An LLM is a lossy zip file of the internet; tools and context make it useful.
Yes. A new chat wipes the context window back to zero. Leftover tokens cost you twice: the model can be distracted by them, which lowers accuracy, and every extra token makes sampling the next one slightly slower and more expensive. Treat that window as precious working memory.
The free tier gets GPT-4o mini, a smaller model with fewer parameters that writes less well and hallucinates more. Plus, at $20 per month, gets 80 GPT-4o messages every three hours. Pro, at $200 per month, adds unlimited GPT-4o, o1 Pro mode and deep research.
Pre-training compresses the internet into the parameters and costs many tens of millions of dollars over roughly three months, so it is rerun rarely. GPT-4o was pre-trained many months, maybe a year, earlier, and that date is its knowledge cutoff. Recent facts need tool use instead.
Thinking models, tuned with reinforcement learning, pay off on difficult math and code, where they raise accuracy, but they can emit tokens for several minutes. For travel advice there is no benefit. Start with the fast non-thinking model and switch when the response looks weaker than it should.
ChatGPT recognised the question as recent and searched by itself, and Grok searched too; Perplexity does this by design. Claude answered from its April 2024 last update and has no search tool. Gemini 2.0 Pro Experimental lacks real-time access, while 2.0 Flash answered and showed sources.
It is internet search plus thinking, rolled out over tens of minutes, listed under ChatGPT's Pro tier at $200 per month. On the CA-AKG question it thought for five minutes across 27 sources and produced a cited report. Perplexity and Grok have versions; treat any report as a first draft.
Without a tool the model simply predicts the digits. ChatGPT wrote Python and Claude wrote JavaScript, and both were correct. Grok 3 has no interpreter, did it in its head and ended in 060 instead of 120. Gemini 2.0 Pro got one right, then failed a harder one.
Put the actual text into the context window. For The Wealth of Nations, written in 1776, the chapter comes from Project Gutenberg, gets pasted into Claude, and is summarised before questions follow while reading. PDFs can be dragged in, though a 30 megabyte paper exceeded Claude's chat length limit.
The phone app has a microphone icon that transcribes speech into text, and about 80% of queries on the phone go that way. The Mac desktop app has no such icon, so a system-wide dictation app is used, SuperWhisper bound to F5. Product and library names are still typed.
Handing control to Cursor's Composer, an autonomous agent over your codebase, telling it what to do and hoping it works. It runs commands and edits many files at once: asked for confetti on a win, it installed React Confetti, rewrote App.tsx, added CSS and a sound file.
“I am a one terabyte zip file.”
“And so think of the tokens in the context window as a precious resource.”
“And so again, use it as a first draft.”
“And vibe coding just refers to giving control to Composer and just telling it what to do and hoping that it works.”
“Practical Uses of ChatGPT.” https://www.youtube.com/watch?v=EWvNQjAaOHw. Transcript summary by WhipScribe, https://whipscribe.com/transcript-pages/practical-uses-of-chatgpt.html.