Learning Library

← Back to Library

Mimetic Defense Against AI Hype

Key Points

  • The speaker defines “mimetic defense” as the habit of questioning and counter‑acting meme‑like ideas—especially AI‑related hype—that spread like mind viruses and shape perception before facts are considered.
  • He highlights common misconceptions, such as the belief that a single ChatGPT query uses huge energy (when in fact watching an NFL game on a big TV consumes far more) and worries about water usage, noting that major cloud providers are moving toward water‑positive data centers.
  • Citing an IBM study, he points out that about 75 % of corporate AI projects miss ROI targets, illustrating how media hype about AI’s transformative power obscures the high failure rate and the need for realistic expectations.
  • Real‑world examples like CLA’s failed AI‑driven customer‑service rollout, which forced the company to rehire human agents, demonstrate why a strong mimetic defense is essential for avoiding costly hype‑driven decisions.

Full Transcript

# Mimetic Defense Against AI Hype **Source:** [https://www.youtube.com/watch?v=Wd1fC1AZ6EQ](https://www.youtube.com/watch?v=Wd1fC1AZ6EQ) **Duration:** 00:07:34 ## Summary - The speaker defines “mimetic defense” as the habit of questioning and counter‑acting meme‑like ideas—especially AI‑related hype—that spread like mind viruses and shape perception before facts are considered. - He highlights common misconceptions, such as the belief that a single ChatGPT query uses huge energy (when in fact watching an NFL game on a big TV consumes far more) and worries about water usage, noting that major cloud providers are moving toward water‑positive data centers. - Citing an IBM study, he points out that about 75 % of corporate AI projects miss ROI targets, illustrating how media hype about AI’s transformative power obscures the high failure rate and the need for realistic expectations. - Real‑world examples like CLA’s failed AI‑driven customer‑service rollout, which forced the company to rehire human agents, demonstrate why a strong mimetic defense is essential for avoiding costly hype‑driven decisions. ## Sections - [00:00:00](https://www.youtube.com/watch?v=Wd1fC1AZ6EQ&t=0s) **Mimetic Defense Against AI Myths** - The speaker describes “mimetic defense” as recognizing memes as mind‑viruses that skew perceptions of AI—such as exaggerated energy‑use concerns—and advocates clear, factual analysis to counter these misconceptions. - [00:03:58](https://www.youtube.com/watch?v=Wd1fC1AZ6EQ&t=238s) **Evaluating AI Model Viability** - The speaker stresses the importance of proof of cost, proof of time to utility, and tracking signal‑to‑buzz to determine whether expensive frontier models like GPT‑4.5 are genuinely practical for enterprise use. ## Full Transcript
0:00I want to introduce you to a concept 0:02that I actually practice a lot uh on 0:05this channel uh on my substack is called 0:08mimetic defense. The idea is memes work 0:11like mind viruses and AI in particular 0:14has a incredibly powerful hold on our 0:18imagination right now and AI news 0:20articles often act like viruses. They 0:22infect our brains. They shape our 0:24perceptions before we think about what's 0:26really true or not. Uh I worked on this 0:29uh on my Substack over the weekend when 0:31I wrote about how we have dramatically 0:35misunderstood the energy consumption 0:37that an actual chat GPT query costs. I 0:40don't think many people realize this, 0:42but it is vastly vastly more 0:45energyintensive to run your big screen 0:48TV and watch an NFL football game for an 0:50hour than it is to use a chat GPT query 0:53on the order of hundreds of times. 0:55But people don't think about that 0:57because the they have no mimemetic 0:58defense. The meme of energy consumption 1:01has entered our brains and shaped the 1:04conversation. That's true if you think 1:06about other critiques of AI as well. The 1:08water critique, people don't realize 1:10that major cloud manufacturers for cloud 1:13computing have said their data centers 1:15that they're building are going to be 1:16water positive within 5 years. Uh and 1:19they're working on recycling water right 1:20now. So I could go on and on, but you 1:22get the idea. Mimemetic defense isn't 1:25particularly sexy, but it is super super 1:27important and it's a lot of what I focus 1:29on is just being very clear and factual 1:31about what's really going on. And it's 1:33not just about AI safety or AI 1:35sustainability. Nimemetic defense also 1:38pops up when we think about hype claims 1:41when it comes to AI. A recent study that 1:44came out uh showed that 75% according to 1:47senior executives, this was an IBM 1:49study, uh 75% of projects that uh are 1:53launched inside companies for AI fail to 1:56meet executive ROI targets. They end up 1:59being classed as misses and only 25% 2:02would be considered meeting or exceeding 2:05ROI 2:06targets. That's pretty terrible. Like 2:08that's not great. Now, I for one think 2:11AI is actually high leverage enough that 2:13those 25% alone may justify the 2:16investment if you get them right. But 2:19that doesn't make the fact that we have 2:20a 75% miss rate on AI acceptable. And it 2:24certainly isn't the story that most of 2:26us see in the news. In the news, most of 2:31us see headline after headline after 2:33headline with AI is taking jobs, AI is 2:36taking over this and that and the other 2:37thing. And we miss stories like CLA's 2:39where CLA committed to firing 700 CS 2:43agents in 2024 and then they had to 2:46rehire customer success because their 2:48vaunted AI agent did not do the job that 2:51it was supposed to 2:53do. Again, we had no mimemetic defense 2:56against that kind of hype. We were 2:59vulnerable to it. We were susceptible to 3:01it. So, I'm a practical guy. I want to 3:04tell you what are some of the principles 3:06I go 3:07for that are principles that underly the 3:11medicic defense that shape a lot of how 3:13I do my YouTube that shape a lot of how 3:15I do my substack. Number one I predose 3:19with reality checks. I am really honest 3:22when I give you claims like this. CLA 3:24has claimed right 700 agents and so on 3:27will be replaced by AI and then I walk 3:31through the missing context like you'll 3:33see me do this I'll say okay this is the 3:34claim that was made this is the context 3:36that we didn't talk about I think that's 3:38really important predosing with reality 3:40checks builds your mimemetic immune 3:42system principle two is demanding some 3:45proofs if there's a big claim I tend to 3:47say show me a demo not a slide this is a 3:51big problem that Devon had when it 3:52launched it eventually launched, but 3:54when it launched, it was just a video. 3:57It wasn't actually a workable 3:58demo. Show me a proof of cost. This is 4:03actually one of the big issues that 4:04OpenAI has right now with their frontier 4:06models. They're having to put 4:09GPT4.5 out of beta, out of production 4:12because it the rumor is it was so 4:14expensive to run. The compute and output 4:17tokens were so expensive. 03 is also a 4:20very expensive model. And so when we 4:22talk about intelligence, we need to talk 4:23about what is the cost of that 4:24intelligence. This is actually an area 4:26where Google has been really relentless 4:28about cutting costs and making it easier 4:31for 4:31developers. The third one I'll call out 4:34is proof of time to utility, especially 4:37if you're talking about enterprise 4:39architecture and enterprise change 4:41management. It is not fast to make these 4:44changes. And promising that it will be a 4:46change overnight is not honest. It's not 4:50transparent. it's not true. And so I 4:53tend to actually look pretty hard at 4:54proof of work, proof of cost, and proof 4:56of time to utility because I want to 4:58know, right? Like is this something that 4:59actually is going to stand up or not? 5:02The third principle, so we talked about 5:04predosing with reality checks. We talked 5:06about demanding some proofs. The third 5:08principle is tracking signal to buzz, 5:10right? Is this 5:13only mimemetic headlines or is there 5:16actually something here where I can read 5:18a case study? I can get into the 5:20substance. This is where I actually 5:22really want to appreciate what Claude 5:24and OpenAI have done in exposing model 5:26cards. I think it's a good choice. I 5:28wish that Grock would do it. It's really 5:30important to have solid technical 5:32insights that underline claim technical 5:34advancements. 5:37The fourth piece of mimemetic defense 5:39you can adopt in your personal immune 5:40system is stress testing second order 5:43effects. So you can ask yourself, learn 5:45to ask yourself the question, what if 5:47this piece of hype is real? What is the 5:49second order effect? Let's say the demo 5:51works. What happens when Devon writes 5:54code out of its training data set that 5:57is somebody else's copyrighted code on 6:00accident? What happens when Devon puts a 6:04PR into 6:05production and nobody checked it because 6:08they trusted Devon and now there's a bug 6:11in production. Is Devon liable? Is the 6:14engineer liable? Now, I am sure that 6:16Devon's lawyers have written up the 6:18terms of service to answer those 6:20questions. So, I don't mean to pick on 6:21Devon per se. I'm not saying they have 6:23particular gaps there, but it's an 6:25example of the kind of question that you 6:27need to ask when you get these hypy 6:29headlines. It helps you to build your 6:30mimemetic immune system. The fifth one I 6:34want to call out is closing with 6:35constructive 6:37skepticism. If you're evaluating a meme 6:39that comes through that's hyper hyper 6:42attentiongrabbing that just gets into 6:44your lizard brain really 6:46fast, get yourself into the mode of 6:49saying, "Here's what would make me 6:51change my mind." Ask yourself, "What 6:53evidence am I looking for that would 6:55make me change my mind on this 6:56attention-grabbing headline?" And and 6:58name it. and then be on the lookout for 7:00it. It trains you to be a bit more of a 7:03critical thinker. I doubt that you'll be 7:05surprised. If you follow my channels 7:07anywhere, you know this is kind of how I 7:09work, but I wanted to expose it. I'm 7:11basically trying to build an AI hype 7:14immune system here. I want us to get 7:16from AI hype to AI productivity. And 7:19that means building an immune system 7:21against the worst parts of AI hype. So, 7:24I hope these five principles of 7:26mimemetic defense are helpful. I may do 7:28some more writing on this, but I wanted 7:30to at least call them out uh and share 7:32them here.