OpenAI has released GPT-Live, a new generation of voice models that now powers ChatGPT Voice. It started rolling out globally yesterday across iOS, Android and the web. The headline change: it can listen and speak at the same time. Every previous version of Voice worked in turns. You talked, it waited for silence, then it answered. That is why it kept butting in when you paused to think, and why the whole thing felt a bit like a walkie talkie. GPT-Live processes what you say continuously, so you can interrupt it mid-sentence, pause to gather your thoughts without it jumping in, or tell it to stay quiet and just listen. It even gives the small acknowledgements a real person does, the odd "mhmm" so you know it is following. The second change is the one I think matters most for business use. When you ask for something that needs real work, a web search, proper reasoning, digging through a file, it hands that job to GPT-5.5 in the background and keeps the conversation going while it runs. You can also pick a reasoning level: Instant for quick answers, Medium or High when you want it to think harder before it speaks. A few practical details worth knowing. GPT-Live-1 becomes the default for paid plans, with a mini version as the default for free users. It can now show visual cards while you talk, things like weather, stocks and sports. It is better at ignoring background noise. And at launch it does not support video or screen sharing, so if you use those, the older voice modes are still available. Here is my take on why this deserves your attention. Voice has been the feature most business owners try once and quietly abandon, because talking to a turn-based bot feels like effort. The interesting shift is that voice can now hold a natural conversation while real work happens underneath it. That starts to look less like a gimmick and more like thinking out loud with an assistant. The obvious first test: next time you are driving or walking, talk through a business problem with it for ten minutes and see what you come back with.

Most people accept ChatGPT's first answer, tweak it a bit, and move on. The single biggest upgrade you can make is to stop treating the first draft as the answer and start treating it as something to be marked. Here's the move. You give ChatGPT the task, then you hand it a rubric, the same criteria you'd use to judge the work yourself, and you make it score its own draft against that rubric before you ever see it. Then it rewrites to fix its lowest scores. Say you're writing a cold email. Most people prompt: "Write a cold email to a marketing director offering our service." You get something generic. Instead, try this: "Write a cold outreach email to a marketing director. Then score your own draft from 1 to 10 on each of these: 1) does the first line earn the second, 2) is it about them not us, 3) is there one clear ask, 4) would a busy person read it in under 15 seconds. Show the scores, then rewrite to fix anything under 8." Now you're not hoping for a good email. You've told it what good looks like and made it run the editing pass you'd normally do yourself. Two things make this work. First, the rubric is where your expertise goes. You know what a good email, landing page, or proposal needs, so you encode it once. The model is far better at applying a clear standard than inventing one. Second, asking for scores forces it to actually evaluate instead of just rephrasing, and it will usually catch its own weakest spot before you have to. Save your favourite rubrics and reuse them. A good-email rubric, a good blog-intro rubric, a sales-call-summary rubric. Over time that's a quiet quality system running on every task. What's a task you'd want a rubric for? Tell me the task in the comments and I'll help you build the criteria.

I just fell in love with AI, I created my very first Oracle Shuffler Artifact. All images created on CHATGPT and all coding with Claude. Tada. I hope I am allowed to share this here: https://sacredselflove.netlify.app/

Most prompting advice tells you to write the perfect instruction up front. For everyday business tasks, I think that's backwards. You're guessing at what matters before you've really thought it through, and you end up with a vague answer because you handed it a vague brief. Here's a simpler move: tell ChatGPT to interview you first. Instead of asking for the finished thing, add one line to your prompt: "Before you answer, ask me up to five questions that would help you give a better response." Now ChatGPT does the hard part. It surfaces the gaps: who the audience is, what the goal is, what tone you want, what you've already tried. Then you just answer in plain language. The final output gets built on real context instead of assumptions. This works best on the messy, higher-stakes stuff: a sales email, an awkward client reply, pricing for a new offer, a job description. Anything where the quality depends on details only you know. Two tips. Cap the questions. Five is plenty, or it spirals. And if a question doesn't matter, just say "skip that one." You're steering, not filling in a form. It feels slower for about ten seconds. Then the answer comes back sharper than anything a one-shot prompt would have produced. What's the last thing you asked ChatGPT for that came back generic? That's exactly the kind of task to try this on. Drop it below. Jason 🙌