AI & Startups October 09 - October 15

Gen AI Product News

Meta shares method to watermark AI images

Meta and Inria just released a new technique to create invisible watermarks on images from open-source AI models. Their Stable Signature method traces generated images back to the original model — and potentially offers a solution to one of the more heavily critiqued elements of AI creations.

Google launches healthcare-specific generative AI

Google just launched new features in its Vertex AI search platform, combining the tech with the medically-focused model Med-PaLM 2 to aid health groups in unifying clinical data.

The details:

• New tools in Vertex enable queries across siloed sources like EHRs, clinical notes, medical images, and genomic data.
• Natural language questions can be combined with structured search for intuitive exploration of patient records.
• Google touts security measures like encryption and HIPAA compliance for protecting sensitive health data.
• Med-PaLM 2, trained on medical data, can provide responses to clinical questions linked to specific patient info.

Adobe beefs up creativity tools

Adobe just unveiled major upgrades to its suite of AI-powered creative tools like Photoshop and Illustrator, promising more realistic imagery and new generative abilities.

The details:

• The latest version of Adobe's AI system, Firefly, was trained on 70% more image data, enabling much higher quality visual generations.
• New controls allow users to better fine-tune AI output, like uploading a photo to mimic its style.
• Illustrator can now auto-generate vector graphics from text prompts, while Express gains AI capabilities for quick social media creations.
• Check out our AI Training section for an in-depth guide!

Adobe demos AI-powered video editing

Adobe just revealed an experimental tool at the company’s MAX Conference called Project Fast Fill, featuring a new way to edit videos using AI generation and text prompts, and it’s insane!

The details:

• Fast Fill is similar to Google's Magic Eraser or generative fill for photos, but works with moving video.
• It uses generative AI to fill in missing areas after removing objects/people from clips.
• Text prompts can alter colors, lighting, and more without traditional editing.
• It's still an early experiment, but it could eventually come to platforms like Premiere Pro and After Effects.

ElevenLabs releases AI voice dubbing

Voice AI startup ElevenLabs just launched a groundbreaking new AI Dubbing tool that translates speech into different languages while preserving the original speaker's voice.

The details:

• AI Dubbing converts spoken audio/video into over 20 languages while still maintaining nuance and emotion.
• The tech combines voice cloning, synthesis, and text-processing research by ElevenLabs.
• It handles multiple speakers well, differentiating between voices and music/noise.
• The release comes on the heels of a similar launch from Spotify, which utilized OpenAI’s Whisper tech.

Replit opens AI coding tools to masses

Replit announced free, open access to its AI-powered suite of coding tools — including debugging, autocompletion, and text-to-code features. The company believes putting AI in every developer's hands is key to empowering software's next billion creators.

Meta's AI celebs face backlash from confused fans

Meta recently unveiled AI chatbots mimicking celebs like Kendall Jenner and Snoop Dogg. However, giving them different names like "Billie" and "Dungeon Master" has mainly spawned confusion, skepticism, and backlash.

The details:

• Additional celebrities with Meta’s alter-ego personalities include Tom Brady, Mr. Beast, Charli D’Amelio, Paris Hilton, and Dwayne Wade.
• It's unclear to some fans if it's the real celebrity or AI, with limited disclaimers — leading to questions of whether the star consented.
• Meta reportedly paid celebs like Jenner up to $5 million for rights to their face and voice — but fans are confused by the name changes.
• Legal and ethical issues could arise if AI says things counter to celebrities' brand relationships.

Foundational Models

Large Language and Vision Assistant (LLaVA)

Researchers from Stanford, UW-Madison and Columbia introduced Large Language and Vision Assistant (LLaVA), a new opensource AI system that could rival GPT-4 for visual and language understanding. It’s open-sourced and completely free to use.

The details:

• LLaVA connects CLIP (an AI system that understands images) with LLaMA, an open-source text AI model.
• In a demo, LLaVA showed it could understand and have convos about images, much like the proprietary GPT-4 system, despite having far less training data.
• Unlike the private GPT-4, LLaVA's code, trained model weights, and generated training data are freely available online.

Demo:

LLaVA Github Repo

Anthropic breakthrough makes AI understandable

Anthropic just developed a new method to interpret the individual neurons inside LLMs like Claude, helping researchers better understand and decode the model’s reasoning.

The details:

• Neurons usually respond to many unrelated concepts, making them hard to understand in isolation.
• Anthropic found a way to decompose groups of neurons into simpler "features" with clearer meanings.
• For example, they extracted DNA, legal text, and HTML features from a blend of 500 neurons — and stimulating a single feature consistently caused expected model behavior.
• Features also link together in systems for complex functions like producing valid HTML.

Gen AI VC

Saronic Ferries , a defense startup building autonomous ships, raises $55M

Multimodal database startup TileDB raises $34M in Series B funding

Cleanlab Raises $25M Series A to Automatically Increase the Value and Accuracy of the World’s Enterprise Data

FluidAI Medical Medical Announces $15M in Series A Funding for AI-Driven Postoperative Monitor at HLTH 2023

Gen AI Global Policy

Global divides emerge on AI regulation

While the EU pushes for strict, unified regulation of AI — not everyone is on board. Countries in Southeast Asia are reportedly taking a business-friendly, hands-off approach to AI regulation — aiming to support innovation while letting local governments shape policy.

Space Force pauses generative AI over security

The U.S. Space Force is pressing pause on its use of generative AI — citing cybersecurity and data handling concerns. The agency says it aims to adopt the revolutionary tech responsibly — but there are concerns that the move could put the division behind rivals like China.