Sign in

Mike Caulfield

@mikecaulfield.bsky.social
36K followers 2.3K following 3.3K posts

Author: Verified: How to Think Straight, Get Duped Less, and Make Better Decisions about What to Believe Online (University of Chicago Press). Researcher, infolit/misinfo/rhetoric/civic reasoning. Currently researching AI as tool for critical thinking.

PostsRepliesMedia
Mike Caulfield @mikecaulfield.bsky.social · 6h
These people do not know what a cookie is
3120
Reposted by Mike Caulfield
Jo @whatsjo.bsky.social · 9h
me: Hey, I can’t sleep. Do you think it’s because of stress? hand puppet: *rolls over on my pillow to look at me* you have had a lot on your plate recently.
436079
Mike Caulfield @mikecaulfield.bsky.social · 02/10/2026
Me, after eating all the doughnuts I brought to work for my colleagues "man, those top pot donuts sure go fast huh"
1170
Reposted by Mike Caulfield
Mike Caulfield @mikecaulfield.bsky.social · 01/10/2026
Built a film fact-checker that checks facts in a film summary by watching the film. I am as surprised as anyone else I managed to do this. mikecaulfield.substack.com/p/fact-check...
mikecaulfield.substack.com
Fact-checking a film summary by having your computer watch the film
My fact-checking AI agent now can stream public domain films, make contact sheets, pair that with transcribed dialogue from the audio stream, and fact-check the film summary.
1245
Mike Caulfield @mikecaulfield.bsky.social · 01/10/2026
Built a film fact-checker that checks facts in a film summary by watching the film. I am as surprised as anyone else I managed to do this. mikecaulfield.substack.com/p/fact-check...
mikecaulfield.substack.com
Fact-checking a film summary by having your computer watch the film
My fact-checking AI agent now can stream public domain films, make contact sheets, pair that with transcribed dialogue from the audio stream, and fact-check the film summary.
1245
Mike Caulfield @mikecaulfield.bsky.social · 30/09/2026
Kind of wild, the agentic film fact-checker I built can now use computer vision to check if a film summary accurately orders a sequence of events. Wasn’t there a prediction by someone that LLMs wouldn’t be able to watch a film and adequately explain it by 2030? We are not there yet but headed there.
19 contact sheets streamed from the public-domain Internet Archive copy. Each tile is stamped with its time in the film (h:mm:ss); the sheet names are the checker's own labels. Click a sheet for full size.

The claim notes cite times in the film rather than sheet names; the 12 source-audit claims that cite a time are listed at the bottom.
2110
Mike Caulfield @mikecaulfield.bsky.social · 29/09/2026
Honestly replace two of the three debates with quizzes
47615
Mike Caulfield @mikecaulfield.bsky.social · 29/09/2026
Most of the increased accuracy on niche questions comes through search, not corpora, and it is good to remember that. mikecaulfield.substack.com/p/yes-when-l...
mikecaulfield.substack.com
Yes, when LLMs don’t use tools they still suck at niche questions
And that’s important for nonobvious reasons
0114
Mike Caulfield @mikecaulfield.bsky.social · 28/09/2026
New Couch to 4k episode on how bad the Neflix thriller The Ice Road is www.youtube.com/watch?v=Z1oP...
youtube.com
The Ice Road (2022): "A bagged salad of tropes"
YouTube video by Couch to 4K
140
Mike Caulfield @mikecaulfield.bsky.social · 25/09/2026
This isn't to let other models off the hook, but I'd do think the release of GPT-4o was a uniquely bad decision, and we will be sorting through its impacts for a while.
1162
Reposted by Mike Caulfield
The Great Pumpkin Papers 🎃🎃🎃 @professormusgrave.bsky.social · 20/09/2026
I think we need to take seriously that ordinary folks aren't meant to have audiences in the tens of thousands or more. This is novel and disruptive in human experiences
221530189
Mike Caulfield @mikecaulfield.bsky.social · 20/09/2026
youtu.be/yCp4VLesfW8
youtu.be
Sorcerer (1977): Your story ended back there
YouTube video by Couch to 4K
040
Mike Caulfield @mikecaulfield.bsky.social · 20/09/2026
might be approaching "was The Postman good actually" levels of film contrarianism
3200
Mike Caulfield @mikecaulfield.bsky.social · 20/09/2026
mikecaulfield.substack.com/p/the-expert...
mikecaulfield.substack.com
The “Experts Excel and Novices Don’t” critique gets the AI educational situation backwards
We want technologies in classrooms that reward knowing things, it’s good actually
260
Mike Caulfield @mikecaulfield.bsky.social · 19/09/2026
It took me four hours to write a definition that finally got my Jaws set right. Four hours! The toughest part is what you think makes up Jaws isn't really what makes jaws jaws and not Jurassic Park or separates it from other Spielberg 1980s fare. I finally came to it four hours in.
170
Mike Caulfield @mikecaulfield.bsky.social · 19/09/2026
I will say there's more truck drivers than you'd think in film 1940 to 1970. It's just a really useful profession for a screenwriter looking to have someone bump into plots that is not a lawyer/cop/doctor etc
4130
Mike Caulfield @mikecaulfield.bsky.social · 19/09/2026
The 1970s truck driving protagonist maximum
Peaking in 1977-78 11 films in two years
3160
Mike Caulfield @mikecaulfield.bsky.social · 18/09/2026
Cop thriller boom in 80s/90s seems over, lawyer mini boom still with us
080
Reposted by Mike Caulfield
Liberty Hardy @libertyhardy.bsky.social · 15/09/2026
Song-a-day! Today's song is "The Queen of America" by Russian Apartments: on.soundcloud.com/TcSuitSLKUMv... (@mikecaulfield.bsky.social I LOVE THIS SONG. I still listen to this song all the time!) What song are you loving these days?
soundcloud.com
The Queen of America (Newer Version)
The Queen of America ================== Well they made up a figure And they called it you And they bought you a market For the things you do And you dropped like a name Right into the scene You're the
151
Mike Caulfield @mikecaulfield.bsky.social · 12/09/2026
If writing is thinking and prompting is writing what follows from that? mikecaulfield.substack.com/p/prompts-ca...
mikecaulfield.substack.com
Prompts can be poetic
It’s worth thinking about an LLM not as something that writes, but something you write into.
2102
Mike Caulfield @mikecaulfield.bsky.social · 12/09/2026
I don't think you should be grading with AI but I also think if you try something like this you've got to do a bit more than dump a scoring rubric into a prompt.
2183
Mike Caulfield @mikecaulfield.bsky.social · 12/09/2026
May have got a little emotional writing my eternal sunshineness judging prompt

**A. A relationship as a constellation of privileged moments** | The characters relationship is told as a series of both small and emotional dramatic moments that together create the feeling of a relationship that in this world was fated to follow this arc, but is not defined simply by its ending. A silent perfect moment can weight against the greatest heartbreak or disappointment, it is all true at once. The idea of parallel universes, time loops, fungible reality, and fate weigh heavy. | A romance subplot revealed over time, fate as a typical romance driver and not an existential reflection on the way that different worlds and times and memories touch and form a whole that is all true at once.
090
Mike Caulfield @mikecaulfield.bsky.social · 11/09/2026
Pretty proud of this list, this comes from a decomposition of eternal sunshine of the spotless mind into atomic elements and produces a list of films that somehow aren't the same but have something similar plot.fyi/q?s=%3Dfilms...
plot.fyi
Films tagged is gondryesque eternal sunshine
Explore similar — and surprisingly dissimilar — films on plot.fyi
2176
Reposted by Mike Caulfield
Mike Caulfield @mikecaulfield.bsky.social · 10/09/2026
QT: they've created a pre crime division Actual article: they perform basic platform threat reporting Guys, this is exactly the shit the right cooked its brain on leading up to 1/6, you might want to chill out a bit
1435
Mike Caulfield @mikecaulfield.bsky.social · 10/09/2026
QT: they've created a pre crime division Actual article: they perform basic platform threat reporting Guys, this is exactly the shit the right cooked its brain on leading up to 1/6, you might want to chill out a bit
1435
Mike Caulfield @mikecaulfield.bsky.social · 09/09/2026
The fact this is not a meme is a failure of the internet
Brian depalma unimpressed face
2140
Mike Caulfield @mikecaulfield.bsky.social · 08/09/2026
The one comment I'll make is when everybody gets to choose the worst proponents of ideas to argue with everybody gets to be right and vindicated. And one problem is the worst proponents of the ideas are the people running the companies, who have become quite shameless.
31598
Mike Caulfield @mikecaulfield.bsky.social · 07/09/2026
I think this post is one of my more important ones if you're allowed to say things like that. I hope it's clear? mikecaulfield.substack.com/p/ai-as-fuzz...
mikecaulfield.substack.com
AI as “fuzzy disintermediation”: some patterns and risks
Some thoughts I haven’t seen elsewhere
1173
Mike Caulfield @mikecaulfield.bsky.social · 05/09/2026
One thing came up in the comments to this which is "people aren't going to be building a bunch of digital artifacts" This is true in one way. I think a lot of vibe-coding is going to be considered a class of digital "crapjects" (look it up) in the future. But it misses a larger point.
2170
Mike Caulfield @mikecaulfield.bsky.social · 05/09/2026
The key skills people are going to need to use LLMs effectively are around evaluation and verification. I think many people see that as "Oh, so I have to be the LLM's copy editor now" but this is because most people don't understand evaluation and verification in any systemic way.
911716
Mike Caulfield @mikecaulfield.bsky.social · 04/09/2026
I feel like if you watch Wyler's The Children's Hour you should be allowed to call into work the next day as "emotionally wrecked'
3100
Mike Caulfield @mikecaulfield.bsky.social · 03/09/2026
I know my questions aren't the norm but I swear I get hallucinations (old style ones!) a quarter of the time I ask AI overview a film question. I was looking up this joke...
2150
Mike Caulfield @mikecaulfield.bsky.social · 03/09/2026
This is fun. You can go on plot.fyi and say "give me matches for this absolutely depressing film, but funny" plot.fyi/q?s=%3Dmatch...
plot.fyi
Films like Sorcerer · comedy
comedy · similar films on plot.fyi
3110
Mike Caulfield @mikecaulfield.bsky.social · 31/08/2026
The fact this was filmed by Vittorio Storaro is just wild.
Footprints on the moon poster
020
Reposted by Mike Caulfield
Claus Wilke @clauswilke.com · 30/08/2026
I'm really frustrated by the constant fear-mongering about AI-designed superviruses. Can we focus on the things that should actually worry us, such as measles? blog.genesmindsmachines.com/p/im-sorry-y...
blog.genesmindsmachines.com
I'm sorry, you're not going to die from an AI-engineered supervirus
God save us from economists who think they can competently talk about biology
412557
Reposted by Mike Caulfield
Mike Caulfield @mikecaulfield.bsky.social · 31/08/2026
New couch to 4k is out, where I try and fail to convince Jim local hero is good and probably also anger local hero fans by saying a lot of them didn't get it youtu.be/XigCx1mhjkA?...
youtu.be
Local Hero (1983) and the projections of place
YouTube video by Couch to 4K
092
Mike Caulfield @mikecaulfield.bsky.social · 31/08/2026
New couch to 4k is out, where I try and fail to convince Jim local hero is good and probably also anger local hero fans by saying a lot of them didn't get it youtu.be/XigCx1mhjkA?...
youtu.be
Local Hero (1983) and the projections of place
YouTube video by Couch to 4K
092
Mike Caulfield @mikecaulfield.bsky.social · 31/08/2026
Every movie set in the past should open with a title card explaining whether they are using nominal or inflation adjusted dollar amounts
0221
Reposted by Mike Caulfield
Mike Caulfield @mikecaulfield.bsky.social · 30/08/2026
Tools matter, techniques matter, skill matters, context matters. If you don't understand this stuff I don't know why you would be in digital information literacy in the first place, but for some reason we've lost these core insights in this endless discussion about "AI". www.npr.org/2026/08/30/n...
npr.org
AI chatbots may be better than search engines in guarding against foreign propaganda
In a test, popular AI chatbots mostly debunked falsehoods spread by other countries and avoided uncritically spreading falsehoods better than search engines. AI summaries above search results fared wo...
1155
Mike Caulfield @mikecaulfield.bsky.social · 30/08/2026
Yes, I am quoted in this. But more generally this is the kind of detailed indepth reporting on AI we need so much more of, reporting that doesn't stop at "I saw this error in a response" but contextualizes when and where such errors occur against a baseline. www.npr.org/2026/08/30/n...
36418
Mike Caulfield @mikecaulfield.bsky.social · 29/08/2026
Runaway, or "Tom Selleck in Blade Runner but for Roombas" plot.fyi/s/Q116468?v=...
plot.fyi
Runaway (1984) — plot.fyi
100 min · RT 48% · 📺 Streaming options available. — Runaway is a 1984 American science fiction action film written and directed by Michael Crichton, starring Tom Selleck, Gene Simmons, Cynthia Rhodes ...
180
Reposted by Mike Caulfield
Mike Caulfield @mikecaulfield.bsky.social · 26/08/2026
Wrote a bit on how I agree with Masley that most concerns about data centers are driven by beliefs specifically about data centers (you should read that piece first) but added three notes mikecaulfield.substack.com/p/note-on-ma...
mikecaulfield.substack.com
Note on Masley’s “I think the data center backlash is mostly about data centers”
Three related points
0258
Mike Caulfield @mikecaulfield.bsky.social · 26/08/2026
I looked up this costume and I think we can drop those quotation marks
3411
Mike Caulfield @mikecaulfield.bsky.social · 26/08/2026
Wrote a bit on how I agree with Masley that most concerns about data centers are driven by beliefs specifically about data centers (you should read that piece first) but added three notes mikecaulfield.substack.com/p/note-on-ma...
mikecaulfield.substack.com
Note on Masley’s “I think the data center backlash is mostly about data centers”
Three related points
0258
Mike Caulfield @mikecaulfield.bsky.social · 25/08/2026
Got the spotify page for my new film podcast up. If people could subscribe, even if you don't plan to be a regular listener, it would help me in this way: maybe search would stop asking people if they really meant "Couch to 5K" open.spotify.com/show/4WGM4oj...
open.spotify.com
Couch to 4k: A Film Podcast
Podcast · Mike Caulfield · If you're looking for a podcast where people overthink films, this is your podcast. You want a podcast that references Bertolucci while talking about American Gigolo, talks ...
193
Mike Caulfield @mikecaulfield.bsky.social · 25/08/2026
I have a trade deficit with Criterion of over $400 this year in Blu-ray purchases, and if Criterion doesn't start buying some stuff from *me* there will be consequences.
0234
Reposted by Mike Caulfield
Ada Palmer @adapalmer.bsky.social · 25/08/2026
New Mexico flipped from fossil fuels to renewables in 5 years. In July 2021 its largest utility drew 83% of its power from coal & gas; this July, renewables supplied 62%. Solar carries the days, batteries release stored sun in the evening, & wind picks up after dark buff.ly/ejt1iFF #ShareGoodNewsToo
yaleclimateconnections.org
New Mexico’s clean energy success story » Yale Climate Connections
In just five years, solar, wind, and batteries overtook fossil fuels, keeping air conditioners running while slashing pollution.
4344133
Mike Caulfield @mikecaulfield.bsky.social · 25/08/2026
Interesting I think. One assumption people (including me) have made is if data centers are overbuilt that it would not result in cheaper compute for very long bc the equipment would age out so fast. Is that not true?
3141
Reposted by Mike Caulfield
Andres Guadamuz @technollama.bsky.social · 21/08/2026
This paper feels massive for some image lawsuits specifically, and I wonder if it can be replicated with text. If you remove an artist from a model, the output will remain the same. www.nature.com/articles/s41...
nature.com
Outputs of generative diffusion models are often unattributable - Nature Communications
The authors investigate the attribution of diffusion model outputs to training data and show that generated samples can become unattributable when models are trained on large datasets. They introduce ...
2298
Mike Caulfield @mikecaulfield.bsky.social · 24/08/2026
Look, what I will tell you is this, you know how NYT headlines lie about the things you support? They lie about the things you oppose too. If you get your worldview from shared NYT headlines you will be misinformed.
1232