Matt Silverlock

Writing Python in 2025 is good, actually thanks to `uv` 😎

GitHub - elithrar/shootout.py: Run prompts against multiple Large Language Models (LLMs) and capture structured output. Run prompts against multiple Large Language Models (LLMs) and capture structured output. - elithrar/shootout.py

code: github.com/elithrar/sho...

I wanted a way to benchmark a bunch of LLMs from the CLI.

Wrapped @simonwillison.net’s llm library + had it ingest a config file specifying the models you want. It'll run them async (if the plugin supports it) and write the results out to JSON so you can review/eval/process.

Code, instructions to build it locally + deploy to your own @cloudflare.social account here: github.com/jonesphillip...

When one of the PMs on the team *builds a macOS app* that uses Workers AI to auto-transcribe notes from any system process, summarize them, and do it all privately within your own Cloudflare account 😎

Love this app from Notion (faces.notion.com). My previous avatar was from a Notion inspired Figma.

(Ship fun things!)

More reading in 2025, starting with these:

Case in point: I like AI for prototyping, but… are we headed towards a monoculture because that’s all our tools support?

Do other libraries stand a chance? Do we get to try out new ideas?

I’m hoping the next big advance in LLMs enables more real-time knowledge (that isn’t just function calling).

StarCraft (1998)

GitHub - matthewlynch/react-router-cloudflare-d1 Contribute to matthewlynch/react-router-cloudflare-d1 development by creating an account on GitHub.

Finally got around to open sourcing the template I use for React Router + Cloudflare

github.com/matthewlynch...

“podcasts” moving to video platforms and forgetting about audio only listeners: 🫠

DROP TABLE Dogs;

How do you get your clients into everyone else’s editors? 🙃

Right! Using the LSP is a far more accurate approach vs RAG. Tool calling the LSP to validate autocompletions and/or refactors would be huge.

Curious to understand who is pushing in that direction with the glut of AI-enabled editors.

Right, but that doesn’t help the masses: you now have to distribute (and incorporate!) those fine tunes to users too.

The most popular models/platforms don’t allow you to BYO LoRAs, so we’re back at square one.

USB-C is just the physical connector.

It’s entirely legal to ship a product with a USB-C connector that only supports USB 2.0 transfer rates and 15W PD, and even that’s arguable.

FWIW: I think LLM-enabled editors are powerful tools, but the “knowledge latency” of the most popular/capable models is pretty poor.

RAG helps but isn’t a true fix.

How do we experiment with new ideas, frameworks & ship new APIs as the reliance on LLMs increases?

Specifically: the current gap between training data (April 2024 for Claude!) means that there’s far more friction for users to try something new or different.

tried writing this blog 3 times but it’s really boring, so here’s the short version:

React won. Every mainstream f/w now implements its ideas, and developed their own ecosystems. The ones that didn’t, lost and died.

This implies there’s no _strong_ reason to not pick one of the other winners.

This includes scaling our metadata layer, driving our new Rust-based storage subsystem, and designing for our massively distributed network.

Expecting a decade+ of experience for this one.

Email me (silverlock @ cloudflare) if you’re serious.

If you’re an experienced engineering leader with proven capability in building, scaling and maintaining modern, bare metal storage infrastructure: come talk to us here at @cloudflare.social.

to me it’s similar to talking about programming languages, a good CLI, or even a great ratcheting wrench: an appreciation for great tools and QoL improvements that can come with them!

GitHub - elithrar/dotfiles: dotfiles for my macOS & Linux environments ⌨️ dotfiles for my macOS & Linux environments ⌨️. Contribute to elithrar/dotfiles development by creating an account on GitHub.

config & dotfiles here: github.com/elithrar/dot...

Swapped from iTerm 2 -> Ghostty. Super minimal config. Liking it so far: it really does just work + feel native.

… and before the ISO 8601 crowd replies: ijmacd.github.io/rfc3339-iso8...

RFC3339

When the docs engineering team picks the only correct date format:

Being a good writer has always been valuable, but in an era of bland AI slop I suspect that being able to write with an authentic, human voice is effectively becoming a super power

Half crispy (11 mins), half chewy (12.5 mins) 👨‍🍳

We’ve been working on making the new Model Context Protocol from @anthropic.com *really* easy to build against 👷

Lazy self-installing Python scripts with uv I frequently find myself writing my own short command-line scripts in Python that help me with day-to-day tasks. It’s so easy to throw together …

uv is really really really close to replacing about half a dozen tools (and making python the default scripting language)

treyhunner.com/2024/12/lazy...