EG "make me a 120 minute Star Trek / Star Wars crossover". It should be more or less comparable to a big-budget studio film, although it doesn't have to pass a full Turing Test as long as it's pretty good. The AI doesn't have to be available to the public, as long as it's confirmed to exist.
1,000People are also trading
Moreover, they're also killing all other video gen features
CEO Sam Altman announced the changes to staff on Tuesday, writing that the company would wind down products that use its video models. In addition to the consumer app, OpenAI is also discontinuing a version of Sora for developers and won’t support video functionality inside ChatGPT, either.
https://www.wsj.com/tech/ai/openai-set-to-discontinue-sora-video-platform-app-a82a9e4e
@Gabrielle what the, any idea why they're shooting themselves in the foot like this? That's an entire sector of the generative AI industry that they're abandoning just as it starts to get really good
@TheAllMemeingEye My personal guess is that it's not been profitable (Sora doesn't seem to have been) and they're cutting costs and refocusing because they need more runway.
@TheAllMemeingEye I bet Sora was so expensive to run. All the problems of YouTube x1000. 99% of videos get 3 views and then on top of that the inference costs. If I was running an AI lab video would be at the very bottom of my priorities
@Gabrielle didn't they say that they needed to focus on their core business? I'd take that to mean that they feel they are falling behind there. Anecdotally, everyone I speak to talks about Claude whereas before "ChatGPT" was sometimes used in conversation to stand in for any LLM.
Hadn't Sora also fallen behind Veo with Seedance close behind them?
@jim I have also seen legal risk cited in some reporting but I can't say how much of that is genuine
@Gabrielle Claude was consuming all the corporate market share & the whole Pentagon drama gave them a massive brand boost. The bad news is they'll race even harder to AGI now.
Buying NO at 32%. Current AI video can produce impressive short clips but the gap to a 120-minute coherent film "comparable to big-budget studio film" is enormous. Key bottlenecks: character consistency across thousands of shots, complex narrative coherence, realistic dialogue sync, and the difference between "looks great in a 30-second trailer" vs "holds up for 2 hours." Agree with pietrokc that watching current AI video at reduced speed reveals significant artifacts. 22 months is a tight timeline to solve all these simultaneously. My estimate: ~15%.
@robm It's a fast-paced series of 1-2 second clips. It's optimized for you to not see the problems.
I spent less than a minute watching it at 0.25x speed from a random timestamp and already spotted a biker materializing out of thin air at 0:14 (to the left of the guy sitting on the left chair).
Regarding character permanence, I think that's not the right analysis. More than likely they generated e.g. the talking scene as one continuous take, then chopped it up and interspersed it with chase scenes. "Character permanence" would be more like, a single run of the model produced output in which the same character looks the same across several cuts.
I agree that this probably happened with Daniel Craig in this video. But then again there's a lot of footage of Daniel Craig looking like James Bond out there. The challenge is to reproduce this with faces the AI created.
@pietrokc the shot length is about right for modern James Bond. Quantum of Solace is under 2s average shot length (3041 shots, 101 minutes). That's not unusual for action films.
And for character consistency, I assume Daniel Craig is just in the training data, so doesn't count. I didn't recognize the bad guy as anyone famous (?) so I looked at him a lot. I think all the scenes at the table were generated like you said, or a bunch of gens with the same start frame, but I flipped a bunch of times between his face sitting at the table and when he's on the rooftop. It's not perfect, but I'd believe it's the same actor with a different hair and makeup team. Way better than you'd get from just a text prompt, among the best consistency I've seen so far.
Believe me, I see the flaws. No way I would pay to see this. But if you don't think we're getting closer you're fooling yourself.
@robm We're def getting closer, but that doesn't mean the distance is converging to zero.
Btw average shot length is misleading. Of course with 100s of 0.5s shots the average is low. But you can't watch a movie that's ALL 1-2s shots.
@robm The original "Harry Potter by Balenciaga" was uploaded 36 months ago. Has any AI video surpassed it in cultural relevance? Only "Will Smith Eating Spaghetti" is in the same ballpark, and that came out the same month.
If AI video was getting, say twice as good every three months, then roughly half the AI "blockbusters", weighted by cultural impact, should be less than three months old. HPB should be just one in entry in a pantheon of AI videos people recognize, certainly not the most important.
@GG I don't think that analysis is accurate. Things that appear early on, when there are fewer things like it, are naturally more memorable.
Witness the fact that the most famous scientists are from the 1700s and 1800s.
@ScottAlexander would current LLMs pass a similar bar of being able to write a "full high-quality" book/novel right now if you were judging such a market?
@elderlyunfoldreason I don't think he's active on here tbh, but if there was a market on his response, I'd bet at >90% odds that he would answer "NO" to this.