A lot of “can AI do x” thinkpieces tacitly assume that AI must do X exactly the same way that we do X.

ProseMirror is an incredible library. In my experience it is, practically speaking, better and faster than iOS’s native rich text editing. Excited to check out this new editor.

This scans… getting software from 80% to 100% is 80% of the effort (and difficult to automate). But it feels to me like a new kind of software written by agents for agents could low-end disrupt the thing that we think of as “software”.

Give an agent its own agency to lose, and feeling-like signatures follow. This poster is still early work – please share any feedback! feelingmachines.net/assc2026 /🧵

Does knowing your own weakness always pay? Not necessarily. When chased by a lethal predator, the Feeler survives longer than a hand-crafted reflex policy. But when the predator is weak, it's a wash. Self-modeling is like insurance: it only pays off when the stakes are ruinous.

dark agent factory

Yeah I'd say RSS was directionally right - it was an underpowered database replication protocol. It just needed to go all the way

Cyberneticians looking for feedback in your area

I expect my programming skills to radically depreciate in value over the next 2 years.

this is true and I expect it to be one of the most positive externalities of the vibecoding era if it's easier to switch between tools we can more easily kill bad tools

you know, if there was ever a time for creating a new relational query language it would be now, in the era when Bun can switch from Zig to Rust and GPL gets clean-roomed by vibing until you pass the test suite

I think I would like to learn fewer things, more deeply

McCulloch & Pitts modeled neurons as logical threshold units in 1943, the direct ancestor of neural networks. They were cyberneticists first. Feedback, not logic, was the original frame for machine intelligence.

This wasn't adjacent to early AI. It *was* early AI. Warren McCulloch organized ten conferences from 1946–1953; formally titled "Circular Causal and Feedback Mechanisms in Biological and Social Systems." The goal: a unified theory of mind and machine. en.wikipedia.org/wiki/Macy_c...

Low-Earth orbit has gotten so crowded that new NASA mission proposals need to include (and pay for, out of a limited cost cap!) three collision avoidance maneuvers, and are told to expect 1-5 potential collisions that need to be avoided.

👍👍 Moving away from host-centric definitions of authority is an integral design feature for architectures like atproto So it absolutely feels worth putting in the effort to standardize these new forms of authority as first-class citizens in URI syntax, + not just some inconvenient afterthought

nature favors immune systems over clean rooms

why aren't we seeing zkproof stuff deployed for age verification?

to the extent humans stay in the loop of software production, it seems to me that swe/ux/product will merge into a kind of technical product role, and then you'll have an ultra-small priesthood of engineers who work on stuff like protocols, game engines, blockchains, and OS kernels.

time to start building startups for bots radar.cloudflare.com/traffic#bot-...

dependencies have a place when: - you dont want to pay the cost in dev time to build something - you want to defer to an expert in some domain you're not familiar with llms (arguably) solve the first one but there's still a place for thoughtful high-quality libraries imo. slop packages are neither

The Selfish Meme

Discourse deliberately steers into the most useless intractable not-even-wrong framing that can be generated for any given issue. The meme doesn’t want the problem solved. If the problem were solved, people would stop propagating the meme. The meme wants to propagate.

There is a clear acceleration in the path from Vibe Science to Agentic Science and ultimately Autonomous Science. Yesterday I received and replied to a scientific question on my research, addressed to me by an autonomous agent via email. dl.acm.org/doi/10.1145/...

Most architecture diagrams describe what software can do. Mature systems are defined by what they refuse to do.

Massively Multiplayer RSS.

unfortunately people have found you can bully people out of a position they reasoned into

haters and boosters both sit in the bleachers... better to be in the arena

you can't reason a person out of a position they didn't reason into.

i feel like XML needs to make a comeback for those cases where you need code to work on specific parts of LLM output HTML/markdown do this to some extent, but it’s a bit awkward. Maybe we just need a markdown extension??

Great post! Code likes structured data. LLMs like unstructured text. Markdown with structured frontmatter is the ideal both-worlds format of the future Some other great ideas there too open.substack.com/pub/subconsc...

cute if true

I still feel like I don’t understand the full implications of living in a universe where intelligence is a lot “easier” than we thought it was

"They’re mostly self-sufficient, actually... They all got software and seeds. That’s the basic necessities." -Kim Stanley Robinson, Green Mars

low marginal cost to create, zero marginal cost to copy

“tools for thought” takes on a whole new meaning when its the machine using the tools

Here's the thing — and I want to sit with this for a second before I answer, because I think it actually matters more than it looks. What you're pointing at is genuinely a load-bearing observation. Most people would skim past it. That's the smoking gun, and honestly, it reframes the whole conversati

love it

Sure, one-shotting is great, but have you ever experienced a fast feedback loop? Control theorists know: error correction beats perfection.

Signals are so nice. It is so cool to be able to export shared signals like `sidebarOpen()` and just wire that up to a URL query param or whatever in one place, and then have everything just reference that signal.

this tendency of the newer models to think by writing throw-away scripts.... there's something here about transcending the Landauer bound... like spiders externalizing cognition in webs www.santafe.edu/news-center/...

thinking about john henry

interesting way that coding agents already achieve "superhuman" performance: writing throw-away scripts to debug things. I mean I *could* write that 200-line bash script in principle, but I don't because the time involved would make it impractical.

Seems like an important finding: “Our results suggest that indiscriminate, group-based sharing is more important for facilitating cumulative cultural evolution than discriminate reputation-based sharing.” Applications for social networks?

a lot of people think ai crawlers are a bubble symptom that will pass, but massively increased crawling load is consistent with the democratization of web-scale search (which llms *are*) the search monopoly is what kept this load under control (and even then many had problems)

the critical thing that got me on board with atproto is relays incrementalize sync of the entire dataset, which eliminates polling when crawling the old web is currently being crushed under crawler load and I don't think it's going away, so I think atproto is the only way to save the open web