Laya: Uncovering the future trends of the AI entertainment industry

The entertainment industry lacks the most content. What lacks is the speed at which it can make correct decisions in the face of massive content. Short Video platforms upload 500 hours of video per minute, generate tens of thousands of bullet screens per second in live broadcasts, and the UGC platform adds millions of new texts every day. In the past, the solution was to pile up people-auditors worked three shifts, operators marked manually, and recommended algorithm engineers adjusted their opinions until early morning. What System 1 decision models like Laya do is seemingly simple: Give a paragraph of text and return "whether to violate the rules","which interest label they belong to" and "what the user's intentions are", with mathematically calibrated probabilities. But 33 milliseconds, more than 100 languages, and Apache 2.0 are commercially available. Together, these three indicators mean that judgment has become a horizontally scalable infrastructure .

This is interesting. When "judgment" becomes an API with a delay of 33 milliseconds per call of US$0.0001 per call, what will the person who originally made the judgment do?

I know a boss who does live slicing business. Last year, his team had 40 labeling staff, but this year, he has cut 6. Those six people no longer do "labeling" but instead "teach models to label"-they sort out boundary cases where models are uncertain, write few-shot examples and feed them back for fine-tuning. Originally it was "human judgment, machine execution", but now it has become "machine judgment, man takes the bottom, and man teaches machine." Tools do not replace people. Tools force people from "executing judgment" to "defining judgment boundaries". This is the most naked projection of the theme "How tools reshape people who use tools" in the entertainment industry.

Of course there will be a retort: Laya is just a classifier, and generative AI is the future of entertainment. Sora can generate videos, Midjourney can produce maps, and classification models are at most an auxiliary tool. Why are we talking about reshaping?

This refutation sounds reasonable at first glance, but in fact it misses the true constraints of the industry. The bottleneck in the entertainment industry has never been "unable to produce content", but "producing too much content without anyone watching, unable to divide, and dare to publish it." Short Video platform has 600 million yuan a day and tens of billions of recommendation requests every day. No matter how fast a video is generated, it must first pass compliance judgment, interest label judgment, user intention judgment, and copyright risk judgment. These four levels are all classified decisions. Models like Laya don't steal the limelight of generation. What they hold is whether the generated content can reach the distribution link. Without a fast, accurate, and inexpensive judgment layer, no matter how strong the generation model is, it can only be piled up on the hard drive to mold.

Cross-border analogy. When Photoshop was first released in the 1980s, designers ridiculed it for "only retouching pictures but not creativity." Ten years later, designers who couldn't PS were eliminated; twenty years later, designers who could PS but couldn't write Prompt were eliminated again. Each generation of tools evolves, it forces practitioners to a higher level of abstraction-from "pixel-level operations" to "layer-level thinking" to "semantic instructions." Laya's reshaping of content operation, review, and recommendation strategies is exactly this path: explicit, standardized, and iterated judgment . In the past, operations relied on "feelings" to allocate weights, but now they have to write "feelings" into typed questions that Laya can understand, and organize boundary cases into training sets. Operations that cannot write typed questions are like designers who cannot use layers, which will be eliminated sooner or later by tools.

There is a company that makes overseas short dramas and goes overseas. I went to their office for a meal last year. They deployed Laya at the export of each sliced video: China, English, Spanish, Portuguese and Arabic parallel compliance, subject labeling, and payment conversion tendencies in five languages. Originally, 20 operators manually marked Excel tables, manually selected language versions, and manually selected countries for release. Now three "prompt engineers" maintain 200 typed questions, covering 12 countries for distribution. The boss told me: "In the past, when recruiting operators read English CET-4 and CET-6, but now recruiting people first let them write 10 Laya question templates, and if they don't write correctly, pass them directly." Language ability is no longer the threshold, but the ability to define and judge tasks has become a new threshold.

How to calculate business accounts? Laya Apache 2.0 protocol is self-managed, fine-tuned, and has no call limit. Compared with the price of closed-source APIs of a few dollars per million tokens, the marginal cost of self-built reasoning clusters drops to zero on the order of trillions of tokens. But the real money is not saving GPU money, but after pressing the "judgment delay" from hours to milliseconds, the new business model --real-time bidding advertising breaks, second-level compliance shelves of live slices, and user comment areas Real-time emotional intervention and simultaneous distribution of cross-language content. These businesses simply cannot run into ROI in a world of "slow judgment".

I make a bet: within two years, the "judgment center" of the head content platform will become a more core department than the "recommendation algorithm group". The recommendation algorithm solves the problem of "who to show it to", and judges the problem of "dare to give it, what label to put on, and how to cut traffic." Laya is just the first open source benchmark in this wave to reduce the delay to 33ms and make the calibration probability production-usable. There will be more follow-up, but the team that first built "judgment" is already eating the lunch of the old world with the new organizational structure .

To put it bluntly, Laya is not cool. It does not generate videos, write scripts, and draw characters. But it determines whether the generated video can be online and whether the script can pass review