Google’s AI Video Gen Destroys ‘Sora’

Google’s Veo 2 edges out OpenAI’s Sora Turbo in AI video generation benchmarks. Capable of manufacturing 4K-resolution videos, Veo 2 responds to intricate filmmaking prompts, including lens specifications and camera effects, while mitigating common AI flaws like visual “hallucinations” and unrealistic physics.

Benchmarked against Meta’s MovieGenBench dataset, with human evaluators assessing 720p, eight-second clips, Veo 2 outperformed its competitors in each quality and prompt precision. Nonetheless, Google concedes the model stays challenged by intricate motion dynamics and complicated scene composition.

Initial deployments are confined to VideoFX, YouTube, and Vertex AI, with expansion to YouTube Shorts in 2025, embedding SynthID watermarks to mark AI-generated content.

Meanwhile, Imagen 3 pushes the envelope for AI image generation, serving up vibrant color balance, precise texturing, and stylistic versatility. Imagen 3 and Veo 2 can be available via ImageFX and VideoFX, with API and Google AI Studio access rolling out early next 12 months. Read more.

Day 8 of OpenAI’s 12-Day rollout lifts the paywall for ChatGPT’s search, unlocking real-time, web-sourced results for all registered users. The update prioritizes speed and reliability, especially on mobile, and adds integrated maps, voice search, and the choice to set ChatGPT because the default search engine. Results now mix text, visuals, videos, and interactive maps, with demos showcasing practical uses like finding local events, planning trips, and selecting restaurants.

Day 9 rolls out the total o1 reasoning model, but with access limited to Tier 5 developers—those with a minimum of one month of account history and a $1,000 monthly spend. This premium tool, priced at $15 per 750k words analyzed and $60 per 750k words generated, brings advanced capabilities. Key features include a “reasoning_effort” parameter for custom processing depth, function calling, image evaluation, and 60% fewer reasoning tokens for reduced latency.

The Realtime API now supports WebRTC for low-latency vocal AI with noise suppression and dynamic congestion control. Developers also gain “direct preference optimization,” a more intuitive fine-tuning method that ranks outputs over predefined input/output pairs.

Google Labs has dropped Whisk, a generative AI tool that reimagines image creation by specializing in visual inputs as an alternative of the standard text prompts. Using Google’s Gemini 2.0 Flash model, Whisk generates detailed descriptions of your images, that are then fed into Imagen 3 to capture the topic’s essence without exact replication.

The tool is designed for rapid creative exploration, letting users experiment with subjects, scenes, and styles—ideal for brainstorming sessions reasonably than pixel-perfect results. Early testers, especially in creative fields, find it a refreshing change from standard editing tools.

Users outside the US are locked out for now, and delays in image generation have been reported, likely attributable to the influx of users. Currently available in Google Labs, Whisk could possibly be the precursor to something larger. Read more.

Nvidia has launched the Jetson Orin Nano Super Developer Kit, a compact yet formidable AI supercomputer, now available for $249, reduced from $499. This iteration delivers a 1.7x performance increase over its predecessor, boasting a 70% performance gain, achieving 67 INT8 TOPS, and enhanced memory bandwidth of 102 GB/s.

With Nvidia’s Ampere GPU and Arm CPU, the system supports concurrent AI pipelines for demanding applications, resembling generative AI, robotics, and computer vision. The kit integrates with Nvidia’s comprehensive AI software suite, offering tools for vision AI and edge computing. Existing Jetson Orin Nano users will receive software updates to unlock these advancements without the necessity for brand spanking new hardware. Read more.