La Emisora Que Veo#

I think that in recent AI history we haven’t seen a flex as big as Sora. It’s February 2024. After the success of AI image generation, there was massive interest in achieving something equivalent for video. Many startups were trying, and huge amounts of money were being invested.

Quality was improving slowly. The infamous Will Smith eating spaghetti from March 2023, generated with ModelScope,¹ was already behind us. Videos were starting to look better, though many were little more than pans with slight motion. 2023 closed with Runway leading the closed source space with Gen-2² and Stability AI at the forefront of open source with Stable Video Diffusion.³

Then, in February 2024, OpenAI dropped something absolutely shocking: Sora.⁴ It had illogical consistency, illogical quality, illogical actions within clips, and generations lasting up to a minute. Everything was illogical. They didn’t even call it a “video generator”—they presented it as a “world generator.”⁴

How was it possible that a single lab had such an advantage? And let me stress this: video was not a neglected area. There was a lot of money and many teams working intensely in parallel.

The Sora effect was immediate: everyone took a step back to rethink goals and strategies. If there had been doubts about how far AI video could go, now we all knew this level of quality was possible. Startups and the open source ecosystem got to work.

Within a few months, we saw a wave of competitive launches.

Even some large corporations felt compelled to show demos of things that clearly weren’t ready yet—like Veo 1 from Google⁹ or MovieGen from Meta,¹⁰ just to make it clear they “weren’t that far behind.”

Meanwhile, OSS advanced at breakneck speed thanks to techniques inherited from image generation: DiT, flow matching, rectified flow, and more. We got Open Sora,¹¹ LTX-Video¹² and CogVideoX.¹³

You might imagine OpenAI would use that time to dominate the market alone. But… no. Except for a few selected designers, artists, and film directors, no one had real access to the original Sora model.¹⁴ Apparently the model was extremely large and extremely expensive to serve,¹⁵ yet OpenAI claimed that keeping the model without public access was for “safety” reasons and the risks associated with generating video at that level of realism.¹⁴

Ten months after the announcement, in December 2024, ChatGPT subscribers were given access to sora-turbo, a much inferior version.¹⁶ Where was the consistency? Where were the one-minute videos? Where was the quality? Well, there was some aesthetic appeal… but overall, sora-turbo was an outright failure.

The one who truly took all the glory was Google. They announced Veo 3 in May 2025 during Google I/O: a massive leap forward, far superior quality, immediately available, with integrated audio.¹⁷ The king. Social media exploded. Memes, jokes, and viral videos made with Veo 3 flooded the internet. It was Google’s first truly viral AI product—and they won the race against OpenAI.

MidJourney joined the party later, in June, with its characteristic style and aesthetic.¹⁸ It’s astonishing to think that this bootstrapped startup reached $500 million in revenue in 2025.¹⁹

The rest of the year was marked by constant updates from all competitors in the video space: Kling, Hailuo, Luma, Runway… Even Google reinforced its leadership by releasing Veo 3.1.²⁰

Integrating AI Video into Audiovisual Content#

AI-generated video has reached a level good enough to be consumed, but it still hasn’t fully integrated into all types of audiovisual content.

Cinema—feature films, series, documentaries—has not adopted it yet. The reasons could be many: small quality details, duration limitations, lack of fine-grained control in proprietary models that outperform OSS… or perhaps simply the inherent slowness of the film industry. Honestly, I don’t know.

YouTube has adopted it slightly, mostly as illustrative material. Where AI video truly found its refuge was in short-form content: TikTok, Reels, Shorts. Short enough to be generated with one or two prompts. Trivial enough that its mistakes don’t even warrant complaint.

The Infinite AI Slop Machine#

Short-form platforms filled up with AI videos. We knew that, at some point, someone would launch a social network exclusively for AI-generated video. The question was who—and when. The answer came from two unexpected places: Meta and OpenAI, almost in the same week, in September 2025.

Meta went first with Vibes, an AI video feed inside its Meta AI app.²¹ Since they didn’t have a competitive in-house video model for multimedia generation, they decided to use models from MidJourney and Black Forest Labs.²²

OpenAI, on the other hand, launched a full-fledged, standalone social network powered by the also-new Sora 2,²³ a very strong model with a clear focus on social media content. Like Veo 3, it could generate audio and video together.²³ But the feature that turned it into a viral phenomenon was cameos: anyone could scan themselves and ask Sora 2 to generate a video featuring them in any situation.²³

Within days of launch, the internet was flooded with videos of celebrities doing absurd things. It wasn’t impossible before—but now it was easier than ever… and free.²³

The ins and outs of Sora’s level of success in the months after its launch, its user base, or the internal dynamics of that social network, are beyond my knowledge. I could dive into mega-labyrinths to bring you information, but I’m not going dumpster diving.

References#