Problem
To know the world one must construct it. -- Pavese
In 1972 Alan Kay (Kay, 2011, p. 11) conceptualised the Dynabook, a portable device that empower users of all ages to explore and learn. More than just a portable computer, the Dynabook was envisioned as a dynamic, interactive medium for learning, creation, and self-expression, that could adapt to users’ increasing skills and need.
As we progress into the century, software has become comoditised, serving as the engine of transformations that transcend every corner of our life. Simultaneously, we’ve seen exponential growth in machine learning (ML) systems’ capabilities, mainly through the general push of large language models (LLMs) into the mainstream. As these systems exihibit emergent properties of intelligence, how should we craft interfaces that promote agency and encourage a sense of personalisation through interactions, rather than providing a tool for automation?
Imagine you are an engineer who pursues creative writing as a hobby. You often curate topics and ideas from discussion on social media, then categorise them into themes for your arguments. There are plethora of tools out there that you can use to assist you with these tasks. For those inclined towards more adventurous endeavours, such as running customized models to meet specific requirements, you might find yourself in the land of auto-regressive models: GPTs and friends.
Auto-regressive models excels at surfacing machines’ internal representation of the world through a simple interface: given a blob of text, the model will generate a contiguous piece of text that it predicts as the most probable tokens. For example, if you give it a Wikipedia article, the model should produce text consistent with the remainder of said article.
This heuristic lays the foundation to the proliferation of conversational user interfaces (CUIs), which is obvious given that chat is a thin wrapper around text modality. Yet, CUIs often prove frustrating when dealing with tasks that require larger sets of information (think of support portals, orders forms, etc.). Additionally, for tasks that require frequent information retrieval (research, travel planning, etc.), CUIs are suboptimal as they compel users to unecessarily maintain information in their working memory (for no reason).
Given these challenges,
Link to originaltinymorphseeks to explore alternative interfaces for text generations models to extend our cognitive abilities. This means developing spatial and visual interfaces that allow for non-linear exploration of information and ideas, through writing.
The following includes a few shortcut for capstone portion of tinymorph:
- Checklists for progress tracking.
- Design contains specification for the project.
- Development Plan covers the development plan for the next 8 months of
tinymorph. - Hazard Analysis with regards to existential risks of
tinymorph. - Software Requirements Specification specifies the requirements for
tinymorph.
See also: References | Glossary
Navigation of the site
Shortcut Action ctrl+k search ctrl+n next search item ctrl+p previous search item ctrl+g graph
acknowledgements
This project is based on the work from EleutherAI, Linus for his general exploration in machine-native interfaces, Neel Nanda spearheading the field of mechanistic interpretability, Anthropic Interpretability Team, Google DeepMind Interpretability Team.