LLMs Will Benefit from Scratch Workspaces
Current (mid-2026) LLMs are quite powerful. However, (prior to addition of agent harnesses and loops) LLMs vendors don’t claim to include an effective scratch workspace. Current LLM designs, …
The Devouring of All Knowledge - Ryan Fleck
I was horrified but unsurprised to read reports1 this week of the AI/LLM2 companies buying, butchering, scanning, and shredding rare books at an unprecedented scale. With all online forums, PDFs, articles, and books consumed, the next best source of human writ…
Personal website of Chad Arimura - Buzz in Action
I was having a conversation with friends this weekend about inference as a core property of intelligent systems. In the context of machine learning, inference is the moment a trained model is put t…
Software Engineer
The Power of Focus in Agentic Coding
Alex Graveley, the creator of GitHub Copilot, showed me how I've been thinking about AI coding wrong.
Kimi K3 Architecture Notes
Short architecture note on Kimi K3, including LatentMoE, Kimi Delta Attention, Attention Residuals, NoPE, multimodality, and inference-efficiency choices.
Evals before prompts: building an LLM OCR for KYC
Building an evaluation framework for production-grade document extraction.
Golden hour is almost never an hour. Using the NOAA solar-position algorithm we measured golden-hour length for 27 cities — from ~24 minutes at the equator to about 58 minutes in Berlin in December. Not one reaches an hour.
Evolution can surprise you sometimes.
Zenzic v0.24.0 through v0.26.0 was not a feature sprint. It was a three-step architectural correction: interaction, stabilization, and deterministic global scoring.
A write-up on why I don't give in to the hype and switch to Lean for formal verification of programs.
A QA engineer walks into a bar. Orders 1 beer. Orders 0 beers. Orders 999999 beets. Orders -1 beers. Orders a lizard. Orders NULL beers... Then a customer walks in and asks where the bathroom is. The bar bursts into flames. You probably know this joke. It's a …
I Designed A Custom PCB To Avoid Pressing A Button Three Times
My night light always booted up too bright, so I designed a PCB that presses its button for me.
Six months of trying to make a 1.2B model useful, and the measurement mistakes I made along the way.
Substack writers, you need a website!
Substack is convenient, but treat it as a distribution tool for your content. Your content should live on an independently owned website, which has a domain name you control and own. This way, you …
Guide to the DeltaNet Family of linear attention mechanisms.
Using an open model feels surprisingly good
advanced-context-engineering-for-coding-agents/benchmarking-opus-5-on-slop-code-bench.md at main · humanlayer/advanced-context-engineering-for-coding-agents
Contribute to humanlayer/advanced-context-engineering-for-coding-agents development by creating an account on GitHub.
気象庁
AI Development with Kiro
Anyone who has been writing any kind of code in the last number of years has undoubtedly interacted with AI coding assistants to some extent or another. These coding assistants have been around for a
Basement Programmer is back... Kind of
If you have been following my articles for a while, you will notice that the Basement Programmer blog has been shockingly quiet lately... Many tech Blogs out there start with a bang, and then fizzle o
Our position on open-weights models
Anthropic CEO Dario Amodei on open-weights models
Judge Rejects Google’s Attempt To DMCA Its Way Out Of Being Scraped
Back in December we called out Google for filing a DMCA 1201 lawsuit over companies scraping Google’s results. Almost everything about the lawsuit seemed problematic, not the least of which i…