Mass Adoption

The current landscape of AI is intriguing. For years, society has interacted with AI-enhanced experiences through social media algorithms, e-commerce suggestions, Instagram filters, and facial recognition technology. However, it is the advent of generative AI that has sparked conversations about its mass adoption and implications. The involvement of sports makes it tangible for the general public. For instance, Wimbledon has partnered with IBM to launch AI-powered commentary this year. This initiative will be available on Wimbledon’s app and website, distinct from the BBC’s broadcasts. The AI has been programmed with tennis-specific terminology, preparing it to deliver engaging commentary and deeper insights. This marks a significant leap toward fully AI-driven match commentary, signaling that the future of sports coverage is upon us.

Synthetic Data

As generative AI continues to evolve, researchers face a significant challenge: their models may end up being trained on data that has already been artificially generated. This perplexing issue is discussed in the paper titled "The Curse of Recursion: Training on Generated Data Makes Models Forget," which highlights a phenomenon known as Model Collapse. This occurs when models rely on their own generated outputs, resulting in irreversible flaws and a loss of original content. This concern extends beyond language models to Variational Autoencoders and Gaussian Mixture Models. The study emphasizes the necessity of addressing this challenge to preserve the benefits of large-scale web data training and points to the increasing importance of authentic human interaction data as language models proliferate online.

Governments Are in Full Attention

Even before Wimbledon kicks off, governments are already addressing AI challenges. Recently, the Senate invited leaders from the open-source sector for discussions. Senator Chuck Schumer (D-N.Y.) is poised to lead AI regulation efforts with his “SAFE Innovation framework,” aimed at safeguarding national security and economic stability while promoting AI's potential. He seeks accountability in addressing issues like copyright, misinformation, and bias, urging businesses to clarify AI system outputs in layman's terms.

President Biden has also acknowledged the rapid pace of change, anticipating more transformations in the next decade than in the previous fifty, driven by AI. His administration is actively organizing meetings to devise secure AI strategies.

Check this article on Turing Post: https://www.turingpost.com/p/fod8

New Name in LLMs — Inflection AI

In a noteworthy development, President Biden attended a fundraiser co-hosted by Reid Hoffman, known for founding LinkedIn and investing in significant tech ventures, including OpenAI. Hoffman is also a co-founder of Inflection AI, which recently launched its advanced language model, Inflection-1. This model reportedly surpasses GPT-3.5 and other prominent benchmarks.

Furthermore, Mustafa Suleyman, co-founder of Inflection, proposed rethinking the Turing test. His "modern Turing test" or artificial capable intelligence (ACI) evaluates AI based on its ability to set and achieve complex objectives with minimal human intervention, moving beyond the traditional focus on artificial general intelligence (AGI). Suleyman envisions an experiment where AI transforms $100,000 into $1 million through strategic initiatives, believing this milestone could be reached within two years.

Inflection AI has also been highlighted in discussions around companies competing with OpenAI.

Busted!

In the EU, OpenAI has been criticized for lobbying to avoid being classified as "high risk" under the EU AI Act, successfully securing exemption in the final draft but now facing stricter transparency requirements. Currently, none of the involved companies comply with the draft legislation.

Transparency is crucial! Insights from AI Snake Oil suggest that transparency reporting is technically feasible and can largely be automated, requiring minimal user interaction samples for analysis. A collaborative team from Stanford and Princeton has proposed three regulatory ideas for generative AI.

What’s New in Generative AI

CBInsights has released its annual list of the 100 most promising private AI companies globally. Recently, significant advancements in audio and image sectors have been observed. Meta introduced Voicebox, a new text-to-speech generative AI model, while Google launched AudioPaLM, a multimodal speech model that integrates techniques from both PaLM2 and AudioLM. ElevenLabs, known for its voice cloning technology, recently secured $19 million in funding.

In image generation, Midjourney’s version 5.2 has been released, boasting improvements in aesthetics and coherence. Stability AI launched SDXL 0.9, which produces hyper-realistic images, focusing on enhancing the realism of human hands.

Robotic Achievements

Google DeepMind has announced RoboCat, a self-improving AI agent capable of learning tasks with just a few demonstrations. Carnegie Mellon University researchers have developed robots that learn household chores by watching YouTube videos, utilizing a new algorithm that allows them to learn without being in the same environment as the video content.

Trends in Computer Vision

Andrew Ng shared his observations from the recent Computer Vision and Pattern Recognition Conference (CVPR), noting several emerging trends:

1. Vision transformers are increasingly seen as a viable alternative to traditional convolutional neural networks.
2. Advancements in image editing are giving users greater control in image generation, especially for facial features.
3. Neural Radiance Fields (NeRF) are gaining traction for converting 2D images into 3D scenes.
4. There is heightened interest in multimodal models capable of processing both image and text inputs.
5. Ongoing research on self-driving cars suggests that large, pre-trained transformers could be pivotal.

Papers, Articles, and Releases to Watch:

• “Tune As You Scale: Hyperparameter Optimization For Compute Efficient Training” introduces a new hyperparameter optimization method.
• “How Reinforcement Learning from Human Feedback (RLHF) Works” explains how RL models utilize human feedback for performance enhancement.
• “LLM Survey Report — MLOps Community” discusses diverse LLM applications and deployment challenges.
• “Eliminating Bias in AI/ML” addresses the complexities of bias in AI and ML.
• “Fast Segment Anything” presents FastSAM, a real-time solution for efficient segmentation tasks in computer vision.
• MosaicML has released MPT-30B, an enhanced open-source model with an 8k context length.
• HuggingFace has launched OpenLLaMA, an open-source version of Meta’s LLM LLaMA.

What I Want to Try Next Week:

• GPT-Engineer, a novel tool that generates software or games based on user concepts, refining its output through clarifying questions.

Thank you for reading! Feel free to share with friends and colleagues.

Best, Turing Post