ByteDance's Seaweed 2.0 Rewrites AI Video

Remember Will Smith eating spaghetti? That viral mess of AI-generated chaos from 2023, where his fingers melted into pasta and physics took a vacation? Three years later, Matt Wolfe shows us the same prompt run through ByteDance's Seaweed 2.0—and Smith looks entirely, disturbingly human.

That jump isn't incremental improvement. It's a different category of technology, and it's raising questions about what happens when synthetic media becomes indistinguishable from reality.

What Makes Seaweed 2.0 Different

ByteDance—the company behind TikTok—released Seaweed 2.0 this week with capabilities that set it apart from competitors. The model accepts four input types: text, image, audio, and video. No other commercially available video model currently handles all four, according to Wolfe's analysis.

The outputs are 15-second clips with dual-channel audio and what Wolfe describes as "some of the best lip-syncing and most realistic lip-syncing I've seen from any of these video models yet." Character consistency across shots holds up. Motion graphics work. Celebrity likenesses render accurately, voices included.

Users on X have generated everything from Waffle House fight scenarios to Lord of the Rings alternate cuts to One Piece characters throwing Apple laptops overboard. A developer built an agent that crawls product pages and auto-generates user-generated-content-style marketing videos. The technical execution is solid enough that determining authenticity requires careful examination.

The Copyright Question Nobody's Answering

Here's where it gets complicated: Seaweed 2.0 has no apparent guardrails against generating copyrighted characters or trademarked IP. Users are making Spongebob clips, Marvel heroes, anime characters—content that would trigger immediate legal responses if generated through OpenAI's Sora or Google's systems.

"I think one of the big differences here is that, well, this is from Chinese company ByteDance, who apparently is less concerned with copyrights and trademarks than some of the American companies," Wolfe notes. "Sora and Google and some of these other companies based in the US are not going to be able to get away with allowing people to generate copyright work like this."

This isn't speculation—it's observable reality. American AI companies face immediate litigation risk for enabling IP violations. Chinese companies operate under different regulatory frameworks, both domestically and in terms of international enforcement. US copyright holders can file lawsuits, but actually getting judgments enforced across borders presents substantial practical barriers.

The result: ByteDance can train on and generate content that gives their models capabilities American competitors legally can't match. Whether this constitutes an advantage or a liability depends entirely on your perspective and jurisdiction.

The Speed War Heats Up

While ByteDance dominated video generation headlines, OpenAI and Google both released significant updates focused on inference speed—how quickly models respond to prompts.

Google's Gemini 3 Deep Think launched for Ultra subscribers ($250/month) with benchmark scores that "blew everything out of the water" on reasoning tasks. It achieved gold medal results on 2025 International Physics and Chemistry Olympiad problems and scored 50.5% on the condensed matter theory benchmark—numbers that positioned it ahead of Claude Opus 4.6 and GPT-5.2 on several measures.

OpenAI countered with GPT-5.3 Codex Spark, which uses Cerebras chips to dramatically accelerate code generation. Wolfe tested it by asking for a Vampire Survivors clone. "I gave it a prompt. 50 seconds later, game ready to play," he reports. The game included working mechanics, XP systems, level-up choices, and enemy AI—functional if visually minimal.

In a side-by-side comparison video, Spark generated a playable Snake game in under 10 seconds. The standard model took 47 seconds for the same task. That's not just faster—it changes the workflow. Iteration cycles collapse. Real-time pair programming with AI becomes genuinely real-time.

Both models carry $200-250/month price tags, which positions them as professional tools rather than consumer products. The question is whether that pricing represents the actual cost of compute or an early adopter premium that'll compress as competition intensifies.

Open Source Closes the Gap

Perhaps the most significant development this week came from Z.ai's release of GLM-5, an open-source model that's "catching up to the state-of-the-art models" according to Wolfe's benchmark analysis.

On Humanity's Last Exam—a notoriously difficult academic reasoning benchmark—GLM-5 with tools scored 50.4, beating Claude Opus 4.5, Gemini 3 Pro, and GPT-5.2 with tools. On software engineering tests, it matched proprietary models. On browser automation tasks, it exceeded them.

Running GLM-5 locally requires serious hardware (two M3 Ultra Mac Studios with 512GB RAM each, roughly $20,000 total), but the point stands: open models are reaching capabilities that were exclusive to well-funded labs just months ago.

What makes this particularly interesting is GLM-5's approach to autonomous work. Researchers at EO1 gave it a goal—build a Game Boy Advanced emulator—then stepped back. The model worked for 24 hours, testing its own code, logging results, identifying problems, and iterating solutions without human intervention.

"We're getting to a point now where you don't just give it a task, you give it a goal," Wolfe explains. "The agent makes a plan. It then executes. It tests its plan. It adjusts. And then it keeps going for hours on that loop of iteration."

That's not a better chatbot. That's a different tool entirely.

What's Actually Changing

Strip away the benchmark wars and feature announcements, and three shifts emerge:

First, the gap between "AI-generated" and "real" continues narrowing in video, to the point where casual viewers won't reliably distinguish between them. That creates obvious problems for misinformation, but also opportunities for legitimate creative work that was previously impossible without substantial budgets.

Second, regulatory approaches are producing divergent capabilities. Chinese companies can train on and generate content that American companies legally cannot, which may give them technical advantages in certain domains while exposing them to different risks in others.

Third, the open-source ecosystem is catching up faster than most observers predicted. When open models match or exceed proprietary ones on key benchmarks, the entire economic structure of AI development gets called into question. Why pay $250/month for hosted access if you can run comparable models locally?

None of these trends resolve cleanly. They're tensions that will play out over months and years as technology, law, and economics negotiate new equilibria. But pretending we're still in the "Will Smith eating spaghetti" era would be missing the story.

—Marcus Chen-Ramirez