Story: At last weeks re:Invent 2023 Amazon unveiled a new model in their Titan family of AI models. The new Titan text-to-image model allows users of their Bedrock AI development platform to create new images and adjust existing ones.
Key Findings:
New AI Image Generation Model: Amazon's Titan Image Generator joins a growing list of AI image generators
Advanced Features and Training: The model can seamlessly modify image backgrounds and is trained on diverse datasets.
Intellectual Property and Indemnification Policies: Amid concerns about intellectual property in generative AI, Amazon has committed to protecting customers against copyright violations with its AI indemnification policy.
Watermarking: Images generated by Titan Image Generator will feature a tamper-resistant watermark to counter AI-generated misinformation and abuse.
Pixit‘s Two Cents: We at Pixit love to see that the market of foundational models gets more and more diversified. It is also great to see that more and more companies start to incorporate temper-resistant watermarking in AI generated content, as it is much needed in regard to disinformation campaigns. The fact that big tech with their deep pockets assures customers in protecting them against copyright claims is a two-sided sword. On the one hand it is great for their customers on the other hand it is something new startups or open source projects simply cannot compete with, in turn giving the incumbents more power in the future.
Story: Stability AI, renowned for Stable Diffusion, has recently been in the spotlight due to potential sale discussions and internal leadership challenges. Bloomberg reports that the company is considering selling after Coatue Management, a significant investor, raised concerns about the company’s financials and leadership under CEO Emad Mostaque.
Key Findings:
Leadership Challenges: CEO Emad Mostaque’s leadership and disorganized culture had prompted several senior managers to leave
Call for CEOs Resignation: Coatue Management urged CEO Emad Mostaque’s resignation due to management style issues
Unwavering Commitment: Despite sale rumors, Stability AI remains committed to its core mission of releasing AI models
Pixit‘s Two Cents: The situation at Stability AI presents a classic case of the challenges tech companies face when balancing investor expectations with visionary leadership. We hope that Stability AI will continue releasing state of the art generative AI models.
Story: Another week where Google’s DeepMind has achieved a groundbreaking feat. This time in material science with its AI tool, GNoME, which stands for Graph Networks for Materials Exploration. This deep learning tool discovered an astonishing 2.2 million new crystal structures, including 380,000 stable materials, potentially revolutionizing modern technologies like computer chips, batteries, and solar panels. GNoME's ability to predict the stability of new materials has dramatically increased the speed and efficiency of material discovery.
Key Findings:
Unprecedented Scale of Discovery: GNoME identified 2.2 million new crystals, vastly expanding the database of technologically viable materials, with a focus on 380,000 stable materials promising for future technologies.
Enhancing Computational Approaches: Previously, computational methods could identify 48,000 stable crystals. GNoME's contribution has raised this to 421,000, showing a significant improvement in the accuracy of AI-guided predictions.
Innovative Applications: The discovery includes potential superconductors and lithium-ion conductors, which could revolutionize electronics and improve rechargeable battery performance.
Pixit‘s Two Cents: Deepmind understands to excel with their AI research. Two weeks ago they challenged the status quo in weather forecasting and last week they made big strides in watermarking AI generated content. This weeks discovery has the potential to give the scientific community around material sciences a tremendous boost. However it has to be seen how many of the newly discovered crystals really hold up to the expectations.
Creating high-quality 3D avatars using DiffusionAvatars: Researchers at the Technical University of Munich (TUM) developed DiffusionAvatars, a method for creating high-quality 3D avatars with realistic facial expressions by combining 2D diffusion models and 3D neural networks.
1960’s chatbot ELIZA outperformed GPT-3.5 in a recent Turing Test: In a recent Turing Test study, the 1960s chatbot ELIZA outperformed GPT-3.5 by convincing more participants of its human-like responses, highlighting the enduring effectiveness of early AI models in human interaction. The study may have some caveats though, since GPT-3.5 does not intend to behave as a human.
Kandinsky 3.0 is here: One year after the release of the Kandinsky diffusion model, Kandinsky 2.0, Sber released Kandinsky 3.0 with the goal of improving text understanding and image quality.
Consistent and Controllable Image-to-Video Synthesis for Character Animation: Using a reference image and a pose sequence, researchers from the Institute for Intelligent Computing at the Alibaba Group, found a new method to generate video clips.