Artificial Intelligence (AI) is revolutionizing research by enabling researchers to delegate complex tasks to AI assistants. These tasks include explaining theories, summarizing reports, comparing arguments, and identifying questions for further investigation. However, as research demands grow, a new question arises: should researchers use conversational AI applications, command-line interfaces (CLI), programming environments, or locally hosted models?
The choice depends on the nature of the research problem. For preliminary research, traditional AI applications may suffice. For tasks requiring the processing of vast amounts of documents, running statistical models, simulations, or building repeatable research pipelines, a terminal-based workflow offers greater flexibility. Yet, neither is inherently more intelligent.
An AI application is not just an AI model; it's an interface providing access to one or more models, along with additional capabilities such as web search, document analysis, memory, code execution, and other tools. These applications can choose between models based on the complexity of the task. For instance, OpenAI's GPT-5 system combines faster models with more deliberate reasoning models and a routing mechanism that selects between them based on the nature of a conversation, its complexity, and tool requirements.
On the other hand, a CLI allows researchers to connect an AI model to local files, scripts, databases, programming environments, and other tools. Some command-line agents even let users specify the model and reasoning effort directly. However, running a CLI on a computer does not necessarily mean the AI model is running on that computer. The terminal can operate locally while the model generates its responses on a remote server.
Before deciding how to access an AI model, it's crucial to understand what it has been trained to do. Large language models are typically developed through pre-training and post-training stages. During pre-training, a model learns statistical patterns in large collections of text, code, and other forms of information. Post-training exposes the models to instructions, demonstrations, preference feedback, and other forms of training to make their responses more useful and appropriate.
Developers use various sources for training data, with complete datasets rarely disclosed by commercial providers. OpenAI's GPT-5 system card describes three broad sources: publicly available internet material, information obtained through third-party partnerships, and information supplied or generated by users, human trainers, and researchers. Anthropic's Claude Opus 4 and Sonnet 4 systems use a proprietary mix of publicly available internet information, third-party data, contributions from data-labelling services and paid contractors, information from users who opted into training, and internally generated data. Both companies implement filtering processes to improve data quality and address potential risks.