Summary
Highlights
The speaker introduces how LLMs have dramatically increased their productivity, learning speed, and task completion. They will demonstrate how LLMs are used for dictation, search, synthesizing YouTube videos and podcasts, learning Japanese, and automating tasks.
Typing speeds are limited, but dictation allows for over 200 words per minute. The accuracy of AI-powered dictation (based on OpenAI's Whisper model) is highlighted as super accurate and fast. The speaker uses dictation for long responses like emails and interacting with LLMs.
Whisper Flow is introduced as the primary app for pure transcription. It uses a keyboard shortcut to transcribe speech directly into the current application, formatting the text appropriately (e.g., for emails). This is convenient for drafting emails or detailed prompts for ChatGPT.
A second app, Super Whisper, processes voice input through a custom prompt to summarize or clean up speech. This is highly useful for taking readable notes in Obsidian (e.g., meeting notes, SOPs) even if the initial spoken input is rambling. Super Whisper can also run a local model for offline use and privacy.
Perplexity is presented as the fastest and most efficient search tool, utilizing a voice model for quick queries. It offers better search results than ChatGPT for general web, social media (YouTube, Reddit), and academic papers. It's the most used LLM by the speaker.
The video demonstrates two methods for summarizing YouTube videos: 'YouTube Summary AI' extension for quick overviews of shorter videos (around 15-20 minutes), and 'Notebook LM' for super long videos like podcasts or lectures. Notebook LM, based on Gemini, acts as an AI research assistant, synthesizing information from multiple sources (including PDFs and e-books) and answering questions.
The speaker explains how they use a custom GPT called 'Sensei' to learn Japanese efficiently. This custom GPT is configured to explain Japanese words in context (not just direct translations) and can process images of text. It's personalized to the user's prior knowledge, even incorporating Chinese similarities. Custom GPTs can be tailored for various specific tasks beyond language learning.
GPT-4 includes a beta feature for scheduled tasks, allowing users to receive daily or weekly emails with specific information. Examples include getting weekly summaries of local events or concerts, custom newsletters with book/documentary recommendations, or reminders of personal goals. It's particularly effective for information that is constantly updated, like real estate listings.
The concept of 'Vibe coding' (coined by Andrej Karpathy) is introduced, enabling users to create full applications using LLMs without writing or even seeing code, primarily through voice commands. The speaker demonstrates building a YouTube channel analysis tool and a playable Tetris game using vzero.dev, highlighting the ability to fix bugs and add features by simply explaining them to the AI.
The speaker concludes by reiterating that LLMs have transformed their productivity workflow, building upon previous productivity tips. They encourage viewers to explore these tools.