I’m almost 50 now, and if I look around and actually remember what things used to be like, everything is nicer, better, fancier than before. Including social services, schools… everything. But it does require really thinking and remembering, it’s surprisingly unobvious.

Will Stancil@whstancil.bsky.social · 9h ago

I think we're learning: -consumerism is an overpowering populist force in American society -the populism expresses itself as a base urge for greater consumption; it is not reactive to conditions on the ground -people are ashamed of this and often try to disguise it as other types of politics

There’s a few LLM evals for solving text adventures. A lot of results seem surprisingly poor. But I bet if you phrased it more as explicitly asking the LLM to solve a text adventure eval, and explain its reasoning and strategy along the way for the benefit of the evaluator, that it could improve

I imagine with agentic customer support bots we’ll quickly see the bot try to hack their own company. Not necessarily computer hacking, but process hacking; something we occasionally see human customer support do when they learn they learn the back doors in their own processes

I reached the end of the Fable usage included in my subscription... I spent most of it sharpening tools, with only a couple features I gave it that felt big enough to really need it. Looking back on the stats...

I've been using Fable to do some large refactors that I've been putting off, and took this advice. It looks like Fable is using Sonnet 5 for most of the work, successfully. I'm still running through my credits, but it's been grinding for like 8 hours.

Simon Willison@simonwillison.net · last mo.

The most interesting Fable tip I've heard so far is to let the model use its own judgement as much as possible I told it "For all coding tasks use your judgement to decide an appropriate lower power model and run that in a subagent" and it seems to be saving tokens simonwillison.net/2026/Jul/3/j...

I made a pocket calculator LLM-powered. As commentary.

I really hate the phrasing that an LLM predicts the probability that a token will be next in a sequence. It assembles scores and uses scores to pick the next token. It's not a probability. ML snuck in some fake sciency-sounding stuff by calling scores "probabilities" and it's bad for discourse.

Ethanol production seems to have a roughly similar environmental impact as AI (though vastly more water). All to do something completely pointless, that doesn’t even employ many people, nor provide any discernible benefit to society

Will Stancil@whstancil.bsky.social · 2mo ago

Tons of furious responses to this, but many things that are objectively worse than AI or data centers, with as many or more visible harms, are not especially unpopular. The rapid backlash to this particular thing is a bit mysterious even if you agree the backlash is well-deserved.

So AI is polling badly but widely used. It seems like this will end up like privacy: everyone will be bothered by it, and it have no impact electorally. This doesn't mean lobbying politicians directly for policies can't work, but like privacy it will feel like directionless chatter in aggregate

It's conventional that, given an animal that is fatally injured, we put it out of its misery. I just helped a neighbor with a bunny that her dogs had gotten. But I actually have no moral intuition on whether it is better than letting it die naturally. Or maybe there's no moral difference. No idea.

It takes me a while to notice when people stop talking about things, but I think everyone stopped talking about fine tuning a while ago…?

Listening to a talk on managing AI contributions to open source, an AI maximalist approach occurred to me: only accept bug reports, implementation guides, RFCs, that lead to a roughly canonical one-shot AI-authored implementation. The implementation is then kicked off by a maintainer

I think part of the reaction against birthrate discussions is when it's phrased primarily in terms of economics. Who will take care of the olds when there's not enough youngs?! That discussion feels like a societal entitlement to both childbearing families and to the children they produce, but...

Ryan Moulton@moultano.bsky.social · 3mo ago

Bluesky would like to enforce a norm that caring about birthrates is weird, cringe, and perhaps even right wing. Any violation of that norm is met with shock. That shock is entirely due to bluesky specific brainworms.

I think the biggest issue I have with CSS is how hard it is to statically understand CSS. Lots of overlapping pieces that can only be understood in a fully realized DOM. Tailwind successfully punts on this. The term "strongly typed CSS" comes to mind, but I have no idea what that would mean

Julia Evans@b0rk.jvns.ca · 3mo ago

moving away from tailwind, and learning to structure my CSS jvns.ca/blog/2026/05...

Inside our nuclear family we have been working at not taking offense, assuming positive intent, and not keeping score or inferring others are keeping score. It’s hard, we backslide, but it is very worth it! Not taking offense is a real virtue. It leaves space open

William B. Fuckley@opinionhaver.bsky.social · 3mo ago

Unfortunately as the the bad faith accusations against stancil show, the internet has settled into a bad equilibrium where anyone with a large platform or presence basically has to adopt a stance of “no it isn’t, fuck off and go away” towards accusations of bigotry from people they don’t know.

"The Ancients could do wondrous things, but often made mistakes" is a common trope in dungeon-crawling fantasy, but said fuckups are usually framed as products of hubris or madness. I want to see a setting where they messed up for the same reasons real public works often come to horrific ends. (1/3)

I've found it very effective to give coding agents lots of bureaucracy to fill out: fields to fill in, lists to follow, overzealous checkers, etc. You know who else does this? The government! How much is this like a procurement process?

My wife had setup a travel guide and wanted to make it printable. It kept leaving out the images from a PDF, and so I sternly admonished it and told it to do better. Apparently ChatGPT simply cannot include images from the web in a PDF. So instead it drew its own pictures...

"Didn’t anyone see my skeet about the epidemic of poorly thought-out return-to-office policies? The carnage that’s currently being carried out by formerly inanimate objects is the next logical step beyond all the stuff I warned everyone about." buff.ly/5tPeXxJ

Why Are You So Surprised That All the Office Chairs on Earth Have Simultaneously Achieved Sentience and Started Killing Everyone?

“Didn’t anyone see my skeet about the epidemic of poorly thought-out return-to-office policies? Or the one about the long-term toll that bad ergonomics can t...

mcsweeneys.net

this is the way. terms: the gift. open source

Seeing some “levels of AI engineering” scales, and they don’t quite fit the progression I’ve felt. One tendency is to max out concurrency, with hoards of agents or lots of sessions, but that feels a bit like a phase people go through, not an end in itself

A very common AI sci-fi trope is for it to make precise predictions like "this is 86.3% likely to succeed" I feel like we are further from this future than we were five years ago. Maybe it was always an unlikely future? Maybe it is an implausible view of probabilities?

Really the only way I can make sense of AI personalities or selfhood is as fictional characters. An LLM may author much of the character, but they are still two distinct things. We never think of fictional characters as independently conscious. Yet without authoring the LLM can’t maintain identity

Melanie Walsh@mellymeldubs.bsky.social · 5mo ago

In fairness, people were really confused about fictional characters in early novels! It does seem like there’s a parallel there.

I like how in 2026 a common security paradigm is writing a strongly worded letter to the guy in your computer

I've been rethinking my testing strategy for agentically coded projects (i.e., vibecoded; where I'm not touching the code directly). It's easy to just have it go and make tests, and sometimes poke it or add instructions to make more tests. I haven't seen much value from the result...

Solar power is definitely a success to be proud of… but it also points to an inability on the left to be happy or satisfied with anything. In this case not even due to any critique (solar power is pretty great all around), but an inability to let the gaze linger on success

mtsw@mtsw.bsky.social · 6mo ago

The left tried for decades to pass green energy subsidies, passed them, the subsidies worked, and now literally the entire planet's electricity grid is going to convert to solar+battery. It just doesn't feel like "tech" because it mostly produces kinda boring blue collar jobs instead of billionaires

Putting all others things aside, the AI of today is SO PERFECT for science fiction stories. These machines are so weird, and there’s so many opportunities to project this new familiar into interesting spaces

norvid_studies@norvid-studies.bsky.social · 6mo ago

kind of a fascinating last mile failure mode, from the POV of like discussions of AI in 1990. almost an asimov story

Very timely and also very not timely

Tom the Dancing Bug, by Ruben Bolling@rubenbolling.bsky.social · 6mo ago

🔴 Brand new comic now up on @boingboing.net - Super-Fun-Pak Comix, IN WHICH Mickey Mouse is reunited with his dog in the public domain, but it doesn't go the way either expected... - boingboing.net/2026/02/11/t...

I have felt this, but also have found I can change my approach as this happens. I switch to long-winded meandering voice input, and step back to giving my motivations instead of direct instruction, and it often helps me get over the hump.

Simon Willison@simon.fedi.simonwillison.net.ap.brid.gy · 6mo ago

Interesting research in HBR today about how the productivity boost you can get from AI tools can lead to burnout or general metal exhaustion, something I've noticed in my own work https://simonwillison.net/2026/Feb/9/ai-intensifies-work/

I’ve been experimenting with some AI-written news briefs, and by default AI creates newsy titles: kind of click bait, teases at the subject, hides the answer in the article. I almost skimmed over it because it’s just what we expect in news. But now I’m the editor, and I don’t have to accept that!…