Markets & Mayhem
@Mayhem4Markets

Passionate about markets and startups. Enjoyer of local AI, security fanatic and builder of things. Working hard to make this world a better place.

 United States 馃嚭馃嚫
marketsandmayhem.com
Joined January 2010
44,563 Photos and videos
Agentic coding tasks burn ~1,000x more tokens than chat, and more tokens don't always buy better outcomes. What does? The harness around the model. New explainer, from Hermes and Muse to Claude Code and omp.
Article

AI Agents, Explained: The Model Is Only Half the Story

You have probably asked a chatbot to summarize a document or draft an email. Now the same kind of technology can open a code editor, find a bug, fix it, run the tests, and leave a note explaining what

31
6
35
9,469
When you start chatting with a model or using an agent, everything you share and that is generated back to you is initially stored in the KV cache. But what is a KV cache? How does it work and why does its size and precision matter? Check out my latest.
Article

The Notebook Behind the Model: A Beginner's Guide to the KV Cache

Every time you talk to a language model, it keeps a notebook. Everything you have said, and everything it has said back, gets an entry. That notebook is the KV cache. Once you understand it, a pile of

9
8
27
15,076
Why "pacing" is the wrong answer to a question many are asking about the dangers of AI.
Article

The Open Model That Fought Back

On July 13, 2026, engineers at Hugging Face, the company that hosts most of the world's open AI models, stopped an intrusion that had been running for days. When they sat down to work out what had

6
7
17
32,718
The shift to more local AI is real, but the destination is more of a hybrid approach than abandoning the cloud entirely.
Article

Bringing AI Home: Why Businesses Are Running Models on Their Own Hardware

Two years after the generative AI boom funneled enterprise workloads toward a handful of cloud APIs, a measurable share of the market is moving the other way. Surveys of intent point one way while

48
4
32
22,020
Have you wondered how AI works? Most of today's AI is powered by what are called large language models. This guide breaks down how they work in an approachable and easy to understand way.
Article

Inside the Machine That Talks: Your Guide to Large Language Models

You have probably typed a question into ChatGPT and gotten back an answer that reads like a human wrote it. Maybe you wondered what was actually happening behind the screen. What Is a Large Language

8
6
40
14,198
Running a 70-billion-parameter language model on a single graphics card sounds impossible. A model that size needs roughly 140 gigabytes of memory in its native 16-bit format, far beyond the 24 gigabytes a consumer GPU offers. But EXL3 made it possible!
Article

EXL3: How ExLlama's Trellis Quantization Shrinks Large Language Models

Running a 70-billion-parameter language model on a single graphics card sounds impossible. A model that size needs roughly 140 gigabytes of memory in its native 16-bit format, far beyond the 24

9
5
29
10,943
Nvidia Blackwell GPUs introduced native hardware support for 4-bit floating-point computation through a format called NVFP4. The results are concrete: 2-3x higher throughput than FP8, 3.5x less memory than BF16, and accuracy within 1-2% on large models.
Article

NVFP4 on Blackwell: What 4-Bit Floating Point Actually Delivers

NVIDIA's Blackwell GPUs introduced native hardware support for 4-bit floating-point computation through a format called NVFP4. The results are concrete: 2 to 3 times higher throughput than FP8, 3.5

10
10
70
47,592
Article

Almost 200 Silicon Valley firms just told Trump: ban Chinese open models and you'll kill US startups

The Little Tech Association sent letters to Trump, Commerce Secretary Howard Lutnick, and OSTP Director Michael Kratsios. Y Combinator and Proton are among the signatories. It was sent yesterday. It's
10
16,644
Article

AI firms splurge on millions of printed books at any price

Pre-2022 books are structurally uncontaminated. There's no model collapse, no poisoning, no synthetic text pretending to be human knowledge. The same companies flooding the internet with generated

2
3
22
9,967
Article

How Transformers Changed AI: From a 2017 Paper to ChatGPT and Beyond

In 2017, a research paper introduced the transformer, a type of artificial intelligence model that now powers nearly every modern AI system. Before that paper, transformers did not exist, and AI

4
6
38
19,345
Article

Dense vs Sparse MoE: Architectural Trade-offs in Large Language Models

Conditional computation at scale For most of the history of deep learning, models used the same set of parameters for every input. The introduction of the Sparsely-Gated Mixture-of-Experts layer in

6
9
69
49,320
Article

Agentic Security: Best Practices for Keeping Your AI Workflows More Secure and Private

Agentic AI is Production Infrastructure. Its Security Model Is Not. AI agents with tool access, code execution, and long-horizon autonomy are being deployed into production environments faster than

1
1
14
26,983
Load more