Everything Is Expensive
Thank you to all of the new subscribers joining us this week. This week, we did receive a question. "Is there a chip shortage? Is it AI-driven?". In short, the answer is "yes," but let's take a more in-depth look by looking at 3 categories that I believe fall into "chips". No potato here keep reading for the meaty part.
CHIPS
| Memory DRAM / DDR5 |
This is the stuff that is used in conjunction with your CPU and GPU to perform tasks. It's immensely important to the AI data centers because it's way faster than pulling information off a hard drive, and AI models need that speed. This put a huge drain on the 3 major manufacturers. What this means is that as a consumer, is since mid-2025, DDR5 RAM for consumers is up 300-400% (ouch!). |
| Graphics cards GPU |
This one is a bit more obvious; it's the backbone of the AI data centers. These huge AI models could not exist without the computational power of GPUs. The GPU that typically gets sold to gamers, those GPUs are now being sold 40-50% above retail. |
| Hard drives HDD |
The last piece of the puzzle for "chips". Are hard drives creeping up? With the latest Claude Mythos model having 10 trillion parameters, it needs ~ 25 TB just to hold a single copy of the model. Keep in mind Mythos is not in wide release right now. But Sonnet 4.7 (100 billion parameters) and Opus 4.7 ( 4 trillion parameters) are in wide release. These models are just getting larger and larger, and it's not just the space requirement for the model but all the outputs and inputs that also require storage space. So to put it in numbers, there's about a 10% increase in HDD retail prices, and many "sales" are disappearing. |
So, subscriber, it's clear to see that yes, chips are in short supply, and yes, it's because of AI.
The Magnificent 7 combined have capex spend this year of ~ $78 billion in Q1 and another $500 - $600 billion of spending through the end of 2026(!!). Not all of that is going into AI infrastructure, but a lot of it is. And this is just 7 companies.
It's startling to even comprehend these numbers. To tie it back to Exploration and the 21st century, all signs point to this time right now being a renaissance of/for/in AI. Which at the moment is.....an age where you are competing with billions in corporate spending for the same GPU, memory, and hard drives. So don't expect to be seeing any relief on computer pricing in the near future.
To leave you with something to chew on, here's the minimum server rack you would need to run Claude Mythos.
HERMES
Below the fold this week is inspired by more reader feedback. The reader writes "I was just reading that people prefer Hermes over Openclaw". This comment took me down the road of exploration once more. So this week, I'm going to give you my rundown of Hermes. In brief, Hermes is, another open-source agent harness, but it came out after Openclaw. If agent harnesses were ice cream, then Openclaw would be vanilla, and Hermes would be, another flavor of vanilla. What I mean to say is that they are both more similar than they are different and both are great. But if you want to know the subtleties in the recipe for each, then you'll have to keep reading.
// until next week
| ◆ Below the fold ◆ |
This area is so you can follow along and setup your own Agentic AI.
THE MODEL HARNESS
A language model is, exactly that, just a model. It has limitations and will always perform at its ability, and as the model parameters go into the billions and trillions, they are getting great. But the big advancement that took these models to the next level is the model harness. The model harness is a collection of Markdown files, tools, memory, and skills. All of these meaningfully improve the performance of any model on benchmarks.
Today we aren't talking about the OGs Claude Code or Openclaw. We are here to talk about Hermes because Hermes is having a moment.
I loaded it up earlier this week. Well, actually, I just prompted Byte to load it up for me. They are both running side by side on the same PC, which I'm happy to report doesn't cause any conflicts as of yet. And they are both running through Telegram chat, which continues to be a quick, easy setup. There was a built-in OpenCLAW migration path (good move), so on install, Hermes picked up all my .MD files from the OpenCLAW instance, parsed through, and loaded the important ones. Onboarding was a breeze, and that is in a time where OpenCLAW is becoming increasingly buggy and inconsistent from version to version.
WHAT MAKES HERMES DIFFERENT
I am going to quote the creators, Nous Research:
"...An autonomous agent that lives on your server, remembers what it learns, and gets more capable the longer it runs."
The interesting part here is how it gets more capable over time. After completing a task, Hermes runs a self-review. This review allows it to consolidate the chat exchange looking for patterns or learnings. When necessary, it creates a skill or memory that can be more easily accessed the next time you need it. In practice, I'm just seeing alerts that say "self-improvement review" from time to time.
ANTHROPIC NOTICED TOO
Is this important and meaningful? Well, just this week at the Anthropic developers conference. They announced they were solving for a problem where an agent makes the same classification error 12 times in a month and has no mechanism to notice because it starts fresh every session. They introduced a state called "Dreaming" that is a scheduled background process that runs between sessions. It reviews past conversations, identifies recurring errors, and good approaches and edge cases and then writes those as a structured learning.
When Anthropic is basically co-opting a core functionality of Hermes, then all signs point to Hermes. It may be short lives but for now, it's HeRmEs MoMenT.
Go try Hermes or OpenCLAW or Claude Code and let me know what you think!
"Would you tell me," said Alice, a little timidly, "why you are painting those roses?"
-- Alice in Wonderland