Tag: llm
個人AI環境を30万円で構築する――画像生成、ローカルLLM、そして計算資源の所有
MIRAの写真生成をきっかけに、Mac miniの構成、Draw Things、ローカルLLMによるPython開発、そして計算資源の所有とクラウド利用について考えた対話記録。
Warum AI‑Slop hinausposaunen
LGTM (Looks Good to Me) - Claude Opus 5.5 music ...
A pocket full of personal software
A year-ish spent building software I want, the way I want with *in a low voice* llms
Process Sandboxing & Agents (in 5min)
Linux sandboxing for the win.
Jev by Example
Exploring Jev, TypeSafe AI’s new AI model that returns probabilities instead of prose.
What a Harness Can Promise That a Prompt Can't
A short follow-up to "What Is a Harness?": two concrete checks a Wikipedia citation harness can make that a chat prompt cannot, and why they cost more work.
Bokmässan – allt mer problematisk
Jag besökte årets bokmässa på torsdagen precis som jag brukar göra. Jag kan konstatera att det numera ocks är väldigt mycket folk på torsdagar. Förr brukade det vara ganska tomt på förmiddagen men nu var de smockfullt redan från klockan 10. De stora förlagen, glasögonförsäljare och andra prylsförsäljare tar också upp en allt större plats.
I (kind of) writing code by hand.
Recently, I saw a petition called we write code by hand. I intent to sign it but in the end I didn't. I wrote 100% of the code in the OneCut project by hand but heavily use AI to help to craft it. I'm not that pure to not using AI at all.
With the release of GPT-6, GPT-6 Sol xhigh seems like a good tradeoff between price and performance when compared with Astra and Fable. Will for sure give it a try going forward.
AI is antithetical to learning
LLM agents can write a lot of code for you very fast, but they also take away opportunities to learn and grow.
ベクトル検索だけだと型番や固有名詞に弱い？ハイブリッド検索（BM25＋ベクトル）で回答精度を向上させる方法（シングルサイト編）
1. なぜ「ハイブリッド検索」が必要なのか？ WordPressの標準検索機能に対し、「表記揺れでヒットしない」「精度が悪い」といった不満を解決するため、OpenAI ＋ Qdrant ＋ Groq による超高速なAI検 […]
A functional taxonomy for LLM inference in agentic tasks
A framework for understanding where LLM inference goes during agentic tasks: Initialization, Reasoning, Orchestration, and Synthesis. Consider "Reasoning Yield" as a way to measure how much inference is spent resolving uncertainty rather than mechanically advancing already-determined execution.
Job-Advice
Wenn Ihr dauerhaft mehrere #KI-Agenten gleichzeitig betreuen müsst: RENNT!
Presidential Administrations, Scored by AI
Three frontier models, same question, same scale. The models approached the scoring differently, but agreed on the ordering: Trump-I is worse than Biden and Obama. Trump-II is worse than Trump-I.
Att Sverige har billig el är lockande för AI-företagen
Allt fler datacenter växer upp i Sverige. Bara en del av dem är dock till för AI, De flesta är för helt vanliga data, servrar för onlinetjänster och molntjänster av diverse slag. Det är i huvudsak US-amerikanska storföretag som Facebook, Amazon, Google och Microsoft som bygger dem.
The irrational effectiveness of the Pi harness
A month of coding with Pi, its small agent loop, and the skills and extensions I've added to it.
Privat programmiere ich ausnahmslos ohne ...
Why your prompt cache hit rate is zero: the silent invalidators
Prompt caching is a prefix match: one changing byte near the top of your prompt drops your hit rate to zero. A timestamp is the obvious culprit - here are the rest.
Am I a Rust developer?
deep knowledge, and why everyone using agents is now a site reliability engineer
DINO (Distillation with No Labels)
Breakdown of DINOv3 model that scales self-supervised learning (SSL) for images to produce our strongest universal vision backbones, enabling breakthrough performance across diverse domains.
they know React
No one ever got fired for picking React, but they're probably wasting a shit load of tokens.
ChatGPT Is a Japanese Internet Degen, or: How Fried Chicken Sent Me Down a Tokenizer Rabbit Hole
A stray piece of 2ch boilerplate in a ChatGPT answer sent me down a rabbit hole through Japanese internet history, tokenizers, training data, and one very weird token.
the expensive energy principle
could a computer scientist pay for an embodied brain
Building a Data Explorer for Historical Brewery Directories
Using an LLM OCR and NER pipeline to make 19th century directories morn accessible
Reducing Ambiguity in a SQL Agent using a domain glossary and RAG
A domain glossary can give an LLM agent context it needs to finish a task, but when that is not enough to understand importance, RAG can reduce ambiguity
No cost, no value
Why generating code makes it meaningless, why writing code by hand has value, and why toil is a critical part of the human experience.
Wikipedia and friends: learning through code
I built friendsof.wiki to surface the open knowledge graph alongside Wikipedia articles, and learned a few things along the way.
Right to repair for AI-generated codebases
we have to think through bad ideas, because they may be our future
How LLM Works
This article tears open that machinery and explains what a language model is doing at a mechanical level - why it produces the outputs it does, why identical inputs produce different outputs on different runs, and what “temperature” actually means beyond “a creativity dial.”
Your LLM doesn't know what day it is: date handling for AI tool calls
If you're building AI features into your product, your model has no idea what day it is - and it will confidently guess. Here's how to fix it.
Limited output is a feature
Great startups don't win because they write more code, they win because they think differently.
A computer can never be held accountable
For better or for worse the things you create with LLMs are your responsibility, just like with any other tool.
Hva enn du gjør, ikke spør meg!
(Jo, spør meg)
Open models are decelerationist
people and planet need open models to win
Chronicler
A PaperMC plugin that tracks server events and generates a dynamic in-game newspaper delivered as a written book.
Woche 29 · 2026
How to stop Claude from saying load-bearing
Hack the text output of Claude Code to make life a little bit sillier.
There's no accounting for AI
or budgeting, or hiring, or anything else you took for granted
A LangGraph SQL agent with per-call observability
Building a text-to-SQL agent on top of the canonical LangGraph tutorials, then routing every model call through my own OTel-emitting proxy so I can see what it actually does
e560 — Sweating the Small Stuff
Photo by nik biziuk on Unsplash Published 6 July 2026 e560 with Andy, Michael and Michael - stickshift & phone handsets with Ian Bogost, camping at EMF Camp 2026 & Mountain Quest 2026, brain cooling, dwarf lemurs, World Cup 2026, USMNT, ThreeLions, the tokenpocalypse and a whole lot more! Andy, Michael and Michael get things...
We Are Starting to Sound Like the Thing We Built
“Stay on the road. Keep clear of the moors.“
HODLing gear in 2026
A Month Without Frontier Models
On AI Religion
- Moon Walker
300baud modem or first broadband?
At what stage of AI model evolution are we at?