Adaptive Intent-based CLI State Machines by Shawn Wang YouTube video by Heroku

This talk on the Netlify CLI is good
www.youtube.com/watch?v=Zueo...

I think it's time to seriously consider moving to decentralized image sharing

An example: Pixelfed

pixelfed.org

Graze.social - Curate Your Bluesky Feeds Design, deploy, and grow feeds of any complexity on Bluesky with Graze.social.

Have you seen www.graze.social

It might be one way to dip your toe and then proceed.

I’ve done some tinkering on Bluesky. I’ve been told the firehose can effectively be consumed on a Raspberry Pi, but I haven’t setup one of my own Raspberry Pi devices to test this theory.

a graph of US data center construction over time, which skyrockets after the launch of ChatGPT

Official data released today shows US data center construction hitting a new all-time high of $31.5B annualized, up 43% from this time last year and 128% since the release of ChatGPT two years ago

4/
- Garbage in, garbage out. Wire it up to interesting data and integrations.
- If a compound system has LLM-level intelligence in the middle, something akin to AI evals are often required to test the overall compound system performance too.

3/
- Much of what we reason about is recent, but LLMs have knowledge cutoffs. Pack your context with all kinds of recent knowledge.
- As software components with downstream effects, either insert LLMs where error is contained (ie. retry priority), or create suggestions for a human to accept.

2/
- API costs can add up fast. The simpler models and local models are very capable. Don't pay tolls if you don't have to.
- Orchestration can be done with things like functions and coroutines. Orchestration frameworks have a lot of downsides.

1/

Some of my AI learnings:

- Vectors can be stored in a lot of places. Vector databases have a lot of downsides.
- Many embeddings can be instruction-tuned. You can fine tune via a special textual prompt on top of the user prompt.

2/ It's simple to achieve setups where a pair of headphones plays music and another pair can work only with meetings and videos. Applications like soundboards, effects layering, and voice chat setups are now straightforward to create.

1/ Finally got around to tinkering with Linux's new sound architecture PipeWire. There's a gui called Helvum where you can patch audio sources (mics, specific applications) into arbitrary outputs (specific speakers, headphones, specific applications).

One of the best software books is about this topic. Check out Stanford Prof. John Ousterhout's "A Philosophy of Software Design". He has this idea of deep modules, and uses the OS filesystem APIs as an example of a tiny interface with a powerful engine. I recommend it.

A Century of Hypermedia

https://buff.ly/4eF0RPU

"Belgian Paul Otlet described streaming audio/video in the 1930s"

#api360 #hypermedia #apiDesign

Pepperidge Farms remembers when whole swaths of the internet staged blackout protests over this.

I'm curious about the idea of slack in the system. Can you quantify or add more detail about how hard people are pushing in high performance scenarios?

New Bluesky tool alert:

If you follow @listifications.app, they will DM you noting when you've been added to a list, starter pack, block list, or someone's blocked you and which user did so.

All those years of CSS specificity puzzles did not prepare me for Ansible's 22 levels of variable precedence.

Reading Bertin feels like taking your thinking from 8-bit to 4K over the course of a few weeks.

Just switched my blog from Ghost to Astro on Vercel and it was a half-day exercise. They've really done an amazing job with it.

Overwhelmed with all the starter packs? Following thousands of people you know nothing about? There's a better way.

Convert starter packs to lists and try them first. Follow the instructions here (not my app):
nws-bot.us/bskyStarterP...

You can pin lists to your home just like you can pin feeds.

I bought them as pomodoro timers, but I just really like the visual cues and fidgetability for a tad more timeboxing and nag in my life. Well that was the 3 minutes I gave myself for this skeet--nice chatting!

Exactly. And I only see this getting better as more people make high quality custom feeds.

My new favorite life hack are these timer cubes. Flipping a cube is much faster than fumbling in an app. Use them for cooking, meetings, reminders of upcoming things. Super handy.

Finally got my allergy drops for Austin, TX! Unfortunately not in time for allergy season. (Please imagine this as the penguin meme)

Pillars of creation by JWST but with googly eyes

New #JWST image just dropped 🔭🐡🧪

The Shift from Models to Compound AI Systems The BAIR Blog

This was one of my favorites.
bair.berkeley.edu/blog/2024/02...

Do you have stats on how well the generated cover letters perform compared to handwritten?

I see them do important but thankless tasks, like upgrading dependencies and making tests faster.

I have these AI apps so far: ChatGPT, Voicenotes, Merlin (bird lookup), PlantNet (plant lookup)

This echoes my ambivalence of content-centric and creator-centric social networks. Unrelated--I've wanted a service where you could contingently escrow-fund a wallet for individuals and projects of your/crowd choosing and it would let them know. There are so many models we haven't tried.