😆 'Max'? I'm not sure that word means what you think it means . . .

I'm about to hit the daily limit in less than a half day of usage just _playing_ with it. 🤷‍♂️

That’s a fun example for MCP! Maybe let people text it (?) 😀

note: was pointed at www.swebench.com

Whether that can be applied in anyone else's configuration... well, I'm looking for a reliable source of information that is not just one or two people's story, and coming up empty.

Everyone seems to have a story of 'I found x works better', which was probably true for them given a particular configuration (model, parameters, system prompt, targeted codebase language/size/shape, context size, assistant tool).

The amount of superstition and anecdotal evidence around how to craft effective context and prompts that improve LLM results when developing software is staggering. 🧵

I have been very surprised how fully featured this Obsidian spreadsheet plugin ( Univer") is. Now I have small local spreadsheets organized with notes and diagrams (Excalidraw plugin), and I really have not run in to any layout/calc I needed that it could not handle.

I have been very surprised how fully featured this Obsidian spreadsheet plugin (
Univer") is.

Now I have small local spreadsheets organized with notes and diagrams (Excalidraw plugin), and I really have not run in to any layout/calc I needed that it could not handle.

Playing around with [building out] this keyboard streak tracking idea some more. Added a way to reset the data, and with only a couple of data points on the chart, my improvement trend line today is impressive. 🤣

Playing around with [building out] this keyboard streak tracking idea some more. Added a way to reset the data, and with only a couple of data points on the chart, my improvement trend line today is impressive. 🤣

After trying out several of the current generation of LLM models and code assistant tooling, I'm looking forward to seeing more techniques for stuffing & increasing the context window, generating embeddings, and deriving codebase roadmaps that better help LLM's to work with larger codebases.

Is Google Search Cooked? + We’re Getting a U.S. Crypto Reserve? + What You’re Vibecoding — Hard Fork “I think Google realizes that this is a once-in-a-generation chance to reinvent the search experience”

re:vibe coding, I thought it was really interesting to hear more examples of what people who are not developers for a living are making with it. i.e. - the last 20 minutes of this episode of the Hard Fork podcast has some fun ones.

Happy Friday! Recently, I was toying around with making a small open source macOS menu bar experiment/app to track keyboard usage 'streaks' (no mouse). Here's what it looks like so far:

Happy Friday!

Recently, I was toying around with making a small open source macOS menu bar experiment/app to track keyboard usage 'streaks' (no mouse). Here's what it looks like so far:

The Cline memory bank pattern is INSANE and not talked about enough. It is enabling me to steadily and safely address tech debt in a code base. As the code base grew, I was getting pretty chaotic results without this 'map' that helps guide the LLM!

The Cline memory bank pattern is INSANE and not talked about enough.

It is enabling me to steadily and safely address tech debt in a code base.

As the code base grew, I was getting pretty chaotic results without this 'map' that helps guide the LLM!

I ‘vibed’ out (with cursor+claude 3.7 sonnet agentic) a next.js backend that uses firestore for persistence, with an API layer, web app, and playwright tests.

It is an unfamiliar stack for me, but the whole process felt more like parenting than coding. Was weird.

Lots of churn and waiting on LLM.

Using Perplexity to sort through campgrounds based on Bortle (dark-sky) scale and driving distance worked alright and saved me _some_ time. 🤷‍♂️

I have been trying Alex Sidebar, Cursor, and Cline lately.

The latest version of Cline is really impressive _when it works_. It can generate tools and execute shell commands in the existing VSCode console, which means you can have it run a build or tests and it can use the results for guidance.

CCMenu (originally named for CruiseControl) can apparently monitor Github workflow status! cool!

I wasn't really looking for a new Terminal app (Ghostty.org), but the satisfaction/nostalgia of an always available 'quick' terminal sliding down from top of screen at a keypress is real.

Some fun background on where Ghostty came from is here: zig.show/episodes/32/

Surprising things I learned: The system prompt additions that power the Cline 'memory bank' feature use a Mermaid spec'd flowchart rather than prose to reduce ambiguity. Cool!! image courtesy of https://cline.bot/blog/memory-bank-how-to-make-cline-an-ai-agent-that-never-forgets

Surprising things I learned: The system prompt additions that power the Cline 'memory bank' feature use a Mermaid spec'd flowchart rather than prose to reduce ambiguity. Cool!!

image courtesy of https://cline.bot/blog/memory-bank-how-to-make-cline-an-ai-agent-that-never-forgets

Horror story in 0 words

#459 – DeepSeek, China, OpenAI, NVIDIA, xAI, TSMC, Stargate, and AI Megaclusters — Lex Fridman Podcast Dylan Patel is the founder of SemiAnalysis, a research & analysis company specializing in semiconductors, GPUs, CPUs, and AI hardware. Nathan Lambert is a research scientist at the Allen Institute for...

This fairly dense convo between a couple of AI practitioners and Lex Fridman, was a fun listen and maybe the only 5 hour (!) podcast episode that has ever been able to hold my attention.

I wasn’t able to do it in one shot, but I did come right back to it ASAP.

overcast.fm/+AAeZyCRnfH8

Does it recording screen recording permission?

That is one of my current nits with the bartender app that does something similar.

Updated my self-hosted @ghost.org blog to latest version.

It's interesting to see how some folks syndicate or make content platforms like Medium or Substack their primary home.

I'm not really looking to monetize, just share, and especially share rough things -- as a way to think through ideas.

Like any metric, it can be gamed... but this particular one is _even more opaque_ than most, since contributions to a private repo are shown in that graph and you can't drill into the detail.

I mean.... this one looks real, right?

When I hear that a company is using a Github contribution graph as some indicator of how 'cracked' (LOL that term!) a developer is, I have to exercise mucho willpower to contain my chuckle. 🧵

Note: icon is from their media kit which can found here: drive.google.com/drive/folder...

Added a new @bsky.app social link to my blog at mattorb.com 🙂

GitHub - mattorb/MakeMeStand: macOS status menu app to automatically raise and lower IDÅSEN standing desk using BLE macOS status menu app to automatically raise and lower IDÅSEN standing desk using BLE - mattorb/MakeMeStand

I wrote an open source _automatic_ Standing Desk macOS menubar app for Ikea Idasen (BLE control to Linak DPG) -- Fun way to experiment with what level of testing can be achieved around BLE code on Mac, and I use it every day too 😄

github.com/mattorb/Make...

Aligning Software Teams This morning, I have been thinking about past software projects in terms of what made them successful balancing risk, learning, achieving results and good vibes. Making sure a team has a shared under...

Aligning Software Teams:
mattorb.com/aligning-sof...

Clarity in Decision Making, from Org to Architecture This morning I was thinking about how the most successful software teams I have worked on had a _very_ shared understanding of how decisions were guided.

Clarity in Decision Making from Org to Architecture:
mattorb.com/clarity-deci...

Model Context Protocol and Local AI What is Model Context Protocol? It could be a game changer for local LLM usage w/tools. I poked at it some this morning to gain some understanding and sketched out these notes along the way. And h...

Model Context Protocol in relation to entirely local LLM usage: mattorb.com/model-contex...