The crown for open LLMs goes to…

Along with: RunwayML Gen-3 Alpha model available for use

Hey there, 

What a summer it has been! The temperatures rose across the globe - several cities, states, and countries saw new levels. You could use the same statement for GenerativeAI and it would be true - new models, better performance, better hardware, and high expectations all happening in the last 3-4 months.

Let’s get to the heat for this week!

What would be the format? Every week, we will break the newsletter into the following sections:

  • The Input - All about recent developments in AI

  • The Tools - Interesting finds and launches

  • The Algorithm - Resources for learning

  • The Output - Our reflection 

Table of Contents

Benchmarks and Leaderboards are always fascinating. They obviously provide us with a way to compare different models, but once they become popular, companies start fitting their models to gain on the benchmarks and hence the leaderboard. We have seen this happen on processors and multiple other places. 

Open LLM Leaderboard from HuggingFace had a different problem though - we saw so much acceleration and investments in AI in the last few years that the leaderboard had plateaued. HuggingFace made changes to how it benchmarks the model. (source)

As per the new Leaderboard, Alibaba’s Qwen2 takes the top spot on OpenLLMs!

RunwayML has recently made its Gen-3 Alpha model available for use. It offers the ability to create hyper-realistic videos from text, images, or video prompts.

So, how is it different from Gen-1 and Gen-2?

  • First, it requires a paid subscription starting at $12 per month. 

  • There is an upgrade in video quality, enhanced speed, fidelity, consistency, and motion. 

  • It includes features like Motion Brush and Director Mode. 

Recently Runway has shared some sample videos that easily show how big an upgrade is this model compared to its predecessor.

Currently, it is focusing on text-to-video conversions with plans to expand to image-to-video and video-to-video transformations. (source)

Meta has introduced a new AI system called Meta 3D Gen, designed to generate high-quality 3D assets from text descriptions rapidly. 

This system includes two components: Meta 3D AssetGen for creating 3D meshes and Meta 3D TextureGen for generating textures. 

Together, these technologies allow for the production of detailed 3D models with realistic textures and materials, enhancing applications in video games, architecture, and industrial design. 

The system promises to integrate seamlessly into existing 3D workflows with its support for physically based rendering materials, improving the realism of lighting and material effects on AI-generated objects. (source)

Amazon is making significant strides in AI by hiring the co-founders of Adept and licensing their AI technologies. The aim is to enhance Amazon's in-house AI capabilities, reducing reliance on third-party AI providers. 

David Luan, Adept's CEO, will join Amazon and report to AI chief Rohit Prasad, signaling a serious commitment to AI development. 

Meanwhile, Adept will concentrate on agent AI solutions, with leadership changes including Zach Brock as the new CEO. This strategic shift allows Adept to focus on its core AI projects without financial strain, suggesting a refocus rather than a retreat. (source)

ElevenLabs, an AI voice startup, has expanded the voice library in its Reader app to include AI-generated voices of iconic Hollywood stars like Judy Garland and James Dean. 

This feature, available on iOS, allows users to hear any digital text narrated in the voices of these late celebrities, enhancing the user experience with a touch of nostalgia. 

The company has secured partnerships with CMG Worldwide to use the voices legally and ethically, focusing on security to prevent misuse. (source)

LLMs have played a significant role in engineering and coding tasks but have been less explored in the domain of code and compiler optimization due to the intensive resources required for training. To bridge this gap, the Meta introduced the LLM Compiler.

This suite of pre-trained models is designed specifically for code optimization and is built on the foundation of Code Llama. 

It is currently available in two sizes, 7 billion and 13 billion parameters, under a commercial license. The compiler improves the understanding of compiler intermediate representations, assembly language, and optimization techniques, and enhances code size and conversions.

It has been trained on over 546 billion tokens of LLVM-IR and assembly code and fine-tuned to better interpret compiler behavior. (source)

CriticGPT, a model based on GPT-4, is developed to enhance the detection of errors in ChatGPT's code output. 

It is designed primarily to assist AI trainers by pointing out inaccuracies, making it particularly useful as ChatGPT's responses become subtler and harder to critique due to advances in AI reasoning. 

The integration of CriticGPT into the Reinforcement Learning from Human Feedback (RLHF) pipeline is underway, aimed at augmenting human capabilities in evaluating AI responses.

Experiments have shown that AI trainers, with the help of CriticGPT, are able to provide more accurate and comprehensive critiques, catching more errors and reducing the occurrence of minor, unhelpful feedback ("nitpicks") and hallucinations. (source)

Just amid the launch of CriticGPT, Anthropic has launched a bit similar but more diverse initiative aimed at developing robust third-party evaluations for AI models. 

This initiative is particularly focused on enhancing the reliability and effectiveness of AI assessments by incorporating diverse evaluation methods. 

These include A/B testing through crowdsourcing platforms and engaging domain experts for red teaming in sensitive areas like national security. 

The goal is to create more realistic and comprehensive evaluation processes that closely mimic real-world AI use, helping to address the unique challenges of assessing AI systems' performance and safety in various contexts. (source)

Character.AI by a16z has launched a new feature allowing users to make calls to AI characters directly.

This innovative feature supports over 1 million user-created voices in multiple languages including English, Spanish, Portuguese, Russian, Korean, Japanese, and Chinese, and is designed to offer smooth calling experiences with reduced latency.

Users can switch between calling and texting seamlessly and can interrupt the AI with a "Tap to interrupt" option.

The company tested this feature extensively before its public release, with over 3 million users making more than 20 million calls. These calls have been utilized for various purposes such as language practice, mock interviews, and enhancing role-playing game experiences. (source)

CEO Mark Zuckerberg has announced AI Studio on Instagram, a platform that enables creators to develop AI chatbot versions of themselves.

Currently, in the early stages, it allows users to interact with the chatbot versions of their favorite creators through Instagram messaging, which will soon be labeled as AI. These chatbots are designed to be helpful and engaging, enhancing both conversations and entertainment.

Safety measures have been implemented to prevent the AI from generating hallucinated messages.

Moreover, Meta is now planning to extend this functionality to small businesses, aiming to create more dynamic interactions within communities. (source)

Researchers at the University of Tokyo have developed a new technique to enhance robots by attaching living skin made from human cells, giving them more human-like expressions such as smiles and grimaces.

This artificial skin, which can self-heal from damages like scars and burns, aims to improve robots' interaction in roles across health care, service, and companionship by making them appear and behave more human-like. (source)

Time Magazine embraces the generative AI through significant partnerships with AI startups including OpenAI and ElevenLabs, to enhance content distribution and model training using its extensive archive, aiming to integrate AI more deeply into journalism and product development.

This partnership also allows Time to access new AI tools that could transform how news is delivered. Concurrently, Time is working with ElevenLabs to implement AI-driven voice narration of articles on its website, enhancing accessibility and engagement through audio. (source)

Apple's strategy of enhancing the longevity of its devices is leading to a reduced frequency of hardware upgrades among consumers. This shift is to focus more on software and artificial intelligence to stimulate growth. 

The newly introduced Apple Intelligence features in the upcoming iPhone 16 and other devices are expected to entice customers to upgrade.

These developments are crucial as Apple's financial growth becomes less reliant on device sales and more dependent on services and AI innovations. 

Additionally, the company's refusal to partner with Meta on AI, despite discussions, highlights Apple's selective strategy in choosing collaborations based on privacy and quality considerations. (source)

Starting July 9th, OpenAI plans to apply stricter rules on how its API can be used in countries that it doesn't officially support. 

This change will mostly impact China, Hong Kong, Russia, North Korea, and Iran, countries not included on OpenAI’s list of supported areas. 

This decision is due to these countries' previous misuse of AI for spreading propaganda and false information. OpenAI hopes this will prevent misuse, especially during sensitive times like elections. However, developers are facing challenges because they currently don’t have a way to easily check where their API traffic is coming from. (source)

Imagine a Professor taking a design course asked the students to build a weather app and someone built an app eerily similar to that of Apple’s Weather App. What should the Professor do? Figma just did that, it shelved its student AI because it learned too much from Apple’s design!

This is something we will see more often going forward and it remains to be seen how we as humanity tackle this challenge. (source)

These days, memes are a powerful tool for capturing public attention and enhancing engagement with your product.

Supermeme.ai offers a seamless solution by generating memes from text or video inputs. Simply upload your content, and the platform provides a meme template on any topic, complete with customizable captions. This makes it easier than ever to create content that resonates with your audience.

How to access the tool:

  • Signup

  • Add a text prompt.

  • Generate meme.

  • Get your humor-based meme in minutes

This was my trial prompt: How does it feel like to sit in a meeting on Monday morning?

  • In a recent interview, Meta CEO Mark Zuckerberg expressed his vision for the future of AI, emphasizing the importance of diversity in AI development over the notion of a single, centralized AI.

  • In the recent episode of Leading with Data, I had an amazing conversation with Dr. Kirk Borne, AI thought leader and Advisor about his views on the pivotal role of data leadership in shaping the future of AI technologies and his experience in driving innovative AI initiatives across various industries.

On a separate note, Google’s carbon emissions have grown 48% in a year! I am pretty sure other big companies (probably except Apple) would see a similar / higher rise!

As an AI professional, I feel a splitting sense of guilt! On one side I cheer for every new development on the GenAI front and the other I can see our actions accelerating climate change globally!

How are you feeling about it? Would love to hear from you.

How do you rate this issue of AI Emergence?

Would love to hear your thoughts

Login or Subscribe to participate in polls.

Join the conversation

or to participate.