← THE TECH STACK · THE RIDE
STOP 01 / 08
How the machine actually thinks
Principles
MEANING BECOMES DISTANCE — SIMILAR THINGS DRIFT TOGETHER
If you're new
The machine never saw your words. It saw numbered chunks — tokens — and answered by predicting the next one.
If you build with it
Your context limit is measured in these, not words — and every reply you’ve ever paid for was metered in them.
If you architect it
cl100k_base BPE, running client-side on this page — the same encoding your bill is metered in.
▶ SHOW MEEXPLAIN ITBUILD IT
Everything you can drop on your foot
Hardware
HBM → THE DIE
MEMORY BANDWIDTH, NOT FLOPS, IS THE WALL
If you're new
Right now your sentence is being multiplied through thousands of chips built for one job: enormous grids of arithmetic.
If you build with it
The bottleneck is almost never the math. It’s feeding it — memory bandwidth, not FLOPs, is the wall.
If you architect it
HBM stacks and the interconnect fabric are why “add one more GPU” never scales linearly.
▶ SHOW MEEXPLAIN ITBUILD IT
The building the models live in
Infrastructure
CHIP→BOARD→SERVER→RACK→HALL→SUBSTATION
If you're new
All of it happens in real buildings that draw real power and shed real heat into water and air.
If you build with it
A GPU bills the same working or idle — utilization, not capability, decides the economics of this entire floor.
If you architect it
Scheduling, cooling, and substations exist to keep depreciating silicon busy.
▶ SHOW MEEXPLAIN ITBUILD IT
The strangest object in the building
Foundation
~300 GB
OF NUMBERS · TRAINED ONCE
translatesearchdraftcodesupportcontractstriagesummarize

  

    
If you're new
One enormous model was trained once, at enormous cost — everything you use is that artifact, adapted.

    
If you build with it
Scaling laws priced this bet before anyone ran it: loss falls predictably with compute, data, and parameters.

    
If you architect it
Pretraining → alignment → quantization; a few hundred gigabytes of weights is the entire object.

  

  

    ▶ SHOW MEEXPLAIN ITBUILD IT
  






  
  
Which one — and how much is true?

  
Models

  
Claude
ANTHROPIC
ChatGPT
OPENAI
Gemini
GOOGLE
Open weights
THE COMMONS
PRICECONTEXTLICENSELATENCYHONESTY

  

    
If you're new
The different names really do answer differently — and choosing one is a buying decision, not a loyalty.

    
If you build with it
Benchmark honesty varies wildly, and deprecation risk is the cost nobody prices in until the flagship retires.

    
If you architect it
Price, context, license, latency, refusal behavior — the five questions that outlive every quarter’s flagship.

  

  

    ▶ SHOW MEEXPLAIN ITBUILD IT
  






  
  
The plumbing around the model

  
Orchestration

  
RETRIEVALTOOL CALLS
THE MODEL
FROZEN · AMNESIAC
AGENT LOOPMEMORY
YOUR QUERY · RETRIEVE → GROUND → ANSWER — WEIGHTS UNTOUCHED

  

    
If you're new
The model can’t remember yesterday and can’t look anything up. The plumbing around it does both.

    
If you build with it
Retrieval quality caps agent quality — and per-step errors compound faster than models improve.

    
If you architect it
RAG, tool calls, agent loops, memory — docked on without touching a single weight.

  

  

    ▶ SHOW MEEXPLAIN ITBUILD IT
  






  
  
Where the stack touches a human

  
Application

  
You: 
THE OTHER SURFACE — THE AUTOCOMPLETE

  

    
If you're new
This is where AI finally touches you: the chat box, the autocomplete, the meeting summary.

    
If you build with it
Placement beats intelligence. Being where the work happens outperforms being ten percent smarter somewhere else.

    
If you architect it
The copilot pattern — embed in the existing surface, win on latency and context capture.

  

  

    ▶ SHOW MEEXPLAIN ITBUILD IT
  






  
  
What to do on a Tuesday

  
The Playbook

  
01Bad promptsPROMPTING
02It doesn’t know our stuffRAG
03It forgot everythingCONTEXT & MEMORY
04Hand it the whole jobAGENTS
05Prove it worksEVALS
06The bill arrivedCOST
07So did the attackersSAFETY
EVERY GUIDE IS ONE YOU CAN RUN TODAY

  

    
If you're new
Start where it hurts: bad prompts first, then context, then handing work to an agent and surviving it.

    
If you build with it
Without evals you cannot prove any of it works — and the bill and the attackers both arrive on schedule.

    
If you architect it
Seven problems in arrival order, each with a guide you can run today.

  

  

    ▶ SHOW MEEXPLAIN ITBUILD IT