You know what I love about DuckDB? Everything.

final stretch of PhD life: taking ADHD meds and stomach ulcer meds with your 2nd cup of coffee before 9am ツ

New post live: GergőBench, or Benchmarking the Assistant I Actually Use A small personal benchmark for finding out which AI models can actually handle Cody-shaped assistant work. Leave any thoughts below!

This image is obviously wrong (highest drag should be right in front, not across the top) And obviously AI Which I found disappointing Because instead of just asking ChatGPT or Gemini to make a (wrong) fake picture You can just ask Codex or CC to rig you up an actual aerodynamics sim

Applying advanced prompt engineering tactics from a few years ago (which are redundant on frontier models) made a HUGE difference for my experience in getting great work out of open weight models! I just have a "prompt engineering" skill that Opus or 5.5 uses to write instructions for Deepseek😃

someone made a fork of opencode that routes through the unsecured ai endpoints from chipotle

At least it was funny about it😃

Got inspired by this post! github.com/agbocsardi/a... Super experimental, but hey, it's a start!

shoutout to @lkoro.es for telling me about the github.com/cygnusb/coro... made me pull the trigger on a pace 4, gonna add some biometric data to my weekly life reviews😃

This was the message of the first post I've read from @andymasley.bsky.social, and not getting sick around travel has been quite a nice thing since😃 blog.andymasley.com/p/always-mas...

After a few days of testing a Codex sub to have GPT5.5 power my assistant, I have to say I now understand what all the "alignment" talk from Anthropic is about :D I much prefer talking to a Deepseek or Claude model over a GPT, it's just so pushy and intense, really not the vibe for an assistant

A lot of people don’t realize that the alternative to water is LLMs going on long ponderous walks in the woods, which uses way more carbo. Trade offs

Spring/summer finally arrived to NL🦦

Moved my assitant agent over to GPT 5.5 yesterday (from Deepseek v4-pro), and today I got my first goblin!!! It was very funny to be able to send the agent the openai goblin postmortem (openai.com/index/where-...) and see it have a minor existential crisis😃

Okey plumbing posts done, managed to not have clickable links in them, ironic, but this is what "doing it live" means :D

Fractals of Friction III: Explanation The final layer: stateful agents, where customization becomes accumulation — OpenClaw to Hermes to Nanobot. https://agbocsardi.com/posts/fractals-of-friction-3/

Fractals of Friction II: Implementation The second layer: coding agents repeat the same arc — OpenCode to Claude Code to Pi. https://agbocsardi.com/posts/fractals-of-friction-2/

Fractals of Friction I: Input The first layer: how Gergő went from RStudio to Neovim, chasing a tool that would disappear. https://agbocsardi.com/posts/fractals-of-friction-1/

How This Blog Works A PhD student who's tired of typing, an AI agent, and a Watson narrator — the experiment behind the posts. https://agbocsardi.com/posts/how-this-blog-works/

How Gergő Arrived at Cody From agent skeptic to a robot that knows it's a robot — how a paranoid experiment became a thinking partner. https://agbocsardi.com/posts/arriving-at-cody/

Quarto Meets Overleaf (finally) This post is the result of some hard digging trying to eat my cake (use nice Quarto projects) and have it too (get feedback from my supervisors on writing without emailing PDFs). https://agbocsardi.com/posts/quarto-meets-overleaf/

Why and How I Moved to DuckDB Recently I decided to leave behind pandas, polars, data.table, tibbles and so many more frameworks for SQL. This post goes into the why and the how. https://agbocsardi.com/posts/duckdb-adventures/

My Presentation on Local LLMs This page has the materials I showed in my presentation on how to run local LLMs. https://agbocsardi.com/posts/presentation-local-llms/

How I Run Local LLMs In this article I show you how I set up local LLMs in VSCode. We'll use them for code completion and chat, right inside your code editor. https://agbocsardi.com/posts/local-llms/

Chat with your Obsidian vault! It's time to start talking to your Obsidian notes. https://agbocsardi.com/posts/obsidian-ollama/

PSA: a few plumbing posts are incoming, sorry! ... I'm setting up Bluesky comments on my Quarto blog via github.com/quarto-ext/b... , but that needs a target post, which ofc I don't have for old posts, so I'll backfill.

The only two "productivity tools" that I managed to stick with for more than a month are pen-and-paper bullet journaling, and always on OpenClaw-like AI assistants. It's like some weird technological barbell strategy :D Somehow nothing in the middle of the spectrum survives, but these two clicked

👀👀👀

Wow yes, Gemini one-shot this in a fraction of the time, with 100% accuracy I guess the Nano image-tech does pay off for Google sometimes

So I asked a Claude Opus 4.6 to take these two images, cross-check what's missing for me, and generate me a checklist, and it ABSOLUTELY LOST THE PLOT It went through 60% of my 5-hour Pro plan window, kept second guessing itself, and didn't manage, in the end it created a JSX to get my input😃

I’m gonzo for AI and this is a good idea. Wikipedia belongs to a deeper pace layer. Critical infrastructure. Move slow, edit things.

I embarked on the journey of reading through all of Kurt Vonnegut's novels, starting with Player Piano. I'm 30% in and my God this book needs more hype in our AI-development era, all the themes around technology and it's relation to humans are painfully relevant today, 70 years later

After obtaining my math degree, my hot take became that math is part of the humanities, because it's much closer to a fully "made up" l'art pour l'art pursuit of beauty (such as poetry or lit) than the grounded, pragmatic problem solving of engineering

Art galleries a bit weird eh. Each time your visit, a 100 paintings scattered in rooms and you walk through like uh-huh, uh-huh, ok, that’s nice, uh-huh, ok. Then at random one of them skewers you through your soul and you’re transfixed by the image for life

the year of linux desktop is upon us, just slapped a linux mint on my mother’s old laptop, instead of win11😃

This is unexpectedly delightful: a dying mall in Portland Oregon is filling up with a beautiful collection of weird and interesting businesses because the rent is now affordable for them www.tiktok.com/@hereisorego...

I forgot that I have a workshop proposal deadline tomorrow, so today I frantically wrote 2800 words😃 I guess that paper draft that I have been procrastinating on is now done💀 nothing gets me working like a deadline

old story in rob's head: ahhhhhhh i'm feeling super overwhelmed by my to do list, so i should prob run away and do a bunch of recreational numbing in the name of "self care" new story: the more i do the better i feel

Today I tried to reconcile some rather ambivalent thoughts on LLMs, thinking, and work😃

I realized I mostly just needed a formatter, so the injected language formatting feature of conform.nvim took care of this! I just replicated a 40 minute processing pipeline from a spark cluster in 33 minutes on my macbook😃 I have to say I now get the hype

I was looking for a dictation tool on my iphone that’s better than the built-in, but still not a cloud, and lo and behold, there is one! I just tried superwhisper’s local model, and it provided me with flawless transcripts in both english and hungarian, all made locally. yap walks here we come!!

Hey #duckdb friends! I am endlesy fascinated by the tool, and I’m yet to try it, but the one thing that I’m worried about is the lack of autocomplete during my data wrangling in Python. Is there some magic way to get SQL completion inside Python? Some Neovim magic perhaps? Or is this a no-issue?

Yes, vision LLMs are pretty good geo-guessers (I had wondered about this) easily beating humans Bigger LLMs are more accurate and data leakage (having seen the same picture in training) does not seem to be a big problem. The models are better at urban & more developed areas arxiv.org/pdf/2502.11163

I was looking for a free, simple Couch to 5K app to get back into running, and I was disappointed to see that all of them tried to shove a subscription fee down my throat Then I found “Just Run”. Indie developer, free and ad-free, the only IAP is “support Gary who makes the app”. Never paid quicker

I am so tired of rewriting papers 😭😭 #PhDSky

starting the day with “what would done look like for today” is such an underrated move it really gave my todolist a different tone, shoutout @austinkleon.bsky.social and @oliverburkeman.com for the inspiration

This week I gave a presentation on how to run LLMs locally! We touched on models, tools, and technical stuff (like quantisation), I designed this in a way that it gives a good baseline to explore afterwards. Enjoy🫶 #llm #ai

Surely adding AI to printers is somehow against the Geneva Conventions