332 community available in the ChatGPT directory
Community discussions, tips, and shared experiences from users of this platform. Join the conversation, ask questions, and share what you've built.
I use multiple AI tools for work and also my side projects, and the annoying part was to track my costs and token usage across tools. Everytime I had to visit each tool and its respective usage setting to check it and I was losing patience and also was getting hit by surprise limits<p>Now I know that there are already free/open-source trackers for Cursor or Claude usage, and they are useful if that is all you need.<p>My problem is broader as I wanted one small place to see tokens, spend, subscriptions and limits across the AI tools I actually use. I was really tired of switching tabs and apps to check the usage.<p>So I built Tokens 4 Breakfast<p>It sits in the Mac menu bar and helps track usage across Cursor, Claude, ChatGPT, OpenAI API, Copilot and other providers. You can also set limit reminders and focus mode to always stay under the budget.<p>Check your usage insights per project and per model in Claude Code to get a even better understanding of where your tokens are flowing
A Wild Idea: Abstract Reality with Ontology<p>Background Large language models (LLMs) debuted with GPT-3 back in June 2020. After roughly five to six years of development, I believe the technology is still in its infancy, with massive room for improvement. A major priority is building powerful models that run on fewer resources, especially capable models that can operate smoothly on CPUs.<p>More crucially, the engineering ecosystem around LLMs is also at an early stage. Many unresolved challenges remain. Among them, hallucinations have become a major bottleneck, greatly limiting LLMs' adoption in real-world production scenarios.<p>I’ve come to a thought: relying purely on natural language to interact with LLMs — the so-called prompt engineering — may have been a wrong direction from the very start.<p>Fundamentally, an LLM is a mathematical computation system that takes linguistic tokens as its basic computing units. Problems like hallucinations and alignment issues are not
WSJ Anthropic is calling for top artificial intelligence labs to weigh slowing the pace of development, suggesting that AI systems are advancing so rapidly that they may soon be able to improve themselves without human intervention in ways that could pose significant societal risks.<p>The ability to slow global AI development would “likely be a good thing,” the company said Thursday in a blog post that disclosed internal data documenting how quickly its most advanced models are improving.<p>The post, written by the head of its internal research institute and head of policy, noted that model advances appear to be on a path toward “recursive self-improvement,” when AI systems can improve on their own without human intervention. Some AI insiders have seen that threshold as a potential marker of danger and enormous societal upheaval.<p>“We believe it would be good for the world to have the option to slow or temporarily pause frontier AI development to enable societal structures and alignme
Not sure if this is happening to others, but sometimes when I paste an image, I get an answer completely different from what I attached. It also sometimes responds in a different language. I'm sure it's not my image, since I ask it to explain what it sees and it describes something completely different. What's going on?
Tested the new Codex agent on 10 real tasks from our backlog. It spins up a sandboxed environment, reads your codebase, makes changes, and runs tests. Compared to Claude Code: Codex is more autonomous but less transparent about what it's doing. Claude Code lets you watch each step and intervene. For straightforward tasks (add a feature, fix a bug), Codex is faster. For complex refactors where judgment matters, Claude Code's interactive approach wins.
Been testing GPT-5 since launch and the improvement over GPT-4o is significant. Reasoning is noticeably stronger — it catches logical fallacies I embed in prompts that GPT-4o missed. The memory feature actually works now, recalling preferences from weeks ago. Code generation is competitive with Claude Sonnet. The only downside: it's more expensive and slower. For complex tasks, worth it. For quick queries, 4o-mini is still better value.
After months of iteration, here's the system prompt structure that consistently gets GPT-4o to follow complex instructions: 1. Role definition (one sentence) 2. Output format (explicit structure with examples) 3. Constraints (what NOT to do — this is crucial) 4. Reasoning approach (step-by-step for complex tasks) 5. Error handling ("If you're unsure about X, say so instead of guessing") The key insight: GPT-4o responds much better to negative constraints than positive instructions. "Never use bullet points" works better than "Write in paragraph form."
Started using ChatGPT voice mode during walks and it's become my primary brainstorming tool. The natural conversation flow helps me think through problems in a way that typing doesn't. Last week I talked through an architecture decision for 30 minutes and it raised three concerns I hadn't considered. The voice quality is incredibly natural and it picks up on context cues like hesitation. The future of AI interaction is voice, not text.
OpenAI quietly upgraded the Code Interpreter backend and the difference is night and day. It now handles 100MB+ files without choking, generates publication-quality charts with proper formatting, and the Python environment includes scikit-learn, statsmodels, and other ML libraries. Uploaded our company's sales data and it built a forecasting model, generated an executive summary, and created a PDF report — all in one conversation.
Hello HN, I have become a great fan of the containers feature in ChatGPT (aka python tool / data analysis) in the last year. For reference, see this article by Simon Willison: https://simonwillison.net/2026/Jan/26/chatgpt-containers/<p>To my mind, this is simultaneously the most powerful <i>and</i> underrated feature of ChatGPT / OpenAI. This thing is like an uber-monster-data-chainsaw-nailgun with a whole galaxy of possible applications.<p>For myself, I've had fun using it to OCR documents to load into a vector store for deep research and semantic search. Or generating sprite sheets with the Gemini API (OpenAI is not good at this for some reason, even gpt-image-1.5) and then using the container tool to cut them up and turn them into animated gifs. Then making a little movie out of those gifs with ffmpeg, which also ships in the container.<p>There's goodies in there like xgboost that let you train a classical ML model, there's
Tested over 200 custom GPTs. The ones actually worth your time: 1. Consensus — searches 200M academic papers 2. Canva — generates designs without leaving chat 3. Zapier AI Actions — connects to 6000+ apps 4. Data Analyst — handles CSV/Excel files natively 5. ScholarAI — full-text research papers with citations 6. Video Insights — analyzes YouTube videos 7. Diagrams — creates technical diagrams from descriptions 8. Code Copilot — specialized coding assistant 9. Wolfram — computational knowledge engine 10. AskYourPDF — analyzes and chats with PDFs Most custom GPTs are wrappers around basic prompts. These actually add real functionality.
Six months into our ChatGPT Enterprise deployment. What worked: customer support team productivity up 35%, legal team document review time cut in half, engineering team uses it daily for code review. What didn't: marketing team found outputs too generic, finance team concerned about accuracy of numbers, HR stopped using it after a policy interpretation was wrong. Key lesson: AI amplifies good processes and exposes bad ones. You need training and guardrails.
Used ChatGPT with my university students for an entire semester. What works: Socratic tutoring (asking it to quiz you and explain wrong answers), debugging code with explanations, summarizing research papers, brainstorming essay outlines. What fails: generating essays directly (students learn nothing), doing math without showing work (they can't verify), replacing office hours (it gives confident wrong answers about course-specific policies). AI literacy is the real skill to teach.