Lightnews — Scholar-powered news

Anders Sandberg @arenamontanus.bsky.social · 10h

I did not expect that the AI might learn the physics well but could not explain it even to itself.

Anders Sandberg @arenamontanus.bsky.social · 10h

I vividly remembering debating with EY and others back in the 00s how much physics could be inferred from a YouTube clip given to a superintelligent actor. I still think you need more than just a few seconds, but we are starting to see a (wide) upper bound here.

1 1

Anders Sandberg @arenamontanus.bsky.social · 10h

But the fact that we do get emergent models, generalisation, and problem-solving in what is "just" predicting sequences of frames is a big thing. Should not be surprising after the LLMs, but it takes many learning samples for us humans...

1

Anders Sandberg @arenamontanus.bsky.social · 10h

It is when we get models that can actually answer "why did the water go right?" or explain why the reflection on the spoon shifted that things get really interesting (and scary). That likely requires new architecture. Maybe. Never bet against emergence.

1

Anders Sandberg @arenamontanus.bsky.social · 10h

I think the deep problem still remains that different parts of the model have no good access to each other: the video model likely cannot explain color mixing and viscous fluid dynamics. ("Likely": this is where clever prompt engineering may reveal surprising things!)

1

Anders Sandberg @arenamontanus.bsky.social · 10h

The problem is as usual reliability: when it works, it works. But it can also fail for random reasons. Specialized models can clearly beat it performance-wise... but as in the case of LLMs, the general foundation model matters.

1

Anders Sandberg @arenamontanus.bsky.social · 10h

Wiedemer, T., Li, Y., Vicol, P., Gu, S. S., Matarese, N., Swersky, K., ... & Geirhos, R. (2025). Video models are zero-shot learners and reasoners. arXiv preprint arXiv:2509.20328.
video-zero-shot.github.io

Video models are zero-shot learners and reasoners

Video models like Veo 3 are on a path to become vision foundation models.

video-zero-shot.github.io

1 1

Anders Sandberg @arenamontanus.bsky.social · 10h

The as always breathless video: www.youtube.com/watch?v=spn_...

DeepMind’s AI Just Solved Video Generation In A Way Nobody Expected

YouTube video by Two Minute Papers

www.youtube.com

1 1

Anders Sandberg @arenamontanus.bsky.social · 10h

Paper and video makes an important point: video generation contains an amazing amount of world knowledge and emergent strategic thinking that in the past one could write entire SIGGRAPH (or AI) papers about how to model for one particular case. A sweet demo of the Bitter Lesson.

1 2

Anders Sandberg @arenamontanus.bsky.social · 12h

I think taking papers (and ideas) really seriously is surprisingly uncommon. As Clarke put it, there is often both failure of imagination and failure of nerve.

Reposted by Anders Sandberg

Forethought @forethought-org.bsky.social · 1d

We’re hiring!

Society isn’t prepared for a world with superhuman AI. If you want to help, consider applying to one of our research roles:
forethought.org/careers/res...

Not sure if you’re a good fit? See more in the reply (or just apply — it doesn’t take long)

1 2 5

Reposted by Anders Sandberg

Toby Ord @tobyord.bsky.social · 1d

The other evening I attended the launch of David Edmonds' book on Peter Singer's Shallow Pond. I was quite struck when he called it 'the most influential thought experiment in the history of moral philosophy' yet with no influence for its first 30 years…
🧵
press.princeton.edu/books/hardco...

Death in a Shallow Pond

From the bestselling coauthor of Wittgenstein’s Poker, a fascinating account of Peter Singer’s controversial “drowning child” thought experiment—and how it changed the way people think about charitabl...

press.princeton.edu

1 1 7

Reposted by Anders Sandberg

Institute for Ethics and Emerging Technologies @ieet.bsky.social · 1d

Join us in Paris June 11-12, 2026 www.thehastingscenter.org/hastings-cen...

1 2

Anders Sandberg @arenamontanus.bsky.social · 6d

Defending what we think is the right and good online every day forever is necessary but tiresome.
I suspect many will insist everything is the Algorithm's fault instead. It gets us off the hook.

1

Anders Sandberg @arenamontanus.bsky.social · 6d

Robin of course proposes futarchy x.com/robinhanson/... as a structing mechanism. I imagine there are many other possibilities if we just look for them. Similarly, quality-rewarding sub-media exist, but will rarely become mass things.

Robin Hanson on X: ""Perhaps social media is a genuine threat to democracy precisely because it gives us democracy good and hard. Perhaps democracy can’t survive too much democracy." Futarchy would solve (part of) this. https://t.co/5qEF7Ey5Rt" / X

"Perhaps social media is a genuine threat to democracy precisely because it gives us democracy good and hard. Perhaps democracy can’t survive too much democracy." Futarchy would solve (part of) this. https://t.co/5qEF7Ey5Rt

x.com

1 1

Anders Sandberg @arenamontanus.bsky.social · 6d

Structuring discourse is however necessary for sophisticated action. Just because a lot of people want something does not mean implementing it is a good idea, or that the preferred implementation reaches the aim. Populism is bad at running complex societies.

1

Anders Sandberg @arenamontanus.bsky.social · 6d

From the perspective of an open society this might not be too bad: the problem isn't the bad views but when there is no way of making one's voice heard, getting recognition of an issue, and the ability to change the issue accordingly. Liberal democracy has just been decent at it.

1

Anders Sandberg @arenamontanus.bsky.social · 6d

Note that this model should worry anybody serious about any political ideology: populism has no color. Authoritarians trying to rein in public discourse have their work cut out for them. Democrats wanting to structure discourse still have to deal with a public with "bad" views.

1

Anders Sandberg @arenamontanus.bsky.social · 6d

The standard response is to blame platform providers for optimizing for this, but optimizing for scale and openness likely does the same. The Eternal September started in 1994. Indeed, every step in broadening access has led to a quality issue online.

1

Anders Sandberg @arenamontanus.bsky.social · 6d

When markets become more efficient they often become less diverse, since there are often economies of scale and people have less diverse preferences than they tell themselves. This might be what has happened on social media. Efficient memetic replicators get amplified.

1

Anders Sandberg @arenamontanus.bsky.social · 6d

Most political parties and institutions have mechanisms to weed out crazies, wrongthink and disruption (those that don't, doesn't last). The traditional classical liberal marketplace of ideas might have been implicitly built on an assumed platform of decent society.

1

Anders Sandberg @arenamontanus.bsky.social · 6d

The basic claim is that a lot of the bad views ventilated actually are popular views that until recently were conveniently excluded by elite-run platforms. People are populists by nature rather than some algorithmic distortion.

1

Anders Sandberg @arenamontanus.bsky.social · 6d

This essay aligns well with my long-running annoyance with blaming social media for everything. Maybe the problem is a truly efficient marketplace of ideas instead.
www.conspicuouscognition.com/p/is-social-...

Is Social Media Destroying Democracy—Or Giving It To Us Good And Hard?

It’s easier to blame the algorithm than the bewildered herd.

www.conspicuouscognition.com

1 4

Anders Sandberg @arenamontanus.bsky.social · 8d

To a large degree it is used to support the "cope bubble", the story that AI is not *really* a challenge to the human condition, the normal economy, normal politics, or human safety. It is both a defense of the status quo in society, and a status quo in mindset.

1

Anders Sandberg @arenamontanus.bsky.social · 8d

As a Swede, this is the weirdest and most unlikely part of the entire Prize tradition: *calling people at inopportune times?!* That is simply not done!

4