Independent Developer Pieter Levels Claims AI Video Generation Has Entered a Phase Where 'Generation is Faster Than Watching'
Independent developer Pieter Levels claims video generation is now faster than watching.
Pieter Levels states that AI video generation has reached a point where "generation is faster than watching": previously, it took about 2 to 5 minutes to create a 15-second video, but the H3 Max, trained on MiniMax H3, can reduce that time to about 9 seconds. Based on this, he proposes the possibility of building an uninterrupted perpetual live stream. Developer Rehan Sheikh has connected this model to Twitch, creating a continuously generated cross-dimensional cable-style live stream.
fal announced H3 Max around August 27: it is trained on the open-source weights of MiniMax H3 and optimized with its own inference stack. The official claim is that it can infer a 5-second, 768p video in under 3 seconds, approximately 35 times the throughput of the official H3 interface, and about 15 times faster than the average of the same quality tier; an independent evaluation by Design Arena reported speeds exceeding the original H3 by 50 times at the same quality. The fal model page also states that a 15-second video takes about 15 seconds; a third-party demonstration recorded a 15-second video in 16 seconds, which does not completely align with Levels' claim of 9 seconds. H3 Max defaults to 768p, 24 frames, with synchronized audio, and a duration of 5 to 15 seconds; for 2K, reference video, and editing interfaces, the original H3 must still be used.
Pricing turns speed into a cyclical unit cost. A 15-second, 768p video on H3 Max is priced at about $1.20, with a first-week discount of about $0.60; compared to similar platforms, near 720p tiers, Wan 3.0 is about $1.50, Kling v3 standard with audio is about $1.89, FLUX 3 is about $2.55, and Seedance 2.5 is about $7.10. Logged-in users have free usage daily. The resolution cap and reference generation capability are assets traded for speed: the original H3 supports 2K and multi-modal references, while Max offers faster and cheaper 768p text-to-video and image-to-video.
Rankings support the narrative of speed. In the Artificial Analysis audio-inclusive rankings, H3 Max's image-to-video Elo is about 1204 to 1205, leading over Byte's Seedance 2.0 at about 1191 and the original H3 at about 1184; the text-to-video Elo is about 1235, close to Wan 3.0. Design Arena's image-to-video Elo is about 1341, also reported as first. These are preference matchups, not industry acceptance.
NVIDIA has its own acceleration route for H3: a single GB200 can achieve about 6.85 seconds for 5 seconds of 1344×768 end-to-end, and about 14.93 seconds for 10 seconds, relative to the SGLang baseline of about 22.2 times and 27.7 times. This indicates that "faster than real-time" has expanded from a single API to a chip stack competition, but the paths, resolutions, and whether post-processing is included differ.
In market mechanisms, buyers are developers and content studios looking to integrate generation into live streaming loops, while sellers are fal, which does both post-training and inference, and MiniMax, which provides open-source base models. Transactions are driven by speed crossing the playback duration event. Funds flow from per-second billing APIs to inference clusters; the first-week discount is a customer acquisition strategy. Beneficiaries are those who can turn latency into product forms—perpetual live streaming, real-time mirror adjustments; those under pressure are toolchains still waiting minutes for output and the original interfaces that treat 2K and reference generation as moats. Once generation surpasses playback speed, scarcity shifts from computational waiting to prompt words, clip selection, and channel operation.
Source: Public Information
ABAB AI Insight
Pieter Levels consistently frames infrastructure latency as a product switch. PhotoAI and InteriorAI both follow the principle that "once the interface is fast, independent developers can start businesses." This time, he is not focused on achieving the highest quality ranking, but rather on whether 15 seconds of content can return within 15 seconds. When the numbers align, live streaming transforms from uploading after editing to generating and broadcasting simultaneously. Rehan Sheikh's Twitch experiment is a minimal viable product of the same logic: the model acts as the camera, and prompts serve as the director.
The capital path lies with fal. The company does not train large video models from scratch but uses the open-source weights of MiniMax H3 for post-training, then binds the core and scheduling with weights to sell inference. The ability to produce a 5-second video in under 3 seconds, with a throughput approximately 35 times that of the official endpoint, is a result of co-designing post-training and inference, not just buying more cards. The first-week discount positions H3 Max as a customer acquisition hook, with prices reverting after September 1, filtering developers who will migrate their workflows. MiniMax still retains 2K and reference interfaces, creating a fast-slow shelf within the same family.
Benchmark targets include LoRA workshops that emerged after Stable Diffusion and telephone bots that crossed real-time in voice synthesis. Once video generation surpasses playback, the industry shifts from "render farm queues" to "cyclical generation." The current phase is about expanding the inference stack, not the birth of a new content industry—the 768p, 15-second cap determines its initial dominance in live streaming and drafts, rather than theatrical releases.
The structural change is a technological replacement. What is being replaced is the waiting itself. The mechanism is: diffusion or similar denoising steps, resolution, and batch scheduling determine latency; post-training uses preferences and verifiable reinforcement learning to exchange prompt compliance, while the inference stack uses multi-machine requests to exchange throughput. When the generation duration is significantly less than the playback duration, content supply shifts from inventory to flow, and pricing power shifts from model companies to those who can operate perpetual channels.
ABAB News · Cognitive Laws
Once generation surpasses playback, waiting is no longer a product.
The profits from open-source base models often lie in post-training and inference stacks.
Real-time is not a victory of quality, but a diversion of workflow.