iiinova — where your digital self keeps evolving

understanding a feature was never a pre-req to shipping it, bigcorp junior engineers usually work via tickets and often have v limited awareness of the product as a whole (hence the frequent low morale in SWE). Not saying this is good, just saying it's already bad.

100

hmm maybe your company is more forward-thinking... In my experience most SWE jobs do not include what you are describing in work time, unless you are at high levels where you are expected to be inventing new things etc. Maybe a 'training allowance' to spend on books you can read in your own time.

100

If you let it do everything, yea you will not learn more than you might glean from reading API docs, neither will help you expand your areas-of-expertise. But you can still be curious, and it's an infinitely-patient teacher.

100

By no means am I an 'expert' now, but I know way more than I did - even though I didn't actually write the code. It just takes some curiosity, which doesn't require pre-existing expertise. And I def learned more than if I had gone and used some existing controller library.

100

This time last week I had never used a PD controller and didn't really even know what it is. Now I know a lot more. I didn't just say "implement a PD controller", I had it explain to me what this was, how it applies to the problem, etc. When it didn't do the right thing, it helped me dig deeper.

100

this is why I think if someone is only using (eg) automated coding in an IDE, vs fully-prompted "coding" in claude code CLI, they maybe are missing out. I don't just do coding in claude code, I also have it explain things to me as we go. Not just code things, but like math/physics/etc.

My iPhone just autocorrected “you” to “noncombatant”…seems weird

131

sick burn. I am ruined. (Are we on Twitter? This feels like Twitter…)

000

not really sure if what I meant was clear - in my experience so far, having expertise makes you much more effective with these automated coding tools. So your hypothetical employer will need employees with expertise, and there aren’t enough of those people out there to just go and hire

100

assuming this applies to AI-generated code (no obvious reason it shouldn't), you can't even put an MIT license on it.

051

if you think this is devaluing expertise, you are missing what is happening. Expertise is a multiplier.

200

do you also get paid if you tab-completed the implementation?

100

the robots ported my ModelGrid C++ library to Typescript today, and made a basic little in-browser editor. Thought this tech was going to be dead (it's trapped in an Unreal plugin) but now maybe not... (clearly some face-normal issues, though...)

060

the idea that 'obviously the code has value' is exactly why every bigcorp has to be so hardline about IP. The 10-thousand-th implementation of 'Lerp' uploaded to github does not have value, and nobody should be paid for it.

200

not if it's boost-licensed (most of my libraries are, and sever other very geometry libraries like GTEngine). For this exact reason - that requiring attribution of every snippet of code makes it impossible to combine parts of libraries into something better.

David Levin @diwlevin.bsky.social · 02/03/2026

Reposted by Ryan Schmidt

The Symposium on Computer Animation is Coming! BARCELONA, SPAIN. JULY 8 - 10, 2026! Papers Abstract Deadline: April 10, 2026 Poster Deadline: June 1, 2026 Website: sca.graphics Call for papers : computeranimation.org/instructions... Barcelona > LA (Come on: en.wikipedia.org/wiki/Sagrada...)

sca.graphics

The 25th annual Symposium on Computer Animation (SCA) will take place in Barcelona, Spain, from July 8 to July 10, 2026.

173

I was 100% in the "I guess it's ok for boilerplate but I can't see using it for anything serious" until I seriously tried to use it to do nontrivial things. Even just in stuff like running tests and experiments, the space it opens up is ridiculous.

000

I did not go and find a Bezier library. It just generated what we needed. Same for the procedural meshing. Took it 10 minutes. Most of the time was spent on the PD controller, where it knows way more than I do about standard stuff, but clearly lacks the ability to reason about the problem.

100

Parts of the overall system are borderline 'novel' now, ie will probably merit some explanation in the eventual paper. It didn't generate the novelty. We iterated, and I told it how to put things together in (slightly) novel ways. But I didn't write a single line of code myself.

100

The "research" is at a higher level. But it can't be done w/o all this infrastructure. Would have taken substantial time - probably weeks - to get all this connected, building, working, etc. Vibed 1st version in about 4 hours, iterated for a few more hours on the plane, it's /very/ good now.

100

sure but IMO most research projects involve maybe 10% novel code and 90% integration/etc. Last week I worked on a project where we are using Polyscope, Bezier curves, procedural meshing, Jolt physics simulator, and a custom PD controller to drive the sim. This took me one day to integrate/"build".

(A policy created by a "Director of Open Source" type of person who did way more harm than good)

000

...only if the team was willing to commit to running it as a major project. So 'research code' for many projects that everyone agreed could be open-sourced w/o any risk, was not, because nobody wanted to commit to being an active maintainer

120

IME it's often a sort of organizational inertia, where it's just simpler to treat everything this way, because it's more work to define precise boundaries. But it can also be super dumb, like early-2010's Autodesk was willing to open-source things but...

120

people say you shouldn't commit too early...Codex is looking pretty good, and then there is Kimi...maybe I should shop around

100

maybe a bit strong...we've only known eachother for a month or so

120

Geometry Coding in the Time of Robots | rms80

wrote a post about developing geometry libraries in this new reality where a robot can generate working code for damn near anything. www.rms80.com/blog/2026-03...

rms80.com

Proprietary geometry libraries are cooked, and Open Source has a new purpose

1254

Reposted by Ryan Schmidt

Kelly Turnbull @coelasquid.bsky.social · 26/02/2026

I refuse to do any amount of fact checking on this, I am taking it at face value

Twitter post:
“Autistic Arab @autisticarab
Duke Nukem fans, I bring grave news about his origin. In Duke Nukem Forever, you can see a photo of Duke when he was in the military. In it, he's not holding an M4. He is holding a C8A1. That's right, Duke isn't American. Its terrifying.
He is Canadian.”
Response
“BeBBeroni Bizza @BebberoniB
Can't even be explained as an american with a C8 either, that's the CAF tac vest in cadpat arid”

18542115

Reposted by Ryan Schmidt

Freya Holmér @freya.bsky.social · 27/02/2026

hearing AI company CEOs excitedly telling you how "we can now all finally become managers" without blinking makes me want to pack my bags and move to the moon

1961268

Ryan Schmidt @rms80.bsky.social · 17/02/2026

added Claude Coding Robots to this infinite-canvas app I've been noodling with, because I really don't love a CLI. Being able to be lazy w/ the prompts and use screenshots as context is pretty nice. Full vid on youtube: www.youtube.com/watch?v=1z-r...

040

Ryan Schmidt @rms80.bsky.social · 16/02/2026

yea I tried looking into this today but the ones that are competitive seem to all need a GPU farm

010

Ryan Schmidt @rms80.bsky.social · 16/02/2026

seems true, I ran into exactly this over the weekend with Claude Code SDK, which uses API credits instead of the (apparently super discounted) max plan, cost was bonkers