https://www.testingcatalog.com/video-overviews-settings-spotted-in-notebooklm-ahead-of-i-o/

Looks pretty good

@256 An interesting question is: what ELO would a model have to have to resolve this question positive?

My (completely made up) guess is ~10k ELO

@LoganZoellner This question, and the post it responds to, doesn't make sense. ELO is only a measure of relative ability. If all models are total garbage except for one which is just mildly garbage, the mildly garbage one can achieve 10k ELO.

~~Another reason the original post doesn't make sense is that this is an image-to-video model, whereas this market is about text-to-video.~~ It actually does make sense

@pietrokc hi, text-to-video and image-to-video are very connected things (the former isnt possible without the latter), im not sure why you would say that it doesn't make sense to post about it here.

@256 That's fair, I apologize. I assumed you had missed the distinction, which I see often on Manifold. Removed!

o3:

"Bottom line

Today’s tools dazzle in seconds-long bursts, but stretching that to a polished two-hour blockbuster—in just 2½ years—requires unprecedented leaps in length, consistency, and automated post-production. Absent a visible paradigm shift, I’d fade the 36 % and price it nearer 20 %"

Large NO limit orders at 39% and 48%, expiring in a month @musteval @mckiev @robm

(deleted)

People are imagining that we'll accomplish this capability before the intelligence explosion, but it'll come *after*.

We are unlikely to get a system this capable without first automating AI research

@JonasVollmer my thoughts exactly! Generating a Hollywood-quality movie from a single prompt is certainly something that far exceeds the baseline of “AGI”. It requires the combined efforts of hundreds of humans working for several years.

Are all you yes holders shorting Warner Brothers?

@Joshua You can also short humanity/money, and then shorting Warner Brothers doesn't make sense.

@Joshua Trump’s tariffs are gonna incentivize AI movie creation

@Joshua why? The concept of entertainment is going to transform in the future. If you can create a movie with a prompt, think of what professionals are going to be able to do. Not sure about WB, but hopefully they'll manage to evolve

This happens a decent ways before superintelligence, and that plausibly happens ~2028. Am I missing something important about the market here or is everyone bearish on AI progress?

@Maloew I think I disagree with you about everything here. The main thing is that I don't think there are any promising leads on the taste problem - transformers, even with RL, haven't broken past "hardworking, dedicated, but wholly untalented person" in any task I feel qualified to evaluate them on, including coding and math. There are maybe worlds where we get a kind of superintelligence on this timeline, but those are worlds where the taste problem turns out to be irrelevant and RL on coding is all you need, which are worlds where nobody is even remotely interested in making movies with AI, least of all the AI.

@speck "hardworking, dedicated, but wholly untalented person" is probably a higher benchmark than AGI under my preferred definition of "able to automate 90% of economically valuable work".

@LoganZoellner I agree. I don't think we're all that close to saturating that metric, multimodality is probably needed for that and progress there has been slower than I expected, but it absolutely seems possible to me that we achieve some form of AGI within the next decade, maybe even early in the decade. But that still leaves a lot of creative or systems design work un-automated. I think the main point I want to make is that even if you think some sort of AGI is possible, there are AGIs which are not superintelligent, and there are some reasons to believe that that's what we're on trajectory for.

@speck

> there are AGIs which are not superintelligent

I think it's more accurate to say that intelligence is "spiky". AI is already superhuman in many domains (chess, go, poker). At some point we will cross the line of "well, 90% of tasks can now be automated" but that won't magically imply the other 10% has been.

@Maloew You are missing everything; the whole statement is false.

"Generate a full high-quality movie to a prompt" is already superintelligence. Obviously no human can do that.

this seems like a big update

Single character, multiple locations

@robm is it though?

I watched the example short films on their website. Despite having a human "director" who prompted, aggregated, and added a soundtrack to the clips, none of them have even, what I'd call, a plot, dialogue, or really... anything that we haven't seen before?

@bens I'm talking specifically about consistent characters. You can find my comment from a few weeks ago criticizing the earlier gen4 release for missing on this. Any demos on their website from before today are irrelevant to my point, this is a new feature dropped today.

imho, the 3 hardest things remaining are:

• writing great scripts

• audio generation and sync

• consistent characters

If this is as good as the samples they posted today, then one of those 3 barriers is down.

@robm the example they give of consistent characters on their website has like 3 outfit changes across the same clip and only vaguely looks the same in that it is a girl with dark hair and bangs in every clip. I think this is wool-over-eyes at best, and selective editing more likely.

@bens this one?

https://youtu.be/Eb_SxalzkGg

Here is me in this thread literally making the same critique when that was posted 4 weeks ago.

https://manifold.markets/ScottAlexander/in-2028-will-an-ai-be-able-to-gener#vfe32gtb1nn

Check the stuff from their yt channel today. Things have changed.