Lots of AI Marketing News  🔗 **Source:** [Lots of AI Marketing News](https://share.snipd.com/episode/21215384-28a4-41be-8dc2-4ed9b226ea70) ## Highlights - **OpenAI Lost The Video Arms Race** - OpenAI closed Sora because competing video models and compute efficiency advances left it unable to compete. - Christopher Penn points to Google's TPU ownership and TurboQuant compression as reasons rivals scale cheaper and faster than OpenAI's GPU-dependent stack. Transcript: Christopher Penn I mean, OpenAI is in kind of a really tough spot at the moment. They're continuing to burn cash like crazy. They're kind of a revolving door for talent. And they just raised a whole bunch of money. But their models and what they're doing in the space isn't as good as it used to be. When you look at what competitors are doing, particularly Anthropic, also a lot of people, very understandably, were not thrilled about Sam Altman's behavior when it came to the whole Anthropic Department of War business. How OpenAI pretty clearly tried to cut Anthropic off at the knees. And so the business is facing a lot of challenges. But Sora, as a video model, it was good, but not as good as what's available now. When you look at what's on the market, like Halo, Kling, you look at Google's VO 3.1, then back to a new version of VO 3.1, VO 3.1 Flash just came out last night. These models are just killing it in terms of ByteDance's C-Dance 2.0. C-Dance 2.0 is so good that it immediately spawned lawsuits from major Hollywood production houses because the demo video they made of Brad Pitt fighting Tom Cruise in a fistfight Was incredibly high quality. And obviously, Chinese AI companies are not bound by Western IP. So ByteDance basically told Hado to take a long walk off a short pier. But Sora couldn't keep up. And OpenAI stretched so thin, trying to keep services running, trying to keep building new models. And what's interesting is from a very tactical perspective, it doesn't appear, this is an outsider's view, I don't have any inside information, but it doesn't appear like they're Publishing and producing like the latest and greatest. Google, for example, very recently published a new paper called TurboQuant that showed how you can take open models and dramatically compress them. We're talking six to eight times more efficient. Google basically came up with this algorithm. It's all just math, an algorithm that makes data centers, particularly theirs, huge surprise, six to eight times more efficient. So they can serve, obviously, six to eight times more customers. That's stuff that companies like OpenAI don't have on deck. OpenAI, in particular, doesn't have its own hardware. This is one of the reasons why Google is just eating everything because Google owns its own TPUs, the Tensor Processing, so they can scale way faster. They are so efficient that Anthropoc just signed a multi-year, multi-billion dollar deal to rent more compute time from Google, which is always fascinating. So that's a lot of what's happening. And OpenAI is basically saying, yeah, Sora melts our GPUs. And as a result, we have less capacity to generate, to work on catching up to where the rest of the industry is going. ([Time 0:00:54](https://share.snipd.com/snip/561ae3b4-c3e4-488b-a122-999262114feb)) - **Use CLI Tools To Save Tokens And Time** - Use command line interfaces and APIs to avoid burning token budgets and speed up AI workflows. - Christopher Penn explains terminal tools let AI call programmatic actions directly (e.g., update Google Docs) instead of sending huge model context back and forth. Transcript: Christopher Penn The big thing there is that command line interfaces, which are text-based tools that run in a terminal, very 1983, right? Sort of Reagan era computing, the black screen with the green text, are the hottest thing since sliced bread in the last month or so, because it's turning out that the way that generative AI tools like CloudCode and stuff have been accessing resources through a system called Model Context Protocol, or MCP, is good, but very, very slow. And it burns your usage rates. And as tools like Claude Code throttle down on how much you're allowed to use. Anthropic put out a post recently saying because of the massive influx of customers, we're having to throttle how much people can use during peak hours. So much so that it's like a doubling of the restrictions on accounts. So you burn through your five-hour limit much, much faster now. When you use a command line tool, like the one I was using this morning, it's called Google Workspace command line interface. It doesn't use AI. So it's all programmatic text-based interface, sort of Zork, if you will. And because it doesn't use AI, it doesn't eat into your token budget. So when you're using a tool like Cloud Code, you can say, here's a link to a Google Doc. Here's the GWS command line tool. Update my Google Doc using this tool. So it no longer has to send huge amounts of text back and forth through the MCP. Now it can just open up its own command line and essentially just type in the command line for you to do these things. It is much, much faster. It is much, much cheaper. And most companies now that are sort of AI forward companies have command line tools that are super fast, very lightweight, cost nothing other than the software that you're already Paying for. And as an added bonus, if you're old like me and you're comfortable in a terminal, you can also use them. The interface is plain language. So if you wanted to, you could type in and say, update my Asana with this and use the Asana tool or update this Google Doc with this or create a new Google Doc with this and just typing it into The terminal. One of the things I said recently is that if a company, a software company in particular, does not have an API and a command line tool to talk to it, they're dead. They're toast because what people are doing is they're building harnesses using tools like Claude Code and OpenCode and so many of these other ones that use these command line tools. And it's fast, it's easy. And if your software is so difficult to use because it doesn't have that kind of tool, companies that are moving to an AI forward position are going to say, well, you know what, we're not Going to renew our subscription because this tool is harder to use than the one that we can have AI use. A real good example is stuff like marketing automation software. Marketing automation software is ripe for tools like CloudCode to take on because it's all the same repetitive stuff. Schedule this email campaign, schedule this drip campaign, do this. We all know how to do that. ([Time 0:05:47](https://share.snipd.com/snip/3a9ecfa8-32d5-4774-8ec8-9f7b601c8404)) - **Found A Hidden API Via Browser Network Logs** - Christopher Penn reverse-engineered an undocumented API by exporting the browser network audit trail and feeding it to Claude Code. - Claude identified hidden endpoints and helped craft calls by asking Penn to record five specific interactions to infer internal API keys and endpoints. Transcript: Christopher Penn In a browser like Google Chrome, if you go into the view, it's a view menu and go into the developer mode, developer tools, in there, you will see a tab called network and network tells You what the browser is doing. It tells it, it basically is watching communication back and forth. And so when I had, when I was browsing this piece of software, the network tab allows you to export the audit trail of what you did in a session. And I handed this to Claude Codes. Claude's like, well, there is an API. It's not documented, but there is clearly an API there because they built it for themselves to make their own software work. And I can just use that. And I said, great, tell me the things you need me to do. And so it gave me a list of five tasks to perform and record the interactions. And from that, I was able to extrapolate what the API was, what the API keys were internally that are in plain text in this network tab. And it built basically an API interface for the software that doesn't offer one publicly because it's all visible in the network communications. And so that's kind of what happened there. I don't understand why this particular company does not offer this API because it would improve stickiness of the customer. Like if I had a API that I could write to this software, I'd be inclined to stay a customer longer because now my AI agents can go do that work for me while I'm doing other things. And I can just come back later when it's done. Instead of having to reverse engineer their own API to make their software work. I don't understand that. ([Time 0:12:20](https://share.snipd.com/snip/783deb50-63f4-4832-9958-51edb22b1ad2)) - **Diagnose GEO Before Buying Tools** - Don't buy GEO tools before diagnosing whether you have a GEO problem; attend a short assessment instead. - Christopher Penn is offering a free 25-minute webinar to classify GEO into three phases and advise if tooling is necessary. Transcript: Christopher Penn Free. John Wall That's the most important part. Christopher Penn It's a 25-minute webinar on helping understand if you have a GEO problem to begin with. So assessing what GEO is at a very high level and then helping people walk through the three different phases and understanding, do you have a problem? Because you might not. And if you don't have a problem, then you don't need to go and buy a bunch of stuff. I would discourage buying a bunch of stuff anyway. There's a lot of tools that are promising sort of AI visibility and AI brand monitoring and stuff. I've actually sat through demos and tech demos and stuff of these tools. And when you ask quietly on the side, exactly how do you get this data? And they tell you and you go, but that's not how LLMs work. Your tool is worthless. Your brand visibility thing is worthless. The methodologies that a lot of people are using behind these tools are so bad because they don't understand how AI models predict things to generate text. Our friend Rand Fishkin over at SparkToro did a thing in January where they were testing to see, how many times did you have to put the same prompt in, copied and pasted, to get the same Brand results for the prompt in the same order? For Claude, it took 1,400 tries before they got Claude to return the same results in the same And for Google, it was like 149 times or something along those lines. So a tool saying, you're rank number three in ChatGPT or whatever. No, uh-uh. These systems, I mean, the bottom line is they're all bold. Anything that's promising, they know where your brand is in an LLM is just lying to you. Okay, so we've got that. John Wall And then we've talked before about the Trust Insights page scanner that you can just get a report of how your site's doing. We've actually got the pro version open now that will scan the whole site. That's not a free tool. You have to talk to us about getting that, but definitely check out the single page analysis and go run that. And if you want to learn more, we're happy to talk to you about getting the rest of your site going and there's additional scans and we can also do the work for you if you don't want to deal With the headache of picking that up. I don't know, any other stuff on the pro scanner that is worth talking about? Christopher Penn Well, so I mean, the scanner itself, both versions address phase three of GEO. So remember I said there's three phases. Phase one, you know, with the model itself, this does not address because you can't, I just said, like any tool that says you can is lying. We are not in the business of lying. So this tool addresses phase three of GEO and then phase two. ([Time 0:14:18](https://share.snipd.com/snip/8508e570-1d78-4800-8690-1d80d81c7bdf)) - **Google Cuts Off Pages After Two Megabytes** - Google only indexes the first ~2MB of a page, so heavy headers and JS can hide content. - Christopher Penn recommends refactoring heavy resources into separate files and using AI agents to render and test what Google actually sees. Transcript: Christopher Penn Big thing in that article that was really worth paying attention to is Google said, we will grab the first two megabytes of a page and that's it. After that, we cut off anything that's there. So if you have a page that is loaded up with stuff, JavaScript and things like that, and style sheets and things that consumes that two megabyte portion, Google might not actually see The rest of your page. So like if you're, if the page content is way down, you got all this crap up top in the header that diminishes their ability to see the full content of the page. So they were very clear, like you need to refactor those kinds of pages to move the heavy resources to separate files so that your page is as lightweight as possible. And that's, you know, that's the big takeaway there. And this is some, again, this is stuff that you can do with generative AI, with AI agents. You can plug in a tool like OpenWork or Claude Cowork and say, here's the rendered HTML of my page. This is what Google sees. A, how in danger am I being cut off at the two megabyte limit? And B, how do I fix it? Like I use WordPress. I even on this hosting service or whatever, I've got these plugins help. And the tools can either tell you what to do, or in the cases of tools like OpenWork and CloudCore, they can just do it for you. ([Time 0:17:40](https://share.snipd.com/snip/e81a3505-d6f3-42d4-9101-8d2e420f9438)) - **Early Signals Show Weak iPhone Air Demand** - Early reports suggested the iPhone Air may have poor sales and inventory issues, with Europe seeing a €200 price cut. - John Wall said he hasn't verified claims but found little user enthusiasm online and plans to investigate further. Transcript: John Wall Yeah, so we were kicking around the iPhone Air. I went to the whole online community to see who was using it. And interestingly, I didn't have anybody come up and say, yeah, I've got an Air. I love the thing. And then I did come across some articles this week. I haven't verified them. There was stuff out on threads and it's not verified, but there's talk about them not selling at all. Like the majority of them still being in inventory and not activated. And there, they did a price cut of 200 off in the European market already, I think. Yeah I just so we were definitely not on the mark of kind of like looking at this and saying well who is this really for and is there anybody really looking for this thing so I will dig in a Little bit more on that and verify that stuff before I start really running my mouth but yeah kind of unsurprising that battery life or you know cheaper easier is much better than thinner For thin's sake I've got a bunch of stuff coming together on Reddit. ([Time 0:21:03](https://share.snipd.com/snip/9571a397-c834-4c91-9831-866ae07810a8)) - **Red Team Your Stack With AI Agents** - Run AI-driven red team tests now because autonomous agents make break-ins trivial. - Christopher Penn urges using tools like Claude Code or OpenWork to probe your sites and reduce third-party plugins to minimize attack surface. Transcript: Christopher Penn Not on that. But this past week was a really rough week for the software community in general, because it turns out that AI agents are really, really good at breaking into things. The cybersecurity community has been sounding the alarm on this for a couple of years now. But now with agents like OpenClaw that can act autonomously, it's turning out that what we thought were secure systems were, in fact, not even remotely secure. And everything from injections into major uh software packages to just breaking into people's websites is trivial for a lot of these tools uh so the the short version there is now would Be a great time for you to use these tools to try to break into your own stuff and do a sort of a red team exercise, you can have, you know, Claude Code open work, Claude Code work, whatever, Whoever the tools are in your stack, do that and try to find vulnerabilities and patch them before somebody else does it to you. So that would be my caution and warning about what's going on in software right now. It is very, very bad out there. And it turns out that a lot of systems embrace security by obscurity, but are still basically using password or one, two, three, four, five as the combination for everything. John Wall Yeah, that's right. It's the way to get it. Now, two-factor authentication, I'm falling back on that. As long as at least I have that set up for most things, that's a solid way to go. Is there any other best practices that people should be doing just to, you know, kind of get rid of the lion's share of this risk? Use as few third-party dependencies as possible. Christopher Penn If you have a WordPress site and you've got like 82 plugins, now's the time to start reducing the number of plugins. And again, with tools like Cloud Code, you could say, here's all the plugins that I use on my WordPress site. And here's why I use them. Here's what I use them for. Write me a master plugin with better security practices. Have it generate that. Like I did this over the holiday break for my own website. I took five different WordPress plugins that I was using, just cloned the features that I use the most, got rid of the rest, and then ditched those plugins. And so part of my process is when I do QA and auditing this is to say like, hey, these are the known security issues with WordPress plugins. Make sure that you don't do X, Y, and Z. Check the code for this and stuff to make more robust versions of existing software. ([Time 0:22:50](https://share.snipd.com/snip/9093d913-e30f-49c9-9053-1c501fd39033)) - **Consolidate Plugins To Reduce Risk** - Reduce third-party dependencies and consolidate WordPress plugins into a single secure plugin. - Christopher Penn described cloning needed features from five plugins into one custom plugin to lower risk and simplify maintenance. Transcript: Christopher Penn A WordPress site and you've got like 82 plugins, now's the time to start reducing the number of plugins. And again, with tools like Cloud Code, you could say, here's all the plugins that I use on my WordPress site. And here's why I use them. Here's what I use them for. Write me a master plugin with better security practices. Have it generate that. Like I did this over the holiday break for my own website. I took five different WordPress plugins that I was using, just cloned the features that I use the most, got rid of the rest, and then ditched those plugins. And so part of my process is when I do QA and auditing this is to say like, hey, these are the known security issues with WordPress plugins. Make sure that you don't do X, Y, and Z. ([Time 0:24:30](https://share.snipd.com/snip/4ca8b1d7-d3ac-4abc-99c4-9f755c383594))