If you’re not working on AIO, then you will increasingly be lost to all but those who search for terms like “near me”. SEO is still useful, but more so from a local context. Where one used to celebrate being ranked #1 in search results for “coffee van Hobart” (yes, Bean & Leaf is one of my ventures), it’s only useful for those who are local to Tas. Forget the rest of the world, or general searches.

With the lag/TTL issues that plague real time voice exchanges with AI models, bots are still king. The generation that prefers to test over talking, prefers using bots over voice AI, as the latter is still too “confronting” for them. Kind like how the “👍” emoji upsets them. 🤷🏻‍♂️ In all seriousness though, showing application of AI in a much more simple way (bots) works much better for those who are not sure what AI can offer them. So, bots could be considered the lead in drug for AI uptake. 😂😊

I am using it now. It is faster and more accurate in coding responses than Deepseek & Claude. Am using it to code a AI trading bot and, according to my full-stack dev guy, it is smashing it.

@Zahida A Khan haven’t played with/used it. Heard that there is a lot of self-reporting of its capabilities by Alibaba. No one is sure if it’s training data, but, then again, does anyone really care? 🤔

Hey @Toby Barnes . This is a long post, but I think it is important for everyone as well as you 😊 TL;DR: There's no prompt that resets ChatGPT's "brain". When a thread becomes overloaded with contradictory instructions (even though you think that they are clear), the model becomes stuck and starts giving delays or non-answers. The only real fix is to start a new chat. Put your instructions at the very top of the new thread, paste the chapter (or attach it as a txt file), and then let it run. Don't carry anything over from the old conversations. For each chapter, open a new chat. That fully resets the system and prevents contamination,. ——- Detail/what's happening: You ran into run into thread contamination, or what is known as a corrupted instruction-state in the LLM world. In long, multi-step editorial tasks, the model can get trapped trying to reconcile contradictory earlier instructions, especially when they’ve spanned dozens of message turns. Even after you tell it to “ignore previous instructions”, the underlying thread still carries the weight of earlier constraints, earlier examples, earlier rejected attempts, and the implicit optimisation pattern it learned throughout the task. This creates a feedback loop where the model cannot resolve the conflicting objectives, so you get stalled outputs, time-estimates instead of real work, hallucinated reasoning (“I need 20 minutes… 30 minutes…”), refusal to act, excessive caution, or total non-production. Starting a new chat is the entire fix. A fresh thread wipes all prior turn-history, all prior reasoning scaffolds, and all conflicting instruction chains. Remember the old adage in the business world where you have one subject per email? 🤔 Same for ChatGPT. Instructions: 1. One chapter per new chat. Do not process 20 chapters in one enormous thread. They do not contaminate each other. 2. Give the instructions once, at the top, not multiple times. Repeated guidance mid-flow often forces the model into self-correction loops.

@Julie Helmer Indeed you did. I just added the “why?” 😊