Steve Steiner

Latest open artifacts (#15): It’s Qwen's world and we get to live in it, on CAISI's report, & GPT-OSS update After a quiet month, Qwen is back in full force.

Love this wording: “sloptimized models”

From: www.interconnects.ai/p/latest-ope...

Sadly I tend to let it slide until I find myself typing ALL CAPS back at it.

The direction Joe Eldeman takes on identifying user goals is interesting and seems more plausibly pragmatic with llms.

His specific claim is eliciting ‘Attentional Policies’ to create a ‘Values Card’ allows the kind of specificity needed for designing social software.

nxhx.org

You’re absolutely right!

Is obsequious better?

+1 this is an expectation a UI affordance works as it says on the tin - fortunately it also means it has a much better chance of getting fixed.

COLM 2025: 9 cool papers and some thoughts Reflections on the 2025 COLM conference, and a discussion of 9 cool COLM papers on benchmarking and eval, personas, and improving models for better long-context performance and consistency.

𝑵𝒆𝒘 𝒃𝒍𝒐𝒈𝒑𝒐𝒔𝒕! A rundown of some cool papers I got to chat about at #COLM2025 and some scattered thoughts

saxon.me/blog/2025/co...

Annoyed.
“We invented birth control and the population is collapsing” is factually wrong, and poor thinking.

Natural experiments show exogenously increasing contraception access leaves fertility unchanged. Clear do(C) → no-effect evidence. The causal action is upstream in desired fertility.

I hear ‘grow an AI’ being used because training it is not ‘engineering an AI.’

This is a bad metaphor in a different direction.

‘Brew an AI’ vs ‘Bake an AI’ provides a more informative pair of metaphors about what’s happening.

Clearly we should just use the 1s and the 0s. There are many fewer variants.

Assemblage vibing

Claude —

Ha! No, “truthical” is not a word - you’ve caught me in an inconsistency! I was being overly prescriptive about “dynamical” vs “dynamic”:

“Dynamical systems” is indeed the standard term in mathematics

But it’s not because there’s some ironclad rule that “-al” makes things more technical

Sonnet 4.5's suggestion on fixing the error it made splitting one large file into 10 smaller files.

"let me propose a more efficient approach: git reset HEAD"

GitHub - oOo0oOo/lean-lsp-mcp: Lean Theorem Prover MCP Lean Theorem Prover MCP. Contribute to oOo0oOo/lean-lsp-mcp development by creating an account on GitHub.

This is neat:
github.com/oOo0oOo/lean...

YouTube video by Simons Foundation Kevin Buzzard - Where is Mathematics Going? (September 24, 2025)

At some point he needs to say "Smash that subscribe button"
youtu.be/K5w7VS2sxD0?...

VT220 - Wikipedia

Optimal User Experience- en.wikipedia.org/wiki/VT220

Old school programming.

And yeah i heard that all the time when i worked in robots.

The bar is “like a restaurant” not like a washing machine.

I was curious …

These findings establish robust evidence that viable microbes (or at least their DNA / reproductive capacity) can exist in upper-atmospheric aerosols. A domain we might call the “aerobiome.”

It’s an open question if these microbes merely survive or actively live at altitude.

A close-up photo of a horse sticking its tongue out playfully between two metal bars inside a stable. The horse’s head is tilted, its nostrils are flared, and its eye is wide open, giving a humorous, expressive look. Sunlight filters through trees and fencing in the background.

I did have a friend who half-jokingly asked if we could set up a horse related traumatic incident to help encourage his daughter away from any kind of horse fascination.

Close-up of a white ceramic “mushroom” component from an E61 espresso group head, as used in machines such as the Mozzafiato Evo R. The part has a chipped upper lip and a damaged red gasket or seal seated in a groove near the top. Light mineral scale and coffee residue are visible on the surface, with a red sink and drain blurred in the background.

Not good. We’ve swapped to using the AeroPress while waiting for a replacement.

Things going well, then "I'll make 10K tests."

Claude my friend - What happened to the thing we were supposed to be doing?

"Root Cause: Vocabulary Shift
Phase 1 uses proof vocabulary
Phase 2.3 uses testing vocabulary

The WU1.8 task document diverged from it during the Phase 2 detailed planning."

mixed_script_confusables?

Made some Lean 4 proofs before asking Claude Code to build the implementation in rust.

The rust toolchain was not having this Greek symbol in source code malarky -

let θ = test_thresholds();

I’m a fan of Joe Edelman. His Attentional Policies for eliciting a persons values was interesting before llms, and is incredibly more interesting now.
This new paper is a collaboration. The norms model in it also sounds interesting.
www.full-stack-alignment.ai/paper

EP 325 Joe Edelman on Full-Stack AI Alignment - The Jim Rutt Show Jim talks with Joe Edelman about the Meaning Alignment Institute's paper "Full Stack Alignment: Co-Aligning AI and Institutions with Thick Models of Value."

🎙️ w/ @edelwax.bsky.social on “Full Stack Alignment: Co-Aligning AI and Institutions with Thick Models of Value.”

Pluralism as a core principle in social design, thick models of value, multipolar traps, moral graph elicitation, starting with membranes, Moloch-free zones, co-optation risks, & more.

Is a defibrillator closer in latent space to James Bond or Ethan Hunt?

A print Ad for the Coleco Vision super Action controller set. It is a controller with a pistol grip, 4 buttons, a joystick, and a 12 button keypad, and a roller.

When i first saw this as a child i knew this must be how one programmed video games.

I haven’t looked at it recently, but ~18 months ago DSPy stuck me as the way a ML engineer would think programming happens. It literally sets up a ‘forward’.

Kind of the same way as a kid I imagined atari 2600 games were somehow programmed with a very complicated joystick.

I’m guessing Sonnet 4.5, but i just use Claude Code with the cheap-o subscription. I did toss parts at 4.5 explicitly via GH Copilot.
I need to check this morning if Codex stopped thinking about it and started doing something.

Group flow is a thing. A dance company for example.

This definitely is individual flow.