Video of the Day20:35Alex Finn · Demo
ChatGPT Voice Mode Turns Your Voice Into a Chief of Staff for AI Agents
A creator's real-time demo of ChatGPT's new voice mode as an always-on dispatcher for AI agents across every device he owns.
July 27thThe argument in one line.
ChatGPT 6 Astra's real upgrade only pays off if you change how you work with it: delegate multi-step computer tasks to it, save heavy reasoning for planning, and drop the instruction files it no longer needs.
Read if. Skip if.
READ IF YOU ARE…
- Someone actively using ChatGPT's desktop app or Codex for coding, content, or business tasks who wants more out of the newest model.
- A builder juggling multiple devices who wants one dedicated machine to run long AI coding and testing sessions in the background.
- Anyone trying to decide how to split reasoning-effort settings (low vs. high/max/ultra) between planning and execution work to save money.
SKIP IF…
- You don't use ChatGPT's desktop app, computer-use features, or reasoning-effort controls at all.
- You're looking for a neutral, independent review of ChatGPT 6 Astra rather than one creator's personal workflow opinions.
The full version, fast.
ChatGPT 6 Astra's biggest jump is computer-use inside the desktop app: it can operate a browser or app to finish multi-step tasks, like filling an Amazon cart, and now pauses mid-task to ask clarifying questions instead of guessing. The creator's system: run planning and blind-spot analysis on high, max, or ultra reasoning, then hand execution to low reasoning to save tokens; log ideas and tasks to an external Notion or Linear board so a cheaper pass can execute them later; drop custom agent.md instruction files and most skills since the model needs less hand-holding; and dedicate one always-on desktop as a remote testing machine so long automated test runs don't hijack the computer in front of you.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:34
01 · Cold open: "AGI is here"
The creator cites a tweet calling ChatGPT 6 Astra AGI and promises seven tips that will make the reader feel it, including some that save tokens and money.
00:34 – 03:51
02 · Tip 1: Use computer-use in the desktop app
The CLI has no computer-use functionality, so the desktop app is required. A live demo has Astra add chicken-parm ingredients to an Amazon cart unattended, and the model can now pause mid-task to ask clarifying questions instead of guessing everything up front.
03:51 – 05:37
03 · Tip 2: Budget reasoning effort by task
Astra on low reasoning is framed as smarter and cheaper than the prior model on high. The creator's rule: use high/max/ultra for planning and blind-spot analysis, and low/medium for everyday execution.
05:37 – 07:34
04 · Tip 3: Drop agent.md files and skills
The creator argues instruction files and skills were patches for older models' weak spots, and now mostly add token-eating bloat since Astra infers intent well on its own. Exception: highly specialized, repeatable workflows.
07:34 – 09:28
05 · Tip 4: 3D modeling with Blender and Unity
Astra is described as dominant at 3D modeling for the first time. Workflow: generate concept images in Astra, build 3D models in Blender, assemble the game in Unity, shown against the creator's own in-progress shooter game.
09:28 – 11:21
06 · Tip 5: Build a second brain
A high-reasoning planning pass logs every idea to an external Notion or Linear Kanban board, then a low-reasoning pass loops through the board executing tickets, preserving the plan's quality while cutting token spend.
11:21 – 14:00
07 · Tip 6: Use remote mode on a dedicated desktop
Astra tests its own work so exhaustively (up to 45 minutes per feature) that it can monopolize a computer. The fix: designate one always-on desktop as the build/test machine and control it remotely from a laptop, phone, and tablet.
14:00 – 15:47
08 · Tip 7: Ask about blind spots instead of issuing commands
The creator's closing habit: ask the model 'what are my blind spots,' 'what should we do next,' and 'what's the highest-leverage action' rather than dictating next steps, treating the process as exploratory rather than directive.
15:47 – 16:22
09 · Outro: subscribe and newsletter CTA
Closing restatement that this is a bigger step-change than prior model jumps, followed by a like/subscribe/newsletter ask.
Lines worth screenshotting.
- ChatGPT 6 Astra's computer-use only works in the desktop app; the CLI has no computer-use functionality at all.
- Astra on low reasoning is reportedly smarter and cheaper than the previous model on high reasoning, for most everyday tasks.
- Reasoning-heavy modes (high, max, ultra) are best reserved for planning and finding blind spots, not routine execution.
- Running execution on low reasoning after a high-reasoning plan already exists can preserve most of the quality gain for a fraction of the tokens.
- Every agent.md rule and every skill gets reloaded into context on every single prompt, adding token overhead whether or not it's still useful.
- A model good enough to infer intent on its own turns instruction files and skills from a help into unnecessary context bloat.
- Astra reportedly tests what it builds so thoroughly that a single feature test can occupy a computer for 45 minutes straight.
- Routing AI coding and testing to one dedicated always-on desktop, controlled remotely from other devices, keeps long test runs from taking over your main machine.
- Logging every idea from a high-reasoning planning session into an external Kanban board lets a cheaper model execute the backlog later without losing context.
- Asking a model 'what are my blind spots' or 'what should we do next' instead of issuing direct commands can surface ideas the user wouldn't have generated alone.
Reasoning effort is now your biggest lever
AI WORKFLOW HABITS
Getting more from a smarter model comes down to matching reasoning effort to the task, letting go of static instruction files, and treating the model as a planning partner instead of an order-taker.
02Tip 1: Use computer-use in the desktop app
- Multi-step computer tasks, like filling a cart, filling out a form, or doing research, are worth delegating to an AI agent with computer-use rather than doing them by hand once the agent can operate a browser or app directly.
- An agent that can pause mid-task to ask a clarifying question, instead of guessing everything up front, tends to produce more accurate results on ambiguous multi-step work.
03Tip 2: Budget reasoning effort by task
- Matching reasoning effort to the task, low for routine execution, high or above for planning and blind-spot analysis, is a real lever for controlling both cost and output quality.
- Running a heavy-reasoning planning pass once and then executing on a lighter setting can preserve most of the quality gain at a fraction of the token cost.
04Tip 3: Drop agent.md files and skills
- As a model's baseline reasoning improves, static instruction files and skill libraries that once compensated for its weak spots can become unnecessary context bloat rather than a help.
- Every rule file and skill gets reloaded into every single prompt, so trimming them down reduces token spend regardless of whether they were still doing anything useful.
05Tip 4: 3D modeling with Blender and Unity
- Pairing an AI's image-generation output with open tools like Blender for 3D modeling and Unity for game assembly turns a single prompt into a full asset-and-build pipeline.
- A model trained specifically on a skill like 3D modeling can meaningfully outperform general-purpose competitors on that one task, which is worth testing directly rather than assuming parity.
06Tip 5: Build a second brain
- Writing an AI's planning output into an external board like Notion or Linear lets that context survive outside the chat and be executed later by a cheaper, lower-reasoning pass.
- Separating 'think expensively, once' from 'execute cheaply, repeatedly' is a reusable pattern for controlling AI costs on any multi-step project, not just games.
07Tip 6: Use remote mode on a dedicated desktop
- A model thorough enough to spend 45 minutes testing one feature will tie up whatever computer it's running on, so isolating that workload protects your ability to keep working.
- Running long automated AI test or build sessions on an always-on desktop, controlled remotely from other devices, decouples 'the AI is testing' from 'I am blocked.'
08Tip 7: Ask about blind spots instead of issuing commands
- Asking an advanced model open questions, like what are my blind spots or what should we do next, can surface options a user wouldn't have generated on their own.
- Treating a capable model as a collaborator to explore with, rather than only a tool to issue commands to, is a mindset shift as much as a technique.
Terms worth knowing.
- Computer use
- An AI capability that lets a model directly control a computer, clicking, typing, and navigating apps or websites to complete multi-step tasks on the user's behalf.
- Reasoning effort / thinking level
- A setting (e.g. low, medium, high, max, ultra) that controls how much a model reasons before responding, trading speed and cost for depth of thinking.
- agent.md file
- A project-level instructions file that tells an AI coding agent standing rules and context to follow across every prompt in that project.
- Skills
- Packaged, reusable instruction sets an AI agent loads to handle a specific recurring workflow, similar to a saved macro or playbook.
- Second brain
- An external notes or task board, such as Notion or Linear, where an AI logs its ideas and plans so they persist outside the chat and can be acted on later, including by a cheaper model.
- Remote mode
- A ChatGPT desktop feature that lets one computer be controlled and used for tasks from other devices, so long-running work happens on a dedicated machine instead of the one in front of the user.
Things they pointed at.
Lines you could clip.
00:00
“AGI is here. ChatGPT 6 Astra is officially AGI.”
06:36
“When a model is smarter than humans, why would humans need to give it instructions?”
11:42
“Here's a downside of Astra, if you can call this a downside: it is like the most thorough model of all time.”
15:17
“Ask way more questions than prompts you give.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphor
AGI is here. ChatGPT6 Astra is officially AGI. And if you don't believe me, look at this tweet.
The god of AI, Jensen Huang, declared it. AGI has arrived. Congratulations, OpenAI team.
But here's the thing, though. If you don't use it the right way, you're not going to feel the AGI. I got good news.
In this video, I'm going to give you seven hacks that are going to make you feel the AGI. These are the seven things you need to be doing immediately with ChatGPT6 Astra to get way... better results.
If you don't think this is AGI by the end of this video, when you use these tips, you will believe it. Not only that, some of these tips will also save you tokens and save you money. So you get way better usage out of it too.
This is the most important ChadGBT6 Astro video you're ever going to watch in your life. So let's lock in and get into it. Tip number one is you need to be using the ChadGBT desktop app.
More specifically, taking advantage of the computer use. The biggest improvement out of everything that shipped with Chad GBT6 Astra is its computer use technology. It can use your computer better and faster than you.
But how do you take advantage of this? Well, first of all, you use the desktop app, right? The CLI does not have any computer use functionality in it.
So you need to be using the desktop app for that functionality. But then on top of that, you need to have the mindset of anything you would do with your computer. Just ask ChadGBT6Astra to do it.
So to show you how elite and smart this computer uses, check this out. Chicken parm is my favorite food. I'm gonna do this.
Please use computer use to go to Amazon and add all the ingredients for chicken parm to my cart. I wanna make a delicious chicken parm tonight. I need all the ingredients from Whole Foods.
So I'm going to have a go at all the ingredients for me and hit, then all I have to do is start making it. I'm going to hit enter on this. Look how good this computer use is.
I'll use computer use skill to add ingredients for four servings of chicken parm to your Amazon cart. Well, this is going to be great. So it popped open Amazon on my computer.
Look at this. It's going in. It added the chicken.
Now it looks like it's going to add mozzarella. So it's able to think and also use my computer at the same time. Now it's searching for Parmesan cheese.
It's going around. You can see it's controlling the mouse and clicking around. It's adding it to the cart.
So anytime you think of things you need to do on your computer that is like multi -step, you shouldn't be doing it yourself. You should be having ChatGPT Astra do it for you. It is as good, if not better, at using computers than human beings.
But there are way more benefits to using the desktop app over the CLI or anything else as well. I like that search 365 marinara sauce. It saved me a little bit of money by using the generic whole.
foods brand. That's great. You need to be using the desktop app because they added a whole bunch of new features to the desktop app just for Astra too.
Check this out. Here's another reason why you need to be using the desktop app. It now can ask you questions while it's working.
So Astra right now is working on a game I'm building. And as it's doing it, it's asking me questions about what it should be building. What should the coil breaker shotguns charge shot do?
Punch through several enemies, knock enemies backward. Let's make it so it knocks enemies backwards. And boom, as I'm working just asking me these questions to get more familiar with what I want it's almost like a right that ask questions as it works doesn't think all the questions up front it does it as it builds which is amazing that's why you need to be using the desktop app it gives you computer usability and it gives it the ability to ask you questions while it works which just makes the entire model so much smarter and here's a quick trick for using computer use the mindset I try to have is anytime I do something on my computer that I don't really enjoy, I just ask ChadGBT to do it.
Go to Codex, go to work, and anytime you have to do something on your computer, add items to your cart, fill out a form, do some research. Instead of you doing it, ask ChadGBT to do it. I guarantee you'll be shocked at what it's actually capable of.
Here's tip number two for getting the absolute most out of ChadGBT 6 Astra, and this one also helps you save tokens. You need to be using Astralow for almost every single thing you do. on low is actually way smarter than five six soul on high and it's significantly cheaper.
Check this out. This is Astra on low. This is soul on high.
It's both cheaper and significantly smarter. So you want to be using Astra low for almost everything, but there's still some things you want to use max and ultra for. I like to use high and above for any sort of planning I'm doing so when I first go into a new thread or I'm about to do something very intense I go into high and I have it do the planning High and above are like the first AI models I've ever used that truly come up with unique novel concepts really consistently.
They're able to think outside the box better than any other model I've ever used in my life. So you want to use high and above for when you're doing any sort of planning, trying to find your blind spots, things like that. Here's a good tip from my friend Udi I really like, which is he uses it to ask it about his blind spots constantly.
I asked to show me my blind spots and refine the ideas into something I can start testing today. You want to be doing this very often with high, max and ultra. You go in, you say, hey, let's build a plan around this.
What are the blind spots we haven't thought of? And you'll get really good results. For everything else, you use low, maybe medium, and you'll be able to do the execution so well if it's basing it off the plans from high and higher.
Next tip I want to give you, and this is a quick one. It's really, really simple. It comes down to this.
Get rid of all your agent .md files. Stop using skills. This model is so smart that I found basically all the previous commands I was giving in my agent .md file were taken care of, didn't matter anymore.
And skills were only slowing this down and eating tokens. This model is so good at figuring out what tools, what it needs to use that you don't really need skills anymore. That's how smart it is.
I think the age of agent .md files and skills are basically over. I think they were only really patches for like the in models, things the model couldn't figure out or things the model was screwing up. The model doesn't really screw anything up anymore.
So you really don't need to give it specialized instructions. When a model is smarter than humans, why would humans need to give it instructions? Everything you add to your agent .md file and everything you add in skills or anything like that, only adds to bloat in your context.
All your agent .md file rules and all of your skills get loaded with every single prompt. So the less you have, the lighter it becomes, the less tokens it eats. And because the model is just so smart at its core now, you really don't need any of them.
All the other YouTubers and content creators that give you this massive list of rule files and skills you need to use, they're just doing that to get you to click and watch. The model is smart enough now where I think... age of rules and skills are done.
I think there are small exceptions here. If you come up with very specialized workflows that are particular to you, like right now I'm building this game. There's a very specific workflow I use for 3D modeling.
You could build a skill for that. But again, the smarter the models gets, the less you need those because it'll just remember those things in its memory. So get rid of your agent .md file.
Get rid of a vast majority of your skills. The model's smart enough. It'll lighten the load and make it much quicker and smarter.
Tip number four is they clearly, for the first time ever, train this model on 3D modeling. It is by far and away, not even close, dominates all the competition, incredible at 3D modeling. And if you haven't done yet, you need to do it.
Just a quick example before I show you exactly how to do it. This is a game I'm building. There's a ton of 3D elements in here.
It designed all of it themselves. All the buildings, the characters, the graphics. It designed it itself.
I didn't have to use any third -party tools or anything like that. It was just all designed with one prompt, this game. Building 3D models and building games is like the most fun thing ever to do with AI.
If you haven't done it yet, you need to do it. The way I would take advantage of all this is by using Blender, which is an open source 3D modeling library, and Unity, which has an amazing CLI. Unity is a free game engine you can use to build your games.
Now, what you do to get the absolute best results is start a project in Codex and then go, hey, I want you to build a game in Unity. I want you to build assets. with Blender and your process for building these models is this.
Use your image generation to build concepts of what you want the 3D models to look like. You can see that right here. Then use those concepts you generate to actually build the 3D models in Blender and then put them into Unity.
This process is flawless. This process has given me by far the best results ever for 3D modeling and game building with AI.
I mean, the results are stunning of what it's capable of. So if you want to get the most out of this model, you need to be doing 3D modeling, build small games, build cool demos, just use this functionality because I promise you'll be blown away. Use Blender, Unity, and then have Astra generate images and put that into Blender.
I promise it'll be the most fun you have with AI and it'll absolutely blow your mind. Next up is using a second brain and this is going to help you conserve on tokens as well and get better results. So Astra on high to max to ultra is so smart, is able to burn so many tokens, think of so many things at once.
What you want to be doing is connecting it to a second brain so that you can have it write down its thoughts as it goes and go back and revisit them with cheaper models. What I mean by that is... this.
So I have this second brain, which is basically a Kanban board in Notion. Notion's completely free. You can build as many of these as you want.
You can also use linear. Linear is great at this too. I kind of use both.
But this is a Kanban board I have for the video game I'm building, the shooting game. And what I did was I had it plan. So I had to go in high mode and I said, hey, plan this game out for me.
And then I said, every idea you come up with, add it to our second brain in Notion. And as you can see, it came up with all these ideas. There's something like 25 or 30 different ideas in here it came up with.
Did it all on high. It's all saved in here. It all has like really intense detail on each one.
And then what I do is I go back in, I go on low reasoning level, and then I say, hey, loop through our entire notion board and just take care of every ticket and build out every piece of functionality. And it just loops through the entire thing and is able to build it all out really well. And it saves tokens.
because it's doing it on low. So you're doing many things here by using a second brain. You're getting more done because it's able to save every single idea it comes up with to this board, but you're also saving tons of tokens because now you can do execution on low and it still remembers everything the high model came up with.
You get way more done, you get it done faster, and you save tons of tokens. You need to set up this second brain. If you're doing this purely for the second brain, linear works really well.
If you want to take advantage of all the other functionality in Notion, Notion works really well for it too. Notion's completely free. But when you set up the second brain, you get way more done.
You're going to save tokens. You need to be doing this. This will help you get way more out of the model and save a lot of money.
The sixth tip I have for getting the most out of this. Is using the remote mode in chat gbt desktop the right way So here's a downside of astra if you can call this a downside is it is like the most thorough model of all time Everything it builds it tests to an absolutely absurd level I'll have it build like a new weapon in my game and it'll spend 45 minutes just testing that weapon Every single possible way it can inside the game.
Here's the downside because the testing is so good because the testing is so thorough It's just constantly using my computer and doing things on my computer at all times, which makes it hard for me to do other things on my computer, right? If the game's constantly popping up and being tested, I can't do other things.
So now more than ever, it is absolutely critical. You use the remote functionality and you choose one of your computers. If you have multiple.
You choose one of your computers to be like your testing computer where all your code lives and everything gets tested. So this Mac studio we're on right now where I'm filming this video, this is where I remote into from all my other devices, right? So I have my Omachi laptop.
I have my iPhone. I have my iPad. I do vibe coding on all of those, but I have them all remote into this Mac studio computer.
What that allows me to do is I can work and give prompts from all my other devices and all the code gets written and tested on this Mac studio. So now testing doesn't distract me. I can fire off prompts from all my devices and it all gets tested on this desktop.
And the beauty is because I'm using a desktop to do this and never gets turned off and never goes into sleep mode. It can just test 24 seven around the clock. So what you want to do again is think about how you're going to do remote.
Choose a desktop computer if you have one. Could be a Mac mini. Could be anything that's just plugged into the wall and always on.
And you have it as your main computer. Then you remote into that from all your other devices. To set up remote, very easy.
Click on your name. You go settings. You go to connections.
And then from here, you can say control this Mac or control your computer from whatever your desktop is. And then add in all your other devices. you will be able to remote in control and all the testing because it is the best model at testing ever.
It'll test on this computer. Then you'll never get distracted and you can write code on all your other devices. The last tip I'm going to give you here is you need to, because this model is so smart, be asking more questions than prompts you're giving.
Here's what I mean by that. I am constantly asking, we talked about this a little bit earlier, what are my blind spots? What haven't I considered yet?
What can we do next that you recommend? Because this model is so incredible at coming up with novel concepts and novel ideas, I don't like it when I come up with next steps. I want the model to come up with next steps for me because it's going to come up with way better ideas than me.
It's going to be able to come up with things I never thought of. It's going to be able to check all my blind spots. I find that when I say, hey, what should we do next?
rather than me telling it what to do next, we get way better results. We move way faster and the app just progresses way better. So here's my challenge to you when you're using it next.
Don't prompt it. Make everything come from you asking, what should we do next? And then you approve its idea.
Never say, hey, let's do this next. It also makes it more fun where you go on this like exploration of the mile. You're kind of exploring with it.
You're kind, it's almost like a where you're just like, hey, what should we do? What do you think we should do next? And you're exploring and coming up with new ideas and brainstorming and being way more creative, which I think is just way more fun of a way to use AI overall.
So my challenge to you again is this, ask way more questions than prompts you give. Constantly say, hey, what are my blind spots? Constantly say, hey, what do you think we should do next?
And constantly ask, what's the highest leverage action we can take at the moment? If you do that and 95 % of the prompts you send are that, I promise you get way better results and you'll just have way more fun using this model. I truly believe we have entered the AGI era.
I truly believe this is like a step gain. Like we had like the Opus 4 or 5 moment a year. ago.
I believe this is a way bigger moment than even that. This is the Astra 6 AGI moment. Take all these tips I just gave you, put them in action.
I promise your mind's gonna be blown by what this model is capable of. If you learned anything at all, leave a like down below, sign up for my newsletter. It's the number one newsletter in AI.
It's completely free. Sign up down below. I send two very helpful newsletters a week.
That is down below. Subscribe and turn on notifications if you haven't already. All I do is make banger videos about AI.
so grateful you learned from me. I'm so grateful you'd watch my videos. Thank you so, so much.
I'll see you in the next one.
The Hook
The bait, then the rug-pull.
The creator opens with Jensen Huang's tweet calling ChatGPT 6 Astra AGI, then makes the case that the label only holds if you change seven things about how you actually use it.
Named ideas worth stealing.
00:34list
7 tips for getting more out of ChatGPT 6 Astra
- Use the desktop app for computer-use
- Run execution on low reasoning, save high/max/ultra for planning
- Drop agent.md files and most skills
- Use Blender + Unity + image generation for 3D modeling
- Log ideas to a second brain (Notion/Linear) before executing on low
- Dedicate one always-on desktop to remote mode for testing
- Ask the model for blind spots and next steps instead of issuing commands
The creator's full checklist for extracting more value from ChatGPT 6 Astra, covering interface choice, reasoning-effort budgeting, instruction hygiene, and workflow habits.
Steal forany listicle roundup of workflow tips for a new AI model release
How they asked for the click.
VERBAL ASK
15:47newsletter
“leave a like down below, sign up for my newsletter... subscribe and turn on notifications”
Standard end-card stack: like ask, free newsletter signup pitched as the number-one AI newsletter, then subscribe/notifications ask, delivered right after restating the AGI thesis.
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
