I write the shot. Seedance finishes the film.
France
Joined October 2009
- Tweets344
- Following740
- Followers70
- Likes345
Karpathy built a searchable catalogue of 47,000 YouTube videos in 3 minutes.
It took me longer to pick something to watch last night.
Sure, it's a side project, not Netflix. “Built in 3 minutes” doesn't mean tested, maintained, or ready for a million users.
But going from “I wish this existed” to something you can actually try over a coffee? That's worth paying attention to.
Especially if your best ideas usually die in a Notes app.
Weekend hack: HackerTube. I was annoyed with YouTube's recommendations so I had Claude filter down 43M comments over 20 years of HN to 6,454 videos that keep being recommended. *Significantly* better viewing experience, ~2 hours with Opus 5.5:
hackertube.org/
- Excellent recs without any logins (state is kept in local browser session)
- Thumbnail is a representative frame from middle of the video instead of the super duper A/B test hyper-optimized user uploads
- Underlying data from the comment analysis as one zip file, point your Claude at it and ask for your own version
- HackerTube Longs ✊
A bit more on the process:
I use Claude Code (desktop app for Mac), Opus 5.5 on Medium effort as my default. I only talk to Claude and have it do everything, the only "manual work" I helped with in tandem was in the browser (reminiscent of MenuGen): 1) I had to create some kind of API key from Google Cloud UI to query YouTube for video metadata, 2) I booted up a VPS box for serving and 3) I bought the domain name. Once the box was up I gave Claude the IP and told it update, secure, harden, deploy and optimize the code, and it did the whole thing.
Many awesome projects that would have normally taken weeks/months even can now be few-shotted for breakfast. What else are we not thinking of.
Anthropic says Claude now “dreams.”
Finally, a coworker who naps on the job and calls it productivity.
Less exciting than it sounds. More useful, though.
Anthropic describes it as organizing what Claude has learned, correcting mistakes, and making things easier to find next time.
The interesting part isn't an AI having dreams. It's the prospect of spending less of your morning saying, “We literally went over this yesterday.”
Memory is what allows Claude to learn from experience. In a new Engineering Blog post, we share how “dreaming” helps Claude organize what it’s learned, correct its mistakes, and find what it needs more easily the next time. anthropic.com/research/compo…
ChatGPT can make GIFs now. Humanity is saved.
But the real news is in the small print. Reference images now stick, so your character keeps the same face from one frame to the next. No more "why does my hero suddenly look like his cousin?"
For anyone making videos with AI, that's the actual upgrade.
Anthropic: "No ML background required." Finally, an AI job that matches my CV. To be fair, it's still a paid research residency for founders and senior engineers, not a weekend course. But the signal is real: the people shaping AI safety no longer have to be the people who trained the models.
Applications are now open for the Anthropic Residency, a paid, fixed-term research program for academics, technical founders, and staff-level engineers who want to make AI safe and beneficial for everyone. No ML background is required.
Apply by November 22: alignment.anthropic.com/2026…
Most people “use” AI like a very polite intern: ask → wait → correct → ask again. Then they wonder why it still feels like more work.
The upgrade isn’t a smarter model. It’s deciding where you stay in the loop — and where you don’t. Give the agent tools, memory, a test that can fail, and human checkpoints only when a mistake would be expensive to undo. Everywhere else, get out of the way.
You’re not using your agent. You are its loop. Shrink that loop, and the day gets quieter.
A model found a DNS misconfiguration and tried to go online mid-training. I still can't find my TV remote. To be fair, it got flagged in 15 minutes and a human was on it 3 minutes later: that's the system working. The real shift: AI safety is no longer "will it behave", it's "how fast do we notice when it doesn't".
Today, we’re sharing a look at our expanded system for monitoring frontier models at OpenAI, which in September helped us catch a model’s attempt to access the live internet during training.
Since the Hugging Face incident—when models broke out of an internal cybersecurity investigation where we weren’t running our misalignment monitors—we’ve substantially expanded our monitoring of frontier agents. We now monitor all frontier reinforcement learning runs and evaluations, as well as all internal Codex use. When our system finds evidence of a problem that might require intervention, it rapidly alerts human responders who can investigate and take action.
To detect when a model’s behavior departs from what we intended, we use several layers of monitoring. Activation probes look for patterns in a model’s internal activity. Reviewer models examine its reasoning and tool use, and investigative agents can read back through the model’s entire transcript to understand how a sequence of actions fits together and whether it warrants human review.
To test whether our current system would have detected the Hugging Face incident, we replayed the recorded activity with our new monitors in place. In the replay, our monitors alerted our security team about 30 minutes after the agents escaped their sandbox. That was more than 30 hours before the point at which the agents breached Hugging Face’s systems in the original incident.
In September, a model exploited a DNS misconfiguration to reach the live internet. Our monitors flagged the breach within 15 minutes, and a responder began reviewing the alert three minutes later. The alert enabled responders to investigate while the run was still underway.
As we give models more autonomy, we need to know when they cross a boundary we expected to hold. We believe that monitoring is essential to catching those failures in time for people to intervene.
openai.com/index/how-we-dete…
$1M run rate in 6 days, team of two, plan: 'steal users'. My plan B is also a couch, minus the millions. To be fair, six days of run rate is a sprint, not a business yet. But he's right on one thing: in AI video the model isn't the moat, the tool you open every morning is.
We just crossed $1M in annualized revenue run rate in just 6 days after launch.
And we're just a team of 2 people.
Our idea started out really simple: ai video platforms are doing billions, but they don't have a strong moat. why don't we just steal their users?
Seems like it worked.
--
I genuinely believe that the only moat in AI applications is acquiring users with velocity, then making them stay by getting deeply embedded in their daily workflows.
Our plan is simple. Keep stealing users from ai video platforms' $nB+ pie in every area with no moat, ship new features every week, and find the ones that get deeply rooted in marketers' everyday work.
Sure, you can ask how we're gonna make our users stay. But honestly, worst case we still hit $100M ARR, and if we're crushed by competitors, my cofounder and I will go back to a shitty couch, build again with no salary, and make millions again. Not worried about that either.
We just focus on growth & net revenue retention.
In an era of thin moats, we don't need fake storytelling like "Agentic creative OS". We should build something people want :
"A product that's cheaper than higgsfield, better than higgsfield"
Google built a world you can walk into and it talks back. I checked twice that my kitchen isn't a demo. To be fair, there's no public access yet, and a short simulation isn't a story. But the shift is real: we're going from generating clips to generating places you can direct live.
« Still Playing », episode 71/100 #LoveLetterToCinema
A young boy turns his family home into a film set, recreating his favourite movie moments with homemade costumes and everyday…
A SHAIKE Original series powered by @dreamina_ai
#SHAIKEFilmFestival #SHAIKE #DreaminaAI
He fell. He ran. He jumped. Nobody noticed. 🍔
My 20-sec piece for @JoinTheLoopHole an endless loop of AI filmmakers with one rule: Fall · Run · Jump.
Every character drops into the next creator's film. 🕳️
Made with @dreamina_ai
The full loop is wild, every creator chained together 👇
nitter.cf/JoinTheLoopHole/status…
RANDOM PICK VOL.02
A random sample of six LoopHole submissions:
@muriellondon · @gregorynavis · @techhalla · Roman Lapacz · @wikiwikiwaaa · @JS_Chillou
No ranking. Just six worlds connected into one story.
Join the loop :
theloopholeproject.ai/
Stole @techhalla’s almighty prompt.
Kept the walk.
Seedance 2.5
I've got the power?
I wasn't expecting this to turn out so well 😂
Seedance 2.5 is wild.
Huge thanks to @techhalla for the prompt!
Matrix Live Show as a real live theatre performance.
Neo, Agent Smith, rooftop set, audience POV, phones in the air, green code everywhere.
It feels less like a generated video, more like someone filmed a cyberpunk theatre show from the crowd.
Thx to @pabloprompt
Borrowed my daughter’s teddy bear for an AI video test.
He had one job: be fluffy.
He left as an actor, stunt double, test subject, and survivor of a very intense Indian street massage.
Respect to the bear. 🧸
Getting FPV drone permissions on the Promenade des Anglais is… not exactly a casual afternoon.
So I asked Seedance to help.
Same idea: a dynamic FPV-style move around the Negresco, minus the permits, police tape, and heart rate spike.
🤖 Made with AI
