distributed-systems - Distributed Thoughts

Distributed Thoughts

Sign in Subscribe

distributed-systems

A collection of 14 posts

The Freeze Function

The Freeze Function

The whole stablecoin fight is about reserves and yield and which bank deposits walk out the door. Nobody is pricing the one feature that actually separates a stablecoin dollar from a bank dollar, which is that somebody can freeze it from a console.

Acceptable for Inference

Acceptable for Inference

Google tested its TPUs for orbit and the memory came back with uncorrectable bit flips at a rate it called likely acceptable for inference. Read that phrase again. We are about to make correctness a per-workload dial set partly by radiation, and the model will not tell you which setting it used.

The ASIC Unbundling

ai-infrastructure

The ASIC Unbundling

Inference is two-thirds of all AI compute. Almost all of it still runs on chips designed for training. The market is finally noticing the mismatch.

Agents Don't Live in Data Centers

Agents Don't Live in Data Centers

Agentic AI is the first workload where the physics of latency and data locality aren't optional. The $700B centralized infrastructure bet was built for a different kind of AI.

When the Incumbents Make Your Argument

When the Incumbents Make Your Argument

Dell Technologies, the company that built the infrastructure for centralization, just made the case for distributed AI. When the incumbents start arguing your position, you are either vindicated or about to be absorbed. Maybe both.

OpenClaw and the Architecture Nobody Noticed

OpenClaw and the Architecture Nobody Noticed

Everyone is writing about the OpenClaw acqui-hire as an agents story. It is an agents story. But the reason OpenClaw worked has almost nothing to do with agents and almost everything to do with where the compute happened.

When in Doubt, Go Up: Orbital Data Centers and the Limits of Architectural Imagination

When in Doubt, Go Up: Orbital Data Centers and the Limits of Architectural Imagination

Two weeks ago, a company filed paperwork with the FCC requesting permission to launch up to one million satellites into low Earth orbit. Not communication satellites. Not GPS. Data center satellites. Solar-powered, laser-linked orbital compute nodes, designed to run AI inference workloads from space. The filing included the

Digital Pheromones: What Ants Know About Agent Coordination That We Don't

Digital Pheromones: What Ants Know About Agent Coordination That We Don't

We are overcomplicating this. The AI industry is pouring billions into agent orchestration platforms, coordination frameworks, and centralized management layers. Every week brings a new architecture diagram with more boxes, more arrows, more middleware. And the results so far? Gartner predicts more than 40% of agentic AI projects will be

The Brownian Ratchet and the Chimpanzee Factory

AI infrastructure

The Brownian Ratchet and the Chimpanzee Factory

Two independent projects solving multi-agent orchestration landed on the same primitives. That's not a coincidence—it's an industry figuring out what agent autonomy actually requires.

The Natasha Problem: Why Your Data Pipeline Only Fits One Person

data-engineering

The Natasha Problem: Why Your Data Pipeline Only Fits One Person

For most folks, you probably don’t think about clothing sizes. There’s a number, you pick it, you try on the clothes, and if they fit, then congrats, you’re that number. But how’d they pick that number? And why does every style/line/person fit slightly differently?

A Picture Is Worth Ten Thousand Tokens

ai-infrastructure

A Picture Is Worth Ten Thousand Tokens

DeepSeek's new OCR model achieves 97% accuracy while using one-tenth the tokens by rendering text as images. The computationally "heavy" modality turns out to be the efficient one, which should make us question some basic assumptions about how we architect these systems.

NVIDIA Built the Most Powerful Chips Ever. Google Trained the World's Best AI Without Them.

ai-infrastructure

NVIDIA Built the Most Powerful Chips Ever. Google Trained the World's Best AI Without Them.

In 2020, NVIDIA released the A100 with 54 billion transistors, 80GB of memory, and 2 TB/s bandwidth. It was VERY BIG and VERY FAST(tm). In 2022, the H100 arrived with 80 billion transistors, the first Transformer Engine optimized for the neural network architecture behind ChatGPT, and performance gains

From Pandas to Upstream Control: The Evolution PyData Needs Next

Data Engineering

From Pandas to Upstream Control: The Evolution PyData Needs Next

The Python data stack gave us Pandas, Dask, and Arrow. But we've created a new crisis: the ingest-it-all-first tax that's drowning us in noise.

WTF Do You Spend $300B On?: The Oracle-OpenAI Deal Decoded

data infrastructure

WTF Do You Spend $300B On?: The Oracle-OpenAI Deal Decoded

Discover the challenges and expenses behind OpenAI's $300 billion Oracle deal, revealing a true AI cost breakdown: 80% data prep, 20% training