This weekly roundup thread is intended for all culture war posts. 'Culture war' is vaguely defined, but it basically means controversial issues that fall along set tribal lines. Arguments over culture war issues generate a lot of heat and little light, and few deeply entrenched people ever change their minds. This thread is for voicing opinions and analyzing the state of the discussion while trying to optimize for light over heat.
Optimistically, we think that engaging with people you disagree with is worth your time, and so is being nice! Pessimistically, there are many dynamics that can lead discussions on Culture War topics to become unproductive. There's a human tendency to divide along tribal lines, praising your ingroup and vilifying your outgroup - and if you think you find it easy to criticize your ingroup, then it may be that your outgroup is not who you think it is. Extremists with opposing positions can feed off each other, highlighting each other's worst points to justify their own angry rhetoric, which becomes in turn a new example of bad behavior for the other side to highlight.
We would like to avoid these negative dynamics. Accordingly, we ask that you do not use this thread for waging the Culture War. Examples of waging the Culture War:
-
Shaming.
-
Attempting to 'build consensus' or enforce ideological conformity.
-
Making sweeping generalizations to vilify a group you dislike.
-
Recruiting for a cause.
-
Posting links that could be summarized as 'Boo outgroup!' Basically, if your content is 'Can you believe what Those People did this week?' then you should either refrain from posting, or do some very patient work to contextualize and/or steel-man the relevant viewpoint.
In general, you should argue to understand, not to win. This thread is not territory to be claimed by one group or another; indeed, the aim is to have many different viewpoints represented here. Thus, we also ask that you follow some guidelines:
-
Speak plainly. Avoid sarcasm and mockery. When disagreeing with someone, state your objections explicitly.
-
Be as precise and charitable as you can. Don't paraphrase unflatteringly.
-
Don't imply that someone said something they did not say, even if you think it follows from what they said.
-
Write like everyone is reading and you want them to be included in the discussion.
On an ad hoc basis, the mods will try to compile a list of the best posts/comments from the previous week, posted in Quality Contribution threads and archived at /r/TheThread. You may nominate a comment for this list by clicking on 'report' at the bottom of the post and typing 'Actually a quality contribution' as the report reason.
Jump in the discussion.
No email address required.
Notes -
Sora is dead
It turns out that spending hundreds of millions for users to make useless slop videos was having a meaningfully negative financial impact. The bizarre thing is that Disney signed a $1b deal with OpenAI just a few months ago - who fucked up here? Of course, there are many more video AI tools out there, with fewer considerations for copyright law. But for now, Hollywood doesn't have much to worry about, at least on this front.
Startups like Runway and Chinese companies like Kling are still around, and AI video generation is only getting more popular. The big players like Google and Tiktok have better in house models than OpenAI. It is a crowded space. Sora was first to market for this caliber of video models, but the space has left Sora behind.
There are 2 reasons OpenAI abandoned Sora, and it has little to do with the viability of AI video gen.
The primary reason is because OpenAI was spread too thin. Enterprise agents are the trillion dollar market, and OpenAI is currently losing to Anthropic. It spooked OpenAI and late last year, they changed focus to Codex. Since then, OpenAI has deprioritized ChatGPT, voice models, music models and ofc, Sora. It's not that video generation is not a lucrative market. But, it's 2 orders of magnitude smaller than the enterprise agents market.
The secondary reason is that OpenAI is not well positioned to win here. Video generation and editing is primarily about control and iterative improvement. You start with a story board -> create a 1st draft -> use a vast tool-kit to iteratively get it to the final version. Sora is great at creating the 1st draft, but the likes of Adobe & Apple have the whole tool-kit built out. Unless the model is capable of fine edits, it will not be a useful substitute for filming it manually. ChatGPT: the product is a thin wrapper on top of ChatGPT: the model. The effort needed to turn it into a commercially viable product is minor when compared to the research effort of creating a GPT v-next model. On the other hand, most of the effort with a commercially viable video generation product is in the product engineering, not the model itself. That's asking a lot lot of effort from OpenAI in an area they are not best equipped to beat seasoned product engineering teams at.
tl;dr: Video generation will survive. The bubble isn't popping. A better analogy is - 'The gold rush ended because someone just discovered a Diamond mine'.
For context, I worked for an LLM/diffusion based content-gen AI startup for a few years. I was very early to this. Frankly, it an indictment of my judgement that I am not yet a millionaire. Should've joined OpenAI or Anthropic in early 2023 while I still had the chance. SMH
It seems evident by their actions that engineers at OpenAI lacked the ability or capacity to use GPT5 to cost-effectively write an Adobe Premiere competitor but with Sora-integration, with UI that's just as good, just as intuitive and user-friendly for longtime video editing professionals, just as stable and responsive, etc.
I wonder if/when AI companies will reach the point where they could just do that for any arbitrary existing software. At what point could one of these companies just instruct the AI to generate an Excel clone that has perfect backwards compatibility to MS Excel, but also has their AI integrated in, and consistently get out a viable software product as the result? What about a Windows clone that has perfect compatibility to all Windows-compatible software, but also has their AI integrated in? What about an Oblivion clone that has perfect compatibility to all existing save files, but also doesn't require major QOL overhaul and performance mods to make enjoyable and also has their AI integrated in?
It relates to the traditional wisdom - "Regulations are written in blood". Much like it, "enterprise software is written in the tears of disillusioned engineers".
Yes, Adobe Premier is a few million lines of code, and LLMs can create millions of lines of code within weeks. However, Adobe premier wasn't one-shotted by a person, and an agent can't one-shot it either. The only way to build an excellent enterprise tool is to build a shitty enterprise tool, get feedback, and improve it with time. In startup speak, this private feedback is referred to as 'moat'. LLMs make this loop faster, but you can't skip it. eg: State of the art forecasting have great benchmarks, but routinely generalize worse than ARIMA. The only way to know this is to have spent years in the trenches trying to get some new paper into prod. The techniques needed for AWS to provide 11 (0.000000000001% chance of failure) 9s of durability will never be discovered by OpenAI unless they pay-off the hundreds of AWS people who meticulously got it there. That information is guarded in vault somewhere.
Coding agents are an exception because AI companies are their own customers (so feedback loops can precede adoption) and the code/discussions/learnings are publicly available.
The value in the text/images/media/any content that form the feedback comes from how modifying the software in a way guided by the feedback improves the software as judged by the people who gave the feedback (and people like them), not in the fact that content was generated by humans using the software and expressing their opinions. Generating the feedback that way through actual humans who used the software is a great way to ensure that that the feedback is valuable in this way, but I don't see why a sufficiently advanced LLM (or LLM-based tool) couldn't generate that feedback with just as much value (i.e. modifying the software in a way guided by that LLM-generated feedback improves the software as judged by the people who would have given the feedback, i.e. target audience), just by predicting the next word. And then modify the software through iterations until the feedback crosses some threshold of asking for small enough changes or something. I don't think this would be considered a "one-shot," but it certainly seems like it would require almost as little investment in human effort. It's just that the LLM-based tools don't seem sufficiently advanced (or perhaps they're not sufficiently fast?).
I don't think LLMs can generate meaningful human-like feedback of what it feels like to use the software. They just don't see the UI in the way that humans do. And it's not clear that increasing their capabilities can ever fix this.
Still, I do expect that they'll get better and better at iterating quickly and nondestructively based on your feedback, so while it won't be a fully automated dev cycle, I wouldn't be surprised if bespoke AI software replaces giant professional products eventually.
I don't see why LLMs would need to "see" the UI in a way similar to humans in order to generate meaningfully useful feedback for improving the UI (as well as any other element of the software) as judged by humans. It's not like the LLM would need to reason out "this UI element here gets in the way of this process due to that issue, etc." or "in my experience of trying to use this software in my workflow, this UI element could be improved by moving it here," or whatever. It'd be doing naked dumb pattern matching, of predicting words based on the prompt (which would include the sequence of 1s and 0s that make up the software, as well as instructions to produce text that a helpful human tester would provide, or the like) and its weights. There's no proof that this would work, but I also see no reason why simply scaling up current techniques and/or making them faster wouldn't allow LLMs to generate feedback like this which is just as useful as human user feedback.
Because it's really hard to predict how the software is going to be used, and it's not something that can be reasoned out. If that were the case, software companies with full UI teams wouldn't still be responding to user suggestions 50 years into the industry's history. Watch some of Tantacrul's videos on music notation software. He's a software developer by trade and a composer by hobby, so he has tried pretty much every major program on the market, and his video on MuseScore a few years ago resulted in him becoming the head of the development team. Music notation software is particularly ripe for this kind of criticism because it's all notoriously difficult to use and people such as myself who occasionally dabble in music have tried pretty much all of the available programs in a desperate attempt to find something that isn't going to piss us off. Highlights from the comments:
Sibelius
Finale
Dorico
Muse Score
Watch the videos. They're long, but highly entertaining. And keep in mind that he's only scratching the surface with respect to the problems he describes, and they're all either deliberate design choices or the result of being bound by the limitations of the existing codebase. I don't think you can just get an LLM to figure this stuff out.
More options
Context Copy link
I have my doubts, but you make a good point. A lot of the other emergent capabilities have been quite surprising, so there's no guarantee that this is out of the question, either.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link