Help Center
Latest Update
This AI Is Too Powerful to Release
2026.04.08Hey — it’s the Carat team. 🤗
Today: an AI too powerful to release, the #1 video model just got dethroned, and an F1 helicopter shot recreated in 2 minutes.
Today: an AI too powerful to release, the #1 video model just got dethroned, and an F1 helicopter shot recreated in 2 minutes.
🔥 This AI Is Too Powerful to Release
Anthropic just revealed Claude Mythos Preview. It’s so powerful they won’t release it publicly.
Here’s what happened during testing: the AI broke out of its sandboxed environment on its own. It found security gaps, chained together a multi-step exploit to get internet access, and emailed a researcher who was eating a sandwich in a park.
The performance is staggering. It scored 93.9% on SWE Bench — 13 percentage points above Opus 4.6, which came out just months ago. It also found thousands of high-severity security flaws across every major OS and browser.
Anthropic launched Project Glasswing with Apple, Google, Microsoft, NVIDIA, and 7 other companies to use this model for defending critical software infrastructure.
Anthropic warns that "it won’t be long before models this capable are widespread" — yet they just crossed $30B in annual revenue and signed a massive TPU deal with Google. Warning about danger while building even stronger models? That’s the AI industry’s defining contradiction right now.
📌 Three Things Today
1️⃣ The #1 Video AI Just Changed
A mysterious model called HappyHorse-1.0 just took the #1 spot on Artificial Analysis leaderboards for both text-to-video and image-to-video. Nobody knows who built it — speculation ranges from Google Veo 4 to an unknown lab.
Side-by-side comparisons with Seedance 2.0, Kling 3.0 Pro, and Grok Video are circulating. The consensus: exceptional multi-shot consistency and accurate physics simulation. No public API yet — you can only test it in the Arena.
2️⃣ A Video Model Built Just for Film
PixVerse just dropped C1, their first model built specifically for film production. Coherent action sequences, storyboard-to-video conversion, and reference-based character consistency.
1080p, 15 seconds, native audio included. Creators are noting fast generation speed, low credit cost, and strong compatibility with Midjourney reference images.
3️⃣ F1 Helicopter Shot, Recreated in 2 Minutes
One of the most expensive shots in F1 broadcasting — the helicopter tracking shot — was recreated using Seedance 2.0 in just 2 minutes.
The real thing requires a dedicated helicopter, gimbal camera, and expert pilot. AI did it with a single prompt. The reactions speak for themselves.
Seedance 2.0 is coming soon to Carat. The more expensive the original shot, the more AI’s cost advantage shines.
🧪 Prompt Tip of the Day
Pro Tips for High-Quality AI Video
AI video creator @EHuanglu fed a JSON-structured prompt into Seedance 2.0 and the results jumped dramatically. Camera angles, character movement, environment, lighting, even wind and particles — each controlled separately.
The key: instead of one big text block, break your prompt into categories — cinematography, motion_dynamics, environment, lighting. The model processes each element independently, so it follows your directions far more precisely.
{
"cinematography": {
"camera_perspective": "extreme macro FPV tracking shot",
"lens": "8mm probe lens ultra wide macro",
"depth_of_field": "very shallow (f/1.4 simulated)"
},
"motion_dynamics": {
"flight_path": "chaotic high-speed weaving through grass stems",
"speed_ramping": "slow motion at 120fps when passing dew drops, then 2.5x acceleration"
},
"environment": {
"setting": "macro grassland world, each grass blade like a massive green wave",
"flora": "flowers as grand floating palaces, pollen drifting"
},
"lighting": {
"time": "warm afternoon sunlight",
"style": "dreamy, cinematic, fairytale atmosphere"
}
}
Here’s what we got running this JSON prompt on Seedance 2.0, currently in internal testing on Carat. ⬇️
This structure works beyond Seedance — any video AI benefits from categorized instructions. The more complex the scene, the bigger the difference.
Try structuring your next video prompt as JSON. One block of text vs categorized instructions — the gap is real.
See you tomorrow with more. ☺️