The REAL potential of generative AI

Habib describes language models as systems trained to predict the next word from large collections of text. Scaling model size and training data improves prediction, and can lead models to acquire knowledge and reasoning abilities that help them complete more complex tasks.

Habib says models can confidently produce false information because they are trained to predict text, not necessarily to be honest. Providing factual context in prompts can reduce hallucinations. Fine-tuning with examples and human preferences can also adapt a model’s responses, tone, and behavior to a particular task.

Developers can fine-tune with background material such as company emails, or with feedback gathered from product use. For a sales email tool, whether a user sends a draft, edits it, or rates it can help improve the model. Humanloop helps teams manage prompt versions, evaluate subjective results, and customize applications for their users.

Habib says tools such as GitHub Copilot already help developers write code faster, including experienced developers who can review and edit suggestions. Over time, developers might spend more effort on product specifications while models handle boilerplate. Longer context windows and models that take actions could expand what applications can do.

Habib sees major potential for startups as models make capabilities that once required research teams more accessible. He also warns that AI could disrupt work, reflect biases in its data and creators, and raise existential safety concerns. He argues that increasing capabilities make careful work on safety and social consequences more urgent.