This is great - it's about time someone updated the discourse on LLM energy usage to reflect that coding agents use massively more prompts than occasional questions to ChatGPT Simon estimates that a day of coding agent usage comes out close to the energy needed to run a dishwasher

For those of you who have followed my brutal 12-year saga of trying to finish The Power Broker (on the other site that must not be named) - I’m on page 899, just 263 pages to go. I’m (for real this time) finishing it in 2026. Less scrolling, more reading (as I type this on a scroll machine)! 💪💪💪

There is a class of AI critics on this website, typically academics, that speak about the tools like it is early 2024. I'm not asking you to start liking AI. But you sound ridiculous because you're describing things that no one experiences anymore. Get updated.

45 likes for a link that 404's. What is happening to us? 😭😭😭

Chain-of-thought around tool-calls leaking into an AI chat response. This actually happens infrequently but regularly in my usage. I wonder if this is a bug or actually something more intrinsic that you might not be able to completely stop an LLM from doing.

Lots of speculation that Google could be displaced by AI chat. Slowly becoming true, but looks like it's actually a lot worse for all the alternatives. 12 week comp. (Personally I've set Perplexity to be my default search engine and don't find Google useful anymore except for navigational queries)

In case people don't realize how domain handles work – though there's a "blue check" next to the name – it's actually not needed. The domain returns to Bluesky a record confirming that the account controls the mayor.nyc.gov domain. If you trust mayor.nyc.gov, then you can trust that this is Zohran 🎉

Bluesky community! This is my official NYC mayor account. I’ll be posting updates on what’s happening at City Hall and how we’re delivering for New Yorkers. It’s great to be here.

Started writing about some major stats issues with an AV story that was published yesterday and found the nuance too difficult to get across in 300 characters. I don't want to encourage quote posting of one point but not the other. Serious proposal: Bluesky but 600 character limit??

This is an interesting example to me of an expert human making a basic arithmetic mistake that the current state of AI does not make. A year later, hallucinations seem way down. The number is a fatality every 79.4M not 123M miles. Perplexity with Sonar or Gemini both get it right. ChatGPT does too.

Small quirk I noticed in inference costs between a smaller model and a larger one. Qwen3 8B is half the price of Qwen3 4B even though theoretically 8B is a lot more computationally intensive. Probably a demand issue with 4B. GPUs need to stay fed, else it is burning money while doing nothing.

I was thinking this morning about how expressions of disdain, disgust, and disagreement are so heavily prioritized and propagated over here. At a minimum, it is simply a psychologically unhealthy information diet.

I'm burned out on sentences that start like: "In full transparency ..." "We need an LoE for ..." "Let's circle back to ..." It's human generated meeting slop.

This is an issue with my family consuming a ton of AI slop www.theatlantic.com/technology/2...

Oldie, goodie:

Is Bluesky blocked in Venezuela? I assume it is alongside X, TikTok, etc. Even if it's not, there is probably self-censorship for fear of imprisonment. One of the glitches in the matrix yesterday in an over-connected world is how social media posts from Venezuelans in Venezuela are very rare.

First-world problems. After all these years, Etsy's personalized recommendations are still trash. You would think years of purchases would change things, but nope.

Been wondering what computer OS I should let my kid use that would both be intuitive and safe for them, but also expose them to computing fundamentals like the file system and basic programming. Chromebooks are garbage, but surprised I am coming to the conclusion probably Linux not Windows or MacOS.

One theme for 2026 is I have been coming to grips with the reality that there has been severe enshittification of public schools through Chromebooks and digitization. Digitization has led to unfettered, dictatorial, and unadaptive control. And *then* there's digital addiction and screentime issues.

Part of the reason why libraries struggle is how consumer-brained we've all become. If you want a book, the default is to buy it, instead of checking if you can borrow it gratis, first. Borrow rates & foottraffic stats crater, austerity hawks jump to defund, repeat.

There have to be other people suffering from this. In case it's helpful: Firefox can end up with severe network slowness with QUIC enabled and a VPN/Wireguard layer. In my case it might be from running Tailscale. Disabling http3 makes it nice and fast again.

NotebookLM is 80% "this is incredible" and 20% "why does it keep doing this bizarre error I tried multiple times to stop now."

Need a story about how enshitification has hit the app experience too.

Corruption has a cost. It pulls the government’s attention away from those who need it most.   That’s why we must fight against it in all forms, at all levels.

I had an argument with my son yesterday and he asked my if I had gotten a piece of info from AI without reading the search results or going to the source to determine if the info was accurate. I had. This video perfectly illustrates that process.

Not gonna lie, it’s pretty annoying that every time I download Threads I see a half dozen new interesting meticulously-executed features that are distinct from what any similar app is doing

Slop (no link) on slop (screenshot of random twitter user) on slop (another post with a take and no link)

"The newspaper industry served as a feeder for our entire news media ecosystem. So when we lose those newspapers, we've lost that journalism. In the early 2000s, we had 40 journalists per 100,000 people. We're now down to 8 journalists per 100,000 people," says @victorpickard.bsky.social.

If you're perplexed how people voted against their interests this year, consider supporting public broadcasting.

🎁 ♥️ The Gift of Room Tone 🎁❤️ Every holiday season, we like to put together a montage of footage from the end of our on-camera interviews with filmmakers, critics, and scholars, highlighting the moments when we capture "room tone"-the ambient sound that our editors need to create a seamless video.

The only true way is the Canadian way (As much as I wish it was a joke, it is in fact not one and this is exactly how it works here)

It's funny IMO that Berkeley, California was one of very first jurisdictions to ban straws/plastic bags -- you know, for the environment -- But soon as we learned the actual leading source of plastic pollution? Crickets. Performative progressivism. www-pbs-org.cdn.ampproject.org/c/s/www.pbs....

The problem with Bluesky seems to me that people just don't subscribe to new people here very often. They make a list when they join and then rarely change it. Compared to my real platform on Reddit, subscriber numbers rarely ever move here. I just don't see how this is sustainable for a platform

Remote workers don't need to be in the same physical space, but the tensors replacing them do.

I want a candidate the centers our kids and our future. No guns. No data centers. No unregulated social media. No more ads for kids. Save the climate and our schools. Make streets safe so kids can play outside and walk. Make our taxes work for us. What is so hard about this???

Unpopular opinion but it’s bad young people don’t talk on the phone anymore. More loneliness and isolation. My boomer parents are constantly on the phone with their siblings and friends

Coastal cities can’t get their acts together on housing. Sun belt cities can’t get their acts together on transportation. The Midwest can’t get its act together on its economy.

Actor Peter Greene has passed. Many years ago, I had been watching The Black Donnellys and was sitting across from him on the subway in NY. Couldn't think of his name. He saw me struggling and asked me for the time to help me out. Rest in peace.

Really hard to follow sensationalist discourse about Nvidia and AI. It is simultaneously: all just a bubble, last week the company has apparently lost to Google, today it must be restricted from selling outdated chips to China, yet China is allowed to train in SE Asia on the latest chips anyways.

Perplexity's default Sonar model is fast but inaccurate and not good. I get the user wanting a quick answer in a lot of cases, but it seems really penny-wise, pound-foolish behavior. So much better to get a high-quality answer with a bit more time.

I have a subscription to Perplexity which lets me switch AI models constantly and compare results. Gemini seems good, but I think Gemini 3 has some trouble with instruction following vs GPT 5.1. Perplexity's RAG-first approach also a big difference maker for answering conventional questions quickly.

log in with your internet handle

Documenting two things about the #atproto Jetstream: 1. If you pass a cursor timestamp parameter, Jetstream only goes back 24 hours. 2. There doesn't seem to be publicly available batch archives. Indiana University has an archive but only for other academics: osome.iu.edu/resources/da...

There is something much more deeply broken with link sharing than Twitter or Facebook's infamous downrankings. 8 likes and a quote tweet on a post with a broken link. People don't read the actual articles anymore.

I say this as someone that roots for Bluesky and thinks they will eventually figure things out but likes are declining again. My hypothesis is there is structurally something that promotes national politics here, and it suffocates the ability to talk about anything else. We need topics. #atproto

This "spotting a bear climbing a fence" video made me doubletake. I think everyone has a spidersense these days something like this has to be AI. The explanation from the author though is it was good old fashioned acting and video editing. One of the last of our era? www.instagram.com/p/DQ7fz1eEi9Q/

Tmux + DBM. Two of the crazier ways to run a production service. You will be missed. cc @lapsu.bsky.social

Lately ChatGPT-5 in reasoning mode is substantially better than Claude or others in accuracy. When source text is obviously incorrect, most models seem to simply accept it, as they are taught not to hallucinate. But GPT-5 is reasoning & catching a lot of issues in source texts the way a human might.

Docker Swarm seems under hyped for some reason.

Two listens on Electric SQL + Phoenix Sync that you might find interesting @lapsu.bsky.social 1. www.localfirst.fm/18 2. www.youtube.com/watch?v=4IWS...