How I use LLMs

LLMs function as highly compressed, probabilistic versions of internet data that predict the next piece of text, rather than intelligent beings. To get the best results, users must treat these models as tools that require distinct strategies—such as selecting the right tier, utilizing external agents, and providing native context—rather than relying solely on their built-in knowledge.

• Core nature — Think of these models as "zip files" containing a compressed recollection of the internet that can hallucinate or hallucinate confidence when they lack information [08:24].

• Working memory — Understand that the "context window" acts as temporary storage for your chat; starting a new conversation wipes this memory to prevent confusion [06:54].

• Pricing tiers — Choose models based on your needs, as cheaper "mini" versions are faster but less capable, while paid, high-end models offer superior reasoning and creativity [19:01].

• Reasoning mode — Enable "thinking" or "reasoning" features for difficult code or math problems to boost accuracy, acknowledging that this requires longer wait times for the model to process [25:26].

• Search tools — Use internet search functions for recent events or niche topics that fall outside the model's fixed training data [33:53].

• Code execution — Utilize built-in coding environments for math or data visualization, but manually verify the code and assumptions, as the model can make mistakes [1:05:55].

• Voice interaction — Use voice modes to interact natively with the model, as speaking is often much faster and more natural than typing [1:23:27].

• Customization — Build custom mini-versions of these tools (Custom GPTs) to save time on repetitive prompts, such as formatting flashcards or translating languages [2:01:08].

• Dedicated software — Switch to specialized coding editors for professional development instead of web browsers to maintain context across your entire project [1:14:51].

• Visual input — Upload screenshots or diagrams to interpret complex data, such as blood test results or nutritional labels, rather than trying to describe them in text [1:42:07].

• How do "thinking" models differ from standard language models in their approach to problem-solving?

• What are the benefits of using dedicated software for coding rather than web-based interfaces?