Power users have flipped to voice prompting. Simon Young, who runs Stitch in the UK and partners with GEOforge, has moved to nearly all voice over the last three months, and Paris Childress is majority voice, speaking the first prompt as stream of consciousness into SuperWhisper rather than typing a polished brief. On episode 15 of The GEO Show, recorded Tuesday, 1 September 2026, they argue that messy spoken context beats a carefully structured typed prompt, that token usage anxiety (TUA) is the bottleneck, that SiteForge has the targets and ContentForge has the payload for earned-path publishing, and that always-on ambient voice will explode token demand.
Power users have flipped to voice prompting. Simon Young, who runs Stitch in the UK and partners with GEOforge, told Paris Childress that over the last three months he has moved to nearly all voice. Paris is majority voice. The first prompt is the long one, spoken. Short follow-ups still get typed.
Paris and Simon recorded episode 15 of The GEO Show on Tuesday, 1 September 2026, the day after the UK bank holiday. The throughline is the flip itself, why stream-of-consciousness voice beats a typed prompt, why token usage anxiety (TUA) is the bottleneck, and why always-on ambient voice will explode token demand. They also land the product line from last Friday: SiteForge has the targets, ContentForge has the payload, and joining them creates an earned path for high information gain publishing.
Paris asked Simon what share of his prompting volume is voice versus typing, and where that is going for most marketers. Simon's answer: nearly all voice, because he cannot type as fast as he can speak, and because talking puts more into the prompt. Claude's voice input still cuts off after about a minute or two, which is the friction on that stack.
Paris put himself in the same cohort. Majority voice. If the follow-up is a short "proceed," he types it. The first prompt is War and Peace. That one is almost always spoken into SuperWhisper. He starts recording before the thought is finished. He talks, takes it back, and lets the model hear the process. A human would not tolerate that conversation. The model does.
The results that I see coming out of a long stream of consciousness voice prompt like that, as opposed to a well-thought-through structured typed-in prompt, it's night and day. I get much better output from my voice prompts.
Simon had just done it. One spoken prompt sent the Stitch studio, GoHighLevel, a live blog check, and three scheduled social posts across three weeks. On screen the transcript looked messy. The chain ran.
Paris burned his Claude weekly tokens on Sunday afternoon. Monday, a UK bank holiday, Claude was unavailable. He would not buy extra credits. He forced the work onto GrokBot, which generated and published episode 14 of The GEO Show while he was at lunch. YouTube, Transistor, and the episode page shipped. The LinkedIn draft was the missing 5%.
Simon named the feeling token usage anxiety, TUA, and called usage limits one of the biggest talking points in marketing in 2027. The products do not warn that a workflow will not finish. They want the wall, then the top-up. Paris was at 21% of a GrokBot week with five days left, enough for this show workflow and not much more. Replit made the same anxiety worse by charging per token and showing the dollar amount while a failed 20-minute job burned $8.
An owned path is publishing on a URL the brand controls. An earned path is a publisher or influencer with more authority carrying the same payload. Paris credited a model-written line from last Friday: SiteForge has the targets but not the payload. ContentForge has the payload but no earned path. Join them and high information gain content goes to media publishers, including recurring exclusives, not only to the brand blog.
Simon is already taking GEOforge blogs and spinning different Facebook, Instagram, Google Business, and LinkedIn posts across three weeks, still pointing at the original URL. Most marketers make an asset once. Automation is what lets that run every day. In a room of more than 100 recruitment-business owners, everyone used AI. About 10% had automated processes. The rest cited GDPR, safety, and data fear. The people who move anyway get the advantage.
The other prompting move on this tape is flipping the script. Simon fed Paris's tone-of-voice skill into Claude, asked it to interview him to 95% confidence, and uploaded a 90-minute GEO talk. The skill came back sounding like him. GEOforge's interview feature is the client version of that: pull knowledge that would never otherwise hit paper or the web, then put a unique point of view where third-party citation can find it.
Simon is waiting on a pocket device so he can talk to Claude and book meetings while walking. Paris pointed at a rumored screenless OpenAI product in the same direction. If everybody has something around a neck, on a watch, or in a bag that is always listening, AI stops being reactive and starts being proactive. Every word around you becomes context. That multiplies tokens. Paris does not know how the market pays for that if price per token stays here, even with open source pushing cost down.
The last third of the tape leaves the prompt and goes to what platforms hold. Simon walked through a UK subject access request, a SAR, including a bank that physically mailed 30 years of statements, and asked what Meta or WhatsApp would actually return. They teed a Google redirect story for Friday. The throughline stays the voice flip. Power users already speak the first prompt. TUA is the meter that will decide how far that spreads.
00:00. UK bank holiday context. Paris was working through the weekend.
00:28. Paris burns his Claude weekly tokens by Sunday afternoon. Monday he is a lost puppy and will not buy extra credits.
01:58. GrokBot generates and publishes episode 14 while Paris is at lunch. YouTube, Transistor, and the episode page ship. LinkedIn draft still missing.
04:43. Simon's Claude skill takes a GEOforge blog and schedules different social posts across three weeks, still pointing at the original URL.
07:11. Room of 100-plus recruitment-business owners: everyone uses AI, about 10% have automated processes. GDPR and data fear block the rest.
09:09. Simon: usage limits will be one of the biggest marketing talking points in 2027. No warning the workflow will not finish.
10:37. Paris describes token usage anxiety. Simon names it TUA.
13:18. Pivot to voice. How much of the prompting load is already spoken?
14:38. Simon: nearly all voice over the last three months. Typing is slower. Talking puts more in the prompt.
16:22. Paris: SuperWhisper stream of consciousness versus a structured typed prompt is night and day.
18:11. Simon's one-shot spoken prompt runs Stitch studio images, GoHighLevel, a live blog check, and three scheduled posts.
19:31. Claude interviews Simon for a brand-voice skill after a 90-minute GEO talk upload.
23:18. SiteForge has the targets. ContentForge has the payload. Join them for earned-path publishing.
28:31. Pocket devices and always-on ambient voice. Reactive AI becomes proactive. Token demand multiplies.
33:22. UK subject access requests, SARs, and what Meta or WhatsApp would actually return.
39:33. Simon tees a Google redirect story for Friday.
39:54. Outro. Next episode Friday, same time.
The throughline stays the voice flip. Power users already speak the first prompt. Stream of consciousness beats typed structure. TUA is the meter. Always-on voice is how that meter explodes.
[00:00:00] Paris: All right, Simon, we're back. How you doing today?
[00:00:04] Simon: Good. Bank holiday weekend here in the UK, so pretty lazy one over the weekend.
[00:00:10] Paris: Oh, nice. I suppose you were working yesterday, were you?
[00:00:13] Simon: Yeah.
[00:00:14] Paris: Did you all have off on Monday? Was that the bank holiday?
[00:00:17] Simon: Yeah.
[00:00:18] Paris: Yeah, I was working yesterday, working through the weekend. I actually burned out all of my Claude tokens around Sunday afternoon. And Monday felt like I felt like a lost child, really. I didn't know what to do with myself because I couldn't use Claude. It was the weirdest thing.
[00:00:41] Simon: You used up your whole month's allocation?
[00:00:44] Paris: No. Weekly. My weekly budget. It resets every Tuesday morning at 9:00 AM. I try to optimize that every week, and I try to finish with a big session on Monday, and that's usually when I'm not working on the weekend. This past weekend I worked a lot, and I didn't manage it correctly, and I burned out all my credits by Sunday afternoon. Then I get in on Monday and Claude was unavailable the whole day. I was like a lost puppy dog. I had to figure out, what am I going to do with this day now that I can't do anything in Claude?
[00:01:22] Simon: No doubt it said, would you like to buy some extra credits?
[00:01:26] Paris: I won't buy the extra credits. I've already done that twice, and I'm not comfortable. My extra credits get maxed out pretty early by the other colleagues, and I'm okay with that. I thought about pulling the trigger and putting in another $50 or $100. I didn't, and one of the reasons I didn't was because I wanted to force myself to get back to GrokBot on Monday. Let me tell you, Simon, GrokBot is the real deal. I told you last time, I have this workflow where I record a podcast episode into Descript, and then I send it to YouTube with all the assets, I send it to Transistor with all the assets, a LinkedIn article for the newsletter that needs a specific type of formatting, and it goes on the company blog as an episode page. All that distribution I had been doing in Claude, and I had been doing it manually with a skill. I have a skill that creates all the assets, but then I have to manually take every asset and actually go into YouTube, LinkedIn, Transistor, et cetera. On Monday I had a breakthrough. I said, let me see if GrokBot can not only derive and generate all these assets, but if it can also publish these assets. And the good news is, it did it. Today, for the first time, I recorded the podcast, about 15 minutes. I have my intelligence brief that comes to me every morning with 10 stories, and I read through those stories. I add my own context, my own opinion to them, and then I stopped recording and I pinged GrokBot: hey, episode 14 is ready. It's in Descript. Go and do it. Execute the skill. I copied the skill over from Claude into GrokBot. We had chatted yesterday with GrokBot saying I want you to not only do what Claude does and deliver the assets, but I want you to publish them. I went to lunch, and I was watching it, and by the time I got back from lunch it had created the assets and it had published the assets. So my dream workflow is now very close. It's about 95% of the way there. I can't find the LinkedIn draft. I don't know where to look for it. But YouTube is done and published. Transistor is done and shipped. That means it's now getting sent through RSS to Apple and Spotify. The blog post is published, and it looks good. It's got the show notes, it has the transcript, it has the image. So it's happening. I am going to be moving my workloads over to GrokBot.
[00:04:20] Simon: Is it expensive though, compared to Claude?
[00:04:24] Paris: It started out at $200 a month, so yes. But then a few days ago they introduced new pricing.
[00:04:32] Simon: I'm in the process of... while we're sat here, Claude is working one of my workflows where I'm testing some GEO stuff in the background to see whether I could get Stitch to another level. I've built a skill in Claude where it's taking the content from GEOforge. It's then rewriting that as different social assets across Facebook, Instagram, Google Business pages, LinkedIn obviously as well. And then as soon as the blog goes up, it's then sharing that, but it's also scheduling future content as well. I've preached to people about, the worst thing you can do in marketing is make an asset and only use it once. When we've got a piece of content that we can talk about more than once, a really good blog that we've produced using GEOforge, we can then create three or four social posts all different. So not churning out and copying, but make different posts and then share them, one this week, one next week, one the following week, but still pointing to that original blog, which I think might be interesting to test, because I think that's going further than what most people would do.
[00:06:25] Paris: Much further. And the fact that you can do that automated means that you can do that not once a week, but you can do it every day. You can move five times faster, and I think very few people are doing that now. Very few. They know it's possible, but I think it still intimidates most marketers even to really roll up your sleeves and solve the automation thing and really max it out. It was more intimidating a few months ago when n8n was the craze and Make.com, when you actually had to build out all the infrastructure. But even now, I think it's still just difficult for people to make it happen. But it is getting easier.
[00:07:11] Simon: To talk to that point, a friend of mine had a networking meeting with a lot of recruitment businesses, and it was their owners that were in the room. He then said, how many people in the room, I think there was over 100 people, how many people in the room use AI? And obviously everybody put their hand up. And then, how many people in the room have automated processes within their business using AI? And about 10% of them put their hand up. And when he then challenged them to say, well, why is that the case? Everybody said, realistically, we're pretty scared about getting it right. And is it safe, and is it compliant, and do GDPR concerns come into it, and what about our data? So I think everyone knows you can do it, but moving quickly, those that do are going to get a huge advantage.
[00:08:08] Paris: Yeah, absolutely. By the way, what I have on the screen here is the updated plans for GrokBot. Two weeks ago you had to pay $200 to get in. I'm on the Ultra plan, but I'm thinking about downgrading. Now you can get in for $20, and it says GrokBot access. Of course that's going to be more limited. I don't know by how much. Then if you go up to $60 a month, which is still not bad, you get higher GrokBot usage limits, and then $200 you get the highest. Let me just tell you where I am, and I've been using it now very heavily in the last couple of days. I'm at 21%, and I reset in five days. With this new workflow I'll be able to absorb this workflow within my token budget, but I probably can't do a whole lot more.
[00:09:09] Simon: That's going to be one of the biggest talking points in marketing in 2027: usage limits across whichever tools you're using. Using Claude, it does feel like my limits come down since I started using it. I can hit a limit now where I probably wouldn't have done before. When I get to the point, if I'm like 11 o'clock in the morning and I've used my credits, it says this resets at 2:00 PM, what am I supposed to do? Start banging my fists on the desk? Because it doesn't say, be careful, Simon, calm down because you might use up your credits, or this workflow you're about to run appears like you're not going to get it finished. They'll never give you those warnings. They want you to hit the wall and then top it up.
[00:10:13] Paris: Yeah. Anthropic's run rate is now at about $70 billion, and they're projected to end this year at an annual run rate of over $100 billion annual recurring revenue. I think that's because millions of businesses like us are experiencing this exact phenomenon. I've gotten hooked on it. I use it every day, and I have all this anxiety now about token usage anxiety. I'm trying to manage it properly. I want to use it all because if I don't use my tokens, I'm leaving tokens on the table. That means I'm leaving intelligence on the table. It means I could have done more this week and I didn't. But then if I hit the wall, then I have to stop immediately, whatever I'm doing.
[00:11:09] Simon: You've just coined a new tradition. You've just coined token usage anxiety. Maybe that's a thing already, but TUA.
[00:11:16] Paris: Yeah. It's real. I'm now checking how my usage meter changes now with almost every task throughout the day, and I see, oh, it ticked up by 2%. That was a big one. I gotta watch it. And if I see it go up by a lot, I really get anxious. I go, oh, I'm now outpacing my token budget. I'm going to hit a wall. I better slow down.
[00:11:39] Simon: I used to use Replit a lot, and Replit was one of the early tools for building code that was really... it seemed quite cool, but the problem was it was charging you per token. It wasn't charging you a monthly fee. It was charging you as you go. And the problem is it then says, oh, I didn't quite manage to get that right. And you're like, well, I asked you the right way. I prompted you to do it how I wanted it done, and then it gets it wrong. And you're like, well, can I have those credits back, please?
[00:12:19] Paris: Yeah. No. I was a heavy Replit user, and I know exactly what you're thinking. I started to hate the tool because I thought that it was conspiring against me by taking longer and doing a more complex job than it needed to so that it could take more of my tokens. And suddenly I see, wow, that job that Replit just took 20 minutes to do and failed, it just cost me $8, and I'm pissed. It's like, yo, I want that back. Where's my refund?
[00:12:46] Simon: And the worst thing about their tool is their tool actually shows you the monetary amount as it's going along, doesn't it? You're seeing it ticking up, going, hey, that was an extra $1.70. Well, no, you got it wrong. Leave me alone.
[00:13:00] Paris: Yeah. Give me my tokens back. I should only pay when you're able to complete the task successfully. If they carry on like that, they're out of business compared to GrokBot and Claude and everything else. I don't think they're going to be able to continue that for too much longer. So Simon, what were we going to talk about? We were going to talk about voice today, and I think it's fitting because what we're doing right now is we're trying to use our voice to create original content, because the insights that are going to come out of this conversation would take us forever to think about and type up. Let me just punt this over to you. How much of your current prompting load, prompting volume, is voice versus typing? What is the trend? Where is that trending? And then where do you see that going for the vast majority of marketing users like us?
[00:14:09] Simon: There was a stat that I read that the first time somebody uses ChatGPT, having never used AI before, I think six months later or thereabouts, their searches on the web have gone down by 50%, as in they are now using an AI to answer half their questions. So that's great to know. Now, if I go back a year or more, I don't see myself doing anything pretty much without Claude or ChatGPT or Gemini now. I use all of them. To answer your question in terms of voice, probably over the last three months I've moved to nearly all voice. So where I can record it and instruct with my voice, it's nearly all that. Because one, I can't type as quickly as I can speak. But also I think you manage to get more into your prompt when you talk to it. The one thing that's a bit annoying with Claude is the limit on the amount it can take through a voice instruction, because you can probably get maybe a minute or two and then it'll cut off, and then you have to give it another prompt. But trending-wise in terms of the general public, that's a whole other question.
[00:15:31] Paris: I think you and I are a bit out in front. I think the majority of people are still typing and still really thinking long and hard about how to phrase it the right way with the type. But I agree with you that speaking, I think it's moving towards voice. Right now I am majority voice, but I still type sometimes. If it's a short one, especially once I'm into the conversation and maybe I get a few shorter answers where I might just have to type in, yeah, proceed, let's do it, or a short phrase, I'll type it because it's about the same as speaking it. But my first prompt, which is usually when I just have to give it War and Peace, I'm trying to give it as much context as possible, I'm almost always using voice. What I do is I just hit the record button. I have a shortcut. I use SuperWhisper. I have my hotkey. And even before I've fully formulated my thought I just start recording, and I actually just start thinking out loud, and then think that it's a stream of consciousness. And I even may say something and then decide, actually, what I just said, no, forget that. That's a little wrong. What I meant to say was this, or let me phrase it a different way. And unlike a friend or another human that would never tolerate such a conversation, because it's like, wow, this guy doesn't shut up, and it's just diarrhea of the mouth, he's just saying whatever comes to his mind, and he's not letting me even get a word in, AI is very patient and sits and listens. And then once you have it, you've got this big block of very unstructured thoughts, but it has your whole thought process. It has the context. And I think that the results that I see coming out of a long stream of consciousness voice prompt like that, as opposed to a well-thought-through structured typed-in prompt, it's night and day. I get much better output from my voice prompts.
[00:18:00] Simon: I've just done one now 10 or 15 minutes ago where I said, right, I've got my GoHighLevel open, I've got Claude, I've got the customer's website, I've got GEOforge, and I want you to go and create... I talked to you about the studio I made for Stitch. So it's going into the Stitch studio, making me 12 images, downloading them all, taking them across to GoHighLevel, dumping them into the media library. Then it's got to create, go to the blog, see if we've launched it, check that it's up on the site. Then it's going to write me three versions of that, schedule three posts, pull in three images, share across three weeks. And like you say, I spoke that prompt in one go, and when it goes up on the screen, it looks, wow, how can you understand what's going on here? But the level of understanding is amazing. It's done it all from one prompt.
[00:18:50] Paris: There was a time where I actually took a lot of care in writing with proper capitalization, with proper punctuation marks. I would start new paragraphs when I thought I needed to pivot. The same way, all the best writing techniques. I used to take the time and the effort, and I'd put those best writing practices into my prompts, and now I just think, well, I'm a little bit sad because am I going to lose those skills? Because those are still, I think, important skills. I don't want to turn into a really sloppy writer, but I'm just wasting time actually. It doesn't make sense.
[00:19:31] Simon: And talking to writing properly, thank you for sending me your tone of voice skill that would then set up the ability for me, in my Claude, to write properly. I took yours, said to Claude, here's Paris's. He's my friend. Let me create me from this. I want something like this. And it then interviewed me for about half an hour. I then uploaded, so I made an hour and a half video all about GEO last week. I uploaded the transcript of that. It went through that, and it came back and said, oh, Simon, you say this quite a lot, and this is the way you talk. And is this right? Is this right? And talking backwards and forwards to it there has then created something that is, well, it is pretty much me.
[00:20:27] Paris: So that was for your Simon brand voice skill creation, right? Yeah, we were chatting about that. That's great. That's another interesting use where if you're not sure what to ask AI or how to phrase your question to AI, you just flip the script and you say, okay, this is my goal. I want to get a skill that captures my brand voice, and I can call on this skill anytime I want to write something that really needs to sound like me. I don't even know how to tell you what to do here, so you ask me questions, you interview me until you're 95% confident that you've got my brand voice nailed, and that you can produce a full set of brand voice guidelines from my voice. And keep asking until you get to 95% confidence. And once you do, then show me what you've got and let's do a few examples together. And it sounds like that's what you did, and you had a back and forth where instead of you asking ChatGPT and prompting it, it was prompting you, basically asking you, are you more this or this?
[00:21:32] Simon: Yeah, and this process is something obviously we're going to be going through with clients. The interview feature in GEOforge is incredible. You can just get it to talk to somebody about their business and pull out all that knowledge that wouldn't necessarily ever go down on paper or on the web. And as we know, as long as you've got that unique perspective on a point of view, we want it to be shared. Obviously we need the value to be there to be shared. Third party citation is everything. I was talking to some people about link building based on that, and the conversation we had last week around using GEOforge to push out content to relevant publishers. I think if we get to the point where we've got influencers that are prepared as well, that could be really cool. I get people reaching out to me to talk about GEO, the way AI search is going. I'm already getting a lot of software products coming to me saying, we want to show you what we can do, what we've got, and I'll say to them, okay, well, I need inside so you can show me how it works and I can test it. And then I'll give you an honest review on what I've found and how it can be improved. And that's one of the fascinating things working with GEOforge and yourself is to say, right, we're already discovering things that we want to see in the tool.
[00:23:18] Paris: This was a big idea that we stumbled into last Friday, and I've already run it up the flagpole internally. Of course, AI gave me this great phrase that I'm going to have to give it credit for, but I'm going to cite it, and that is that SiteForge has the targets, but it doesn't have the payload. ContentForge has the payload, but it has no earned path. And so by joining the two together, you take the payload, the great high information gain content generated from ContentForge, and you deliver it not through an owned path, but through an earned path to media publishers who have much higher brand recognition, higher authority as publishers than you do. So it's going to get that to a bigger audience, and it's going to elevate your brand. You're basically going to ride on the coattails of that brand because you've got the payload. I mean, you've got some proprietary research that you're going to offer exclusivity for that publisher. I actually just got off with Luke right before this talking about this feature, and this could open up doors of recurring publishing. You could develop, whether they be influencers or you could have a group of people that you'd say, all right, I can probably give you one or two or three really high quality stories per month. So once you have those relationships, you can keep going back to that well. And then influencers as well. I've been looking at influencer marketing for GEOforge, and I actually had Claude go out and build me a prospecting list for both YouTube and LinkedIn, 20 people on each, and it did a very good job. The way this game works is usually they charge... influencers will charge you a flat price to do a video. And then you can have some performance kicker, like maybe for every trial sign-up, they will pay you an extra $20 or something.
[00:25:16] Simon: I did a huge piece of work. I was working as the marketing director of a business called Blockovate in crypto, and that was all about collaboration. We were working with hundreds of influencers. And like you say, most of them would be YouTube, Twitter, wherever. And they'll have a flat fee to publish or to review a product, et cetera. But the budgets were ridiculous. You're talking $200,000, $300,000 a week to promote. But you're talking thousands of different publications publishing about a particular crypto as it was at the time. More interestingly, I don't know whether you saw that I put a post on LinkedIn today and I tagged you in as a collaborator. I don't know whether it notified you or didn't, but it allows you now on LinkedIn. It's not something I've done very much of in the past as in tag people in as a collaborator on a post, and I don't know what that actually necessarily gives you. But it could well be interesting to say, okay, GEOforge, I would like to reach out to 5, 10, 20 people who would collaborate on this post because it is of interest for them to do so. Does that then therefore give us more weight on content that we're publishing that way? Be that an article on LinkedIn or a post, maybe that gets cited as well.
[00:27:00] Paris: Interesting. Yeah. If you have multiple collaborators on a LinkedIn article or a LinkedIn post?
[00:27:06] Simon: On a post, yeah. But if the post is referring to an article, so that's where, in theory, you would get what you might call the link juice.
[00:27:13] Paris: Let me pull it up. I want to see if this is what you're referring to.
[00:27:18] Simon: That was about three hours ago, though. No. So that's me just reposting you and I coming on now. Did it not... I don't know whether it notified you, because it was a question marketing post, so it doesn't look like it has.
[00:27:43] Paris: It's not in my recent notifications, no. If it doesn't make it into the notifications, then probably I've missed it somehow.
[00:27:52] Simon: See, I also tagged in, for example, Neil Patel randomly. Let's see if Neil will collaborate on my post. I just wanted to see whether it did anything.
[00:28:06] Paris: It didn't really come up on my radar screen. It didn't get into my notifications. It really should have though, because everything else that you've been doing, resharing stuff, usually that's all coming up in my notifications. It should be a priority. I don't know what's happening there on LinkedIn. All right. Is there anything else to say on the topic of voice AI? Or are you waiting for OpenAI's mysterious screenless AI product, which is going to be voice, most likely all voice operated?
[00:28:57] Simon: It fascinated me, the thing you talked about to have a device that was just going to get thrown on the table when everybody is chatting around, and it does pretty much everything. I will get my pocket device delivered hopefully tomorrow so I can start working in terms of talking to Claude, booking meetings while I'm just chatting away, instructing it to do what I want to do. I just think I'll be able to work so much more quickly, and not forget anything, and get it to do tons and tons... well, burn more credits. Maybe that's where we're going.
[00:29:36] Paris: I mean, the demand for tokens is only going to keep rising exponentially as these types of devices just become ubiquitous. Imagine if society eventually just accepts this and everybody has something that's around their neck, on their watch, or in their bag that's always listening, always watching, always listening, and you just basically have the context of every word spoken by you and around you all day, and maybe along with visual inputs as well, and you have all that context feeding into AI and then giving you certain outputs about, hey, so and so said this at the bar last night, and just a reminder that you said that you were going to call him back at this time. It's all using tokens. And then that's a proactive AI, not a reactive AI. So when AI is slurping up all this context from voice, ambient always-on voice, and then being able to shift from reactive mode to proactive mode, that just is going to multiply the amount of tokens. So I don't know how we're going to pay for this if the price per token remains where it is now. I think open source is definitely driving the price down. But the supply constraints... most of the US is protesting data centers, which I think is very stupid.
[00:31:19] Simon: What, just because of the environmental impact?
[00:31:22] Paris: Yeah. And with, I think, incorrect information about water usage and the impact to electricity, to the electricity grids, local electricity grids.
[00:31:32] Simon: Without being funny, I didn't realize that data centers drank water.
[00:31:38] Paris: They use water for cooling. They have to keep a very cool environment. But they recycle the water. And there was a book that was published six or eight months ago that made a wrong assertion about the amount of water that a data center uses. And it was wrong by an order of magnitude, by a huge amount. I don't know the exact details, but that actually convinced a lot of people and generally convinced the general public that data centers are massive water hogs, when in fact they're not. So that's myth number one is about water usage. And then the second big argument against the data centers is the electricity strain that it puts on a local grid. And at least in the US, the mandate now is that all of these data centers have to bring their own power, and they have to be able to even send any excess electricity that they have generated and not used back into the local grid. So they can actually be net contributors instead of net drainers. And so that's kind of stalled. But if politically now the way to get elected is to be anti-data center, and if they don't allow anybody to build data centers, then all the data centers move to China, and then we're all on China's AI systems, and I don't think that's a good thing.
[00:33:22] Simon: To get back to the voice piece. I was minded to send a subject access data request. I don't know whether you... do you know what that is?
[00:33:32] Paris: No, I don't think so.
[00:33:34] Simon: So in the UK there's the ICO, the Information Commissioner's Office, which is the police of data, basically. Any company that steps outside the regulations that exist would get reported to the ICO, and you have the ability to then be fined huge amounts of money. My wonder was, now if you ask any company in the UK for a subject access data request, known as a SAR, by law they have to give you every single piece of information they hold on you, be that written, electronic, voice, everything. An example of that would be I got into an argument with my bank, and I said to them, subject access data request, please, and they have to give it for free. And I've been with that bank for 30 years, so they had to send me a copy of every single bank statement that I've ever had. Everything. It came in, it was like one of these legal cases where it comes in multiple boxes.
[00:34:44] Paris: It came mailed, physically mailed?
[00:34:47] Simon: Yes. And you have the right to ask for it to be mailed to you physically. They can say, we want to provide it digitally. Or you can say, no, I want a physical copy of it. Anyway, long story short, my wonder was were I to ask Meta or WhatsApp, obviously not headquartered in the UK but have offices here, if I could have a copy of every single piece of data that they hold on me, would I get a copy of all my WhatsApp messages and, more interestingly, the information they hold behind the scenes which says, here is Simon's knowledge base and profile of him and everything he's ever done. Because they've been recording for however long. I bet I'd probably just get a very impolite letter come back saying, get lost, in not so many words.
[00:36:06] Paris: That would include, if they were listening to everything you've been saying, they would have to show that too.
[00:36:12] Simon: And internal emails that mention you. So if you're employed by a business, and let's say you're terminated or even not terminated, you have the right to ask for a subject access data request. And say two of the directors of the business have been emailing backwards and forwards about you, you have rights to have copies of that as well.
[00:36:31] Paris: I didn't know that. Wow. And what a massive headache for the company to have to, in this case, the bank. Imagine how much work was involved in pulling your data and printing it and packaging it, sending.
[00:36:49] Simon: And they all try and squirrel out of it. Every single business will have somebody who is listed as the data controller. And that person legally has to then provide the data to you. I had another case where I asked for this, and they do then have the right to say to you, this is a complex case. We require an extra three months to provide the information. I imagined that if we did it to Meta or WhatsApp and said, right, I want that information, it exists. It has to exist somewhere.
[00:37:30] Paris: Yeah. How much of it they keep is what I would love to know.
[00:37:36] Simon: So if you're listening, Mr. Zuckerberg, please give me a subject access data request and I'd like to know how long you've been recording me for and how much you hold and how much you know about me. There was supposedly, and this is something, this is... I don't know this, so somebody might be able to back up this information or not, but apparently there's a project called Project Silver Water or Silver Lake or something like that, that is owned by Google, which is the bank of information of everybody. That is, they held... like, as a Google Ads professional, I can do certain pieces of targeting. I might be able to target, let's say you've been searching for golf clubs, and then I could put you into an audience which is in-market for golf clubs, and then I could advertise to you because Google knows that Paris has been going onto a few different e-commerce websites for golf clubs. So I could target intent. The question is, how much intent could we target? Because that extra layer over and above the data that we can target or we're allowed into that layer, i.e. general audience segment, beyond that, how much of that is known?
[00:39:12] Paris: I remember that Google made a big fuss about sunsetting cookies several years ago, and then they, in the end, they never did it actually. I think there was pressure, political pressure to do it, but then that softened and then they backed away and then didn't have to do it. But the amount of data is staggering.
[00:39:33] Simon: Just to touch on one more thing before we go, have you seen this redirect that Google are supposedly pushing everybody through?
[00:39:40] Paris: I don't think so. What is that about?
[00:39:43] Simon: There's a load of the SEO guys talking about... I'll have to pull the exact link, but Google, rather than directing straight to a site, they're putting in a Google redirect in between to land on the site. So it's almost like they're filtering traffic out. But maybe we could talk about that in the next episode, because I think it's almost like they're trying to ring-fence content for Google's AI.
[00:39:54] Paris: Okay. Yeah, interesting. All right. Well, I think we're a little over time, so let's wrap up for today, Simon. But great discussion as always. And for those listening, hope to see you all in the next episode, which is going to be on Friday, same time. All right. See you soon, Simon.
[00:40:25] Simon: All right.
