Culture Is Not a Moral Argument
Culture usually has a positive connotation.
Sam's space
Pet projects, random thoughts, notes about books & art.
Latest posts
apps.samu.space: hobby projects
Over the past couple of years I have been building a lot of small apps on the side: language learning tools, AI toys, finance utilities, a few projects around veganism, AI technologies and other topics. They ended up scattered across subdomains, and I kept forgetting which ones still existed. So I put them all in one place: apps.samu.space .
time off
“Sam, you always do what you want!”
Mohambé morning
Hungarian picked up a lot of everyday words from Yiddish, often without the speaker thinking about the trail. Mázli , meló , balhé , haver : they sit in ordinary sentences next to fully “native” vocabulary, and most of the time nobody stops to ask where they came from. That mix is normal; noticing it is a small linguistic hobby.
speak-it: quick spoken-word presentations
I wanted a simple text-to-presentation workflow: type what I need (a one-minute brief, a quick explainer, something to rehearse out loud), get a short script, then hear it instead of squinting at paragraphs. Speak It is that. The backend drafts the text, ElevenLabs reads it back, and I get a player in the browser. It is available at https://speak-it.samu.space/ .
a local dashboard for my chatgpt export
I finally did something sensible with a cleaned ChatGPT export: I turned it into a static, local-only dashboard. Chart.js for the charts, a 2D UMAP “atlas” of conversations plus an optional 3D sibling view (read more about UMAP here: Understanding UMAP ), and semantic search over embeddings stored in SQLite with sqlite-vss . If you have ever stared at a giant JSON dump and wanted a map instead of…
Human vs AI Intelligence Timeline
I put together a report that compares biological evolution, human technological progress, and AI capability growth across time. The goal was to visualize the compression of intelligence development timescales: from millions of years for hominin cognition to decades for modern AI. You can explore it at intelligence-timeline.samu.space .
Reproducing a Context-Related Failure Mode
I spent some time trying to reproduce the “context rot” effect: the idea that LLMs get worse at retrieving information from the middle of long contexts as you add more tokens. The original “lost in the middle” work showed that models tend to favor the start and end of the context and underperform on content in between. I wanted to see if I could reproduce that with a needle-in-haystack setup on…
composer 1.5 vs claude opus on the same task
I’ve been thinking about what actually drives productivity when coding with AI. The obvious answer is model intelligence: smarter models solve harder problems. But there’s another angle. Most of my daily work is not SWE-bench. It’s adding features, fixing bugs, writing tests, refactoring. The bottleneck might not be reasoning depth. It might be how fast I can iterate.
Speculative Decoding Research Overview
I’ve been digging into speculative decoding lately. It’s one of those ideas that feels obvious in hindsight: why run a huge model token by token when a smaller one can guess ahead and you just verify the guesses in parallel? The field has exploded since the foundational paper in 2023 ( Leviathan et al. ), and I wanted to map out what’s out there. Here’s a tour of the papers I reviewed, plus a…