715: Would You Like a LLM With Your Browser?

The hosts discuss the integration of large language models (LLMs) into browsers, mentioning a recent API that allows web developers to download and run models locally without internet access. They explore the idea of browsers downloading models like Chrome and Edge, and how these models are being embedded into browsers, with a focus on privacy and speed. The conversation includes an overview of the evolution of the API over the past year, splitting into multiple APIs such as prompt, summarizer, writer, and rewriter, each tailored for specific tasks. The hosts also share their experiments with creating chatbots that respond in different styles, including emojis and R2D2 sounds, emphasizing making models more conversational and personable.

The discussion shifts to the potential implications of having AI models integrated into browsers. Concerns are raised about model calcification—favoring one model over others, which could hinder innovation and flexibility. They also examine Google's approach with Gemini and the API, questioning how standardized and open these models truly are. Issues around model training data, transparency, and the ethical sourcing of datasets are debated, highlighting the mystery surrounding where AI models obtain their training material. The conversation considers how these models might shape web development, with a focus on potential for browser-specific APIs, and whether the web will fall behind native applications in AI capabilities.

A key theme is the importance of preserving websites and digital content as they fade or are taken offline. The hosts reflect on the transient nature of the internet, citing statistics from the Internet Archive revealing that around 25% of web pages from the past decade are disappearing. They express concern over websites, including personal projects, going offline and the challenges of archiving content, especially media like MP3s and videos. Ideas are proposed for creating a 'website salvage' service that helps people permanently store and host their websites in a responsible, low-maintenance way. The discussion also touches on the impermanence of web content, the costs associated with hosting and archiving, and how to responsibly ensure digital legacies are preserved for future generations.

The hosts discuss AI’s potential for societal good, emphasizing its capacity to empower individuals with disabilities and improve accessibility. They share anecdotes about how AI can help people with limited mobility to communicate and perform tasks more efficiently. The concept of 'purple washing' is examined—using societal causes like accessibility or DEI as marketing tools to bolster company reputation. The debate centers on whether these initiatives are sincere or strategic marketing, citing examples like autonomous vehicles being promoted as accessibility solutions. The conversation underscores the importance of ethical AI development, responsible usage policies, and the need for transparency about training data and motivations of tech companies.