Matthew Siu

interface researcher; currently exploring tools to support human creativity and understanding

[Matthew Siu @MatthewWSiu]

about a year ago, i stopped sharing as much on this platform. it had changed too much from what it had used to be, and many of my mutuals who i used to interact with here had left. or their tweets had stopped showing up on my feed.

but i still wanted a public outlet to share parts of my life online. and so inspired in part by similar experiments by @upcycledwords and @cowjuh, i made a little telegram bot that i could text that would publish the text messages to a feed on my website.

i called it the bulletin. and over the last year, it’s become a record of the thoughts and feelings happening in my life. it’s been nice having a quieter place on the internet to share things to. it feels like a space that i’ve carved out for myself to explore the way i want to show up in the world. i re-read the whole thing recently start to finish, and parts of it felt cringe, but i think that’s just apart of figuring things out.

with the recent changes to twitter, i think i will try sharing more things here again. and so, to re-introduce myself to those who have been following me here, hello! here is a slice of what has been happening in my life over the last year👇

https://t.co/lXih3HyK48

[Matthew Siu @MatthewWSiu]

@enjalot any tips for performance?

[Nick @nickcammarata]

interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had

with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works

https://t.co/zHij5zwKGT

[Matthew Siu @MatthewWSiu]

got my copy from @JimmyRis and looks like he still has copies available for those interested!
https://t.co/iW23uALydo

[james, the giant peach @JimmyRis]

@majamediaco @MatthewWSiu @spencerc99 Glad you found a copy!

For others curious… reprints and originals are still available from the Ted Nelson Archive store at https://t.co/DAj67lVFQ4

Along with other treasures and doodads periodically that we uncover from the archives :)

[Matthew Siu @MatthewWSiu]

wasn't doing much good in the plastic wrap 🤷‍♂️

excited to see more weird, interesting and optimistic stories about computing

[maja 🔭🍒 @majamediaco]

Just opened an original 1974 Ted Nelson ‘Computer Lib / Dream Machines’ from first printing. Special !

[Matthew Siu @MatthewWSiu]

@___________chen still do!

[Matthew Siu @MatthewWSiu]

started doing similar range of motion exercises over the past year after some shoulder injuries sustained doing calisthenics

find them useful for warming up and recovering from minor tears + the body feels really good afterwards

[Justin Mares @jwmares]

Lately I've been diving into what I think is the single most underrated fix for back pain.

It's called Functional Range Conditioning (FRC).

It's essentially strength training for your joints and it basically fixed my back pain (and my brother's) in a few weeks.

Last year, my younger brother tweaked his back doing jiu jitsu. A few days later, he woke up unable to move. He was looking at a fairly long recovery time, PT, the whole bit. I sent him to literally one session with my FRC guy Brian, and after an hour (and a week of consistently doing Brian's stretches), my brother was fine.

I had a similar experience with some light low back pain from too much computer use. After a few weeks of consistent FRC practice, it was gone.

I suspect a lot of people are walking around with some version of this.

e.g. a hip that pinches during a squat or lower back that gets cranky after a flight.

It's not agonizing enough to fix, so we accept it as the "normal" part of being in our 30s and 40s.

The core insight behind FRC is that every joint has two ranges:

1. The passive range - how far a joint goes when something else pushes it.

2. The active range - how far you can move it under your own power.

Try it right now.

Pull your knee to your chest, then let go and try to keep it there. It will drop some amount. That gap between where you can put it and where you can hold it is where most joint issues are.

This gap is driven by your nervous system. Your brain won't let you go somewhere it can't control.

If you never train near your end range of motion, the brain goes "okay, well I guess we don't need this range" and it shrinks.

Stretching won't fundamentally change how your brain thinks you can move. That's what FRC is all about.

(A 2025 review suggested static stretching doesn't actually lengthen your muscles. It mostly just makes them less stiff and increases your tolerance for the stretch, which is a bit depressing given how many hours I've spent on my hamstrings.)

If you want to try it, do 10 minutes of CARs (Controlled Articular Rotations) a day for your hips, shoulders, and spine.

(I left two YT videos in the tweet below for you to check out)

A CAR is moving a joint slowly through the biggest circle it can make (under your own control) with the rest of your body still.

e.g.

- Hips: stand on one leg, lift the other knee, sweep it out to the side and around behind you, then reverse.

- Shoulders: arm straight, trace the largest circle you can from front to overhead to behind you.

- Spine: slow cat-cow stretch

Basically, go slow enough that it's boring, but hard enough that it isn't.

My back pain is mostly gone, for what it's worth. Hope it can help some of you out there.

[Matthew Siu @MatthewWSiu]

@kelin_online gotta write it down 100 times

[Matthew Siu @MatthewWSiu]

@kelin_online @nealagarwal @rsnous @cwervo @FolkComputer @xhfloz @josecastillo yo @xhfloz looking fire

[Matthew Siu @MatthewWSiu]

@vickiheart @loom @zachwaugh @rickmesser aw hi Charlie! 👋

[Matthew Siu @MatthewWSiu]

@maxbittker i like how benchmark companies like @ValsAI are starting to draw these lines that show at a glance the best performing models at certain cost/score tradoffs

thinking this could be useful for runescape bench too

[Matthew Siu @MatthewWSiu]

@DavidSHolz can you share stories?

[Max @maxkrieger]

I built a living town with Opus.
It references "A Pattern Language" for the design. You can toggle architectural details to get a feel for them :)

[Matthew Siu @MatthewWSiu]

@left_pad been meaning to play firewatch and outer wilds

[Matthew Siu @MatthewWSiu]

@maxbittker become a great artist in 10 seconds 👀

[Matthew Siu @MatthewWSiu]

@vincent_yc_li let's goooo

[Matthew Siu @MatthewWSiu]

@theombl what's the story behind the gas station?

[Matthew Siu @MatthewWSiu]

@aidenybai UW！

[Matthew Siu @MatthewWSiu]

@donaldjewkes prefer this one tbh

[Matthew Siu @MatthewWSiu]

This is still a WIP, so any feedback and ideas are welcome

I’m also interested in working with safety/alignment researchers to adapt this to real research workflows and see where it’s actually useful

You can try it here: https://t.co/7kuZOdVmOs

[Matthew Siu @MatthewWSiu]

Search the transcript by keyword or prompt

Then squish the transcript around the results so you can skim them in context

[Matthew Siu @MatthewWSiu]

Built a tool for investigating Anthropic's released Mythos transcript

It gives you an overview of behaviors flagged across the conversation, then lets you jump to key passages in context

https://t.co/7kuZOdVmOs

[Matthew Siu @MatthewWSiu]

The design here draws on some past work done in collaboration with @andy_matuschak

Specifically, when trying to make sense of a pile of long, unstructured documents, it can be helpful to gather the important snippets into one place so that you can see them together

This makes it easier to notice patterns, form hypotheses, and figure out what to do next
https://t.co/i57u2g2cfn

[Matthew Siu @MatthewWSiu]

📣 Excited to share Latticework, a text-editing environment aimed to help synthesize freeform, unstructured documents ✍️

Made in collaboration w/ @andy_matuschak