Text-to-video takes off + Trump on deepfakes

Also: Role of GenAI in the Indian election

Author

Philip Mohr
June 19, 2024

Hey, it’s Philip.

today we have an overview of the new incredible text-to-video models that are being released left and right.

Also Trump talks deepfakes, a post-mortem on GenAI in the Indian election and TikTok embracing AI-Avatars for ads.

For new readers, sign up here

NEWS RUNDOWN

  • Getty Images and Picsart launch a commercially-safe GenAI model solely trained on licensed creative content. (Picsart Blog)

  • Apple announced “Apple Intelligence” at WWDC and integrates OpenAI at OS level (via opt-in). Here is a great summary. Apple’s AI integration apparently won’t output photorealistic GenAI content to avoid deepfake misuse. (Decoder)

  • According to BBC investigations, TikTok has become a major battleground for intentional AI-generated misinformation around the UK election, targeting especially young voters. (BBC)

  • TikTok reveals new features on its ad platform to create own AI clones or use stock AI avatars that can speak multi-languages and create ad creatives via text-to-video. (The Verge)

TOP STORIES

Three new text-to-video models beat OpenAI’s Sora to launch: KLING, Runway Gen-3 Alpha & Dream Machine

The last two weeks were truly insane for the AI video space. We had back to back announcement of really impressive AI video generation models, lets break them down:

(1) KLING by Kuaishou (Chinese) (examples here)

  • it can generate 1080p resolution up to 2 min and is capable of ultra realistic outputs with impressive scene consistency

  • its already publicly available as a demo (for Chinese users only)

(2) Gen-3 Alpha by Runway ML (examples here)

  • it it current version only outputs 10 sec of video, but offers creators more artistic freedom with scene settings and cinematic directions that can be implemented in the prompt

  • Its not yet “out”, but should launch soon

(3) Dream Machine by Luma AI (examples here)

  • It creates high resolution 2 min videos depending on the frames per second and it supports text-to-video and image-to-video

  • the beta is publicly available here

Why its important:

It seems like 2024 might be “the year”, where video models could have their “ChatGPT-moment” and synthetic video creation explodes, while any “AI” disclosure efforts or content attribution systems are still in its infancy.

The video models from Google (Veo) and OpenAI (Sora) have only launched in a closed beta, likely due to fears of misuse (ahead of the US election) and regulatory pressure. AI startups seem to be more open to experiment in public.

Luma, Sora and others do add C2PA metadata and use some proprietary watermark, but as long as the publisher adoption lacks behind, we will be in for an interesting ride till we have industry wide adopted solutions.

TOP PICKS

Featured:

AI Safety & Deepfakes:

  • Trump talks with Logan Paul (!?!) about deepfake risks, how it could trigger a Nuclear War and how he was fooled by a deepfake of himself promoting products. (Time and Impaulsive Podcast)

  • Employees of OpenAI, Google’s Deepmind as well Geoffrey Hinton and others publish an open letter calling for more oversight. (Decoder and righttowarn.ai)

  • OpenAI Insiders Warn of a ‘Reckless’ Race for Dominance. (NYT)

  • Teenager arrested for creating deepfake nudes of 50 female students in Australia. (The Guardian)

  • “I fell for a deepfake Elon Musk’s cryptocurrency scam bait” - (Artifical Ignorance Newsletter)

  • Ad fraud using GenAI content is rising, contributing to 23% YoY growth from 2022 to 2023 with new fraud schemes and an 58% YoY increase in ad fraud on streaming platforms according to DoubleVerify. (Digiday)

  • The Hackerfactor blogs talks about C2PA design flaws and LinkedIn’s C2PA integration (Hackerfactor)

Interesting pieces:

  • Podcast: Leopold Aschenbrenner (ex-OpenAI Superalligment) talks about his time at OpenAI, the next wave to scale AI systems & why he thinks AGI is possible by 2027 - (Dwarkesh Podcast)

  • Steve Endacoot is co-running with his own AI clone “Steve AI” for the general elections in the UK - (The Guardian)

  • DuckDuckGo enables private & anonymous use of ChatGPT, Claude and Meta AI - (Livemint)

  • Imatag’s Blog Article on California Provenance, Authenticity and Watermarking Standards Act (Imatag)

  • The Billionaire Battle For AI's Future (Forbes)