721
The Code-Only Agent
When code execution really is all you need. Exploring what happens when an AI agent can only write and run code.
New
The latest blog posts to take off on Hacker News. 3,091 posts.
722
723
Starting from scratch: Training a 30M Topological Transformer
Tauformer is a topological transformer (see paper ) that replaces dot‑product attention with a Laplacian-derived scalar (taumode) per token/head, then attends using distances in that scalar space. Below is a post-style overview of the idea and the first training signals from a 30M-parameter run.
724
Agent Psychosis: Are We Going Insane?
What’s going on with the AI builder community right now?
725
Crypto grifters are recruiting open-source AI developers
Two recently-hyped developments in AI engineering have been Geoff Huntley’s “Ralph Wiggum loop” and Steve Yegge’s “Gas Town”. Huntley and Yegge are both…
726
Letting Claude play text adventures
Experiments in cognitive architecture.
727
Slop is everywhere for those with eyes to see
How mindless consumption is making our FYPs worse.
728
Reading across books with Claude Code
In which we teach an agent to read syntopically.
729
730
Data is the only moat
How different adoption models drive better applications
731
Raspberry Pi's New AI Hat Adds 8GB of RAM for Local LLMs
Today Raspberry Pi launched their new $130 AI HAT+ 2 which includes a Hailo 10H and 8 GB of LPDDR4X RAM. With that, the Hailo 10H is capable of running LLMs entirely standalone, freeing the Pi's CPU and system RAM for other tasks. The chip runs at a maximum of 3W, with 40 TOPS of INT8 NPU inference performance in…
732
To those who fired or didn't hire tech writers because of AI
Hey you, Yes, you, who are thinking about not hiring a technical writer this year or, worse, erased one or more technical …
733
The Influentists: AI hype without proof
Why we are losing technical rigor to social hype
734
735
LLMs are a 400-year-long confidence trick
In 1623 the German Wilhelm Schickard produced the first known designs for a mechanical calculator. Twenty years later Blaise Pascal produced a machine of an improved design, aiming to help with the large amount of tedious arithmetic required in his role as a tax collector. The interest in mechanical calculation showed…
736
First impressions of Claude Cowork
Plus Fly’s new Sprites.dev addresses both developer sandboxes and API sandboxes at the same time
737
Stop using natural language interfaces
Natural language is a wonderful interface, but just because we suddenly can doesn't mean we always should.
738
The insecure evangelism of LLM maximalists
I just can't help feeling these training wheels are getting in the way of my bicycle commute.
739
Let's be honest, Generative AI isn't going all that well
A sampling of recent news
740
Fabrice Bellard's TS Zip (2024)
Results and speed for other programs on enwik8 and enwik9 are available at the Large Text Compression Benchmark .
741
Reproducing DeepSeek's MHC: When Residual Connections Explode
DeepSeek replaced standard residuals with massive Hyper-Connections, and then watched them explode. I reproduced the 9x signal amplification to understand their fix: the Sinkhorn-Knopp algorithm. Part 1 of 2.
742
Anthropic made a mistake in cutting off third-party clients
Anthropic may have just committed the biggest business blunder of 2026 -- and we're less than two weeks in. To understand why, let's briefly rewind to 2025, the year when agentic AI went mainstream.
743
Which programming languages are most token-efficient?
Comparing token efficiency across 19 popular programming languages using RosettaCode data - from Clojure to C, there's a 2.6x difference.
744
745
AI is a business model stress test
AI commoditizes anything you can specify. It can't (yet) commoditize what requires ongoing operation.
746
My article on why AI is great (or terrible) or how to use it
Senior engineers are best positioned to benefit from AI. We're good enough to avoid slop, and there's so much we can accomplish. I wouldn't go back.
747
MCP is a fad
Model Context Protocol has taken off as the standardized platform for AI integrations, but this popularity will be short-lived.
748
How to code Claude Code in 200 lines of code
The core of tools like Claude Code, Cursor, and Warp isn't magic. It's about 200 lines of straightforward Python. Let's build one from scratch.
749
750
The Q, K, V Matrices
At the core of the attention mechanism in LLMs are three matrices: Query, Key, and Value. These matrices are how transformers actually pay attention to different parts of the input. In this write-up, we will go through the construction of these matrices from the ground up.