Experiments, failures, dalliances, and plenty of unfinished hacks. Leaning hard into the "make myself obsolete ASAP" role.
An AI-powered assistant that can use tools to accomplish tasks. Inspired by HuggingGPT, exploring how far GPT can go with access to APIs like Twilio, Gmail, and Calendar. Uses structured communication between tools and handles freeform natural language requests.
Audio-book replacement: ML-rewrite public domain stories to modern versions (Shakespeare, Sherlock Holmes) with modern references, then generate a highly-detailed play version with dialogue, emotion, music notes, laid out in 3D audio space. Copyright-free audio books of timeless stories made modern.
Inspired by the Reflexion paper, a tool that can write, edit, and speak tech talks tailored to an audience. Two bots - author and audience - critique and incorporate each other's feedback, then read it out in my generated voice. Making myself obsolete ASAP.
A new UI paradigm where UIs build themselves as ephemeral, on-demand, iterable programs that appear out of nowhere, get used, then disappear. The future: ~10% fixed UIs, ~40% conversational UIs, ~50% long-tail on-the-fly generated UIs.
A way to ask questions interactively while reading research papers, tracking conversations across papers down to the specific part that triggered the question. Added polish on the overall experience to make it more and more useful for paper reading.
A nod to the old ELIZA program: a GPT-guided diary to explore what's on your mind, expand it, and write out a narrative as a 1st-person journal entry. Decompress while driving, on the couch, or walking. The most personally engaging project so far.
A pull request tool using GPT-4 to make suggestions on GitHub PRs. Leave a comment and the tool uses the entire file context to create a suggestion that can be committed directly. Make suggestions and get a first pass without switching context.
Summarize HN discussions and extract overall sentiment and keywords automatically. Using the extracted summary + sentiment, generate an embedding and store it with like/dislike data. Use embeddings to predict how likely you'll enjoy other stories on the HN frontpage.
A Swift project for a GPT-4 powered dungeon story where the setting is made up on the spot. The storyteller tracks HP, mana, and stats as you decide what to do. At each step, a comic book image is generated for the scene via Stable Diffusion.
A tool to explore startup ideas by simulating YC interview prep. Create a mind map, and GPT asks questions you may not have thought of. As you answer, the tool follows up the chain, leading to better and better questions. A 90-minute hack that shows how powerful modern LLMs are.
Part of a series where I lean hard into automating away my own tasks, hoping for a phoenix-like rebirth as a new, more powerful, and more useful version of myself on the other side of being automated away.
These are experiments and explorations - some work, many don't, all teach something interesting.
More experiments and demos on YouTube
Watch on YouTube