023. Earlier disclosures placed the end of training data at September 2021 or January 2022. The reports do not line up: ChatGPT has provided different cutoff dates to different users.
The update being rolled out to early users may be related to a training data update, but that connection is not confirmed in the source. A Reddit leak also suggests OpenAI is training a model called "copilot-gpt-4-2," with information current until March 2023. The leak is a separate indication, not confirmation of what the ChatGPT update contains.
The OpenAI developer conference on November 6th was expected to offer more details about the status of GPT-4 models. Until then, the reported cutoff dates and the possibility of refreshed training data remain uncertain. The clearer change is the direction of the product update: GPT-4 can work with uploaded documents and choose among tools, while manual model selection remains available.