OpenAI and Google race to launch multimodal LLM plus AI demos with Sunny Madra

OpenAI and Google race to launch multimodal LLM plus AI demos with Sunny Madra:

Zillow took down their GPT-based interface due to concerns about losing business value and customer relationships.
Fear of GPT becoming the apex aggregator is driving companies to avoid plugins that don't provide direct customer interaction.
Plugins like ChatGPT Enterprise offer a way for companies to integrate generative AI into their existing applications.
Canva is an example of a plugin that uses ChatGPT to generate templates for social media posts or graphics.
The integration between ChatGPT and Canva allows users to request specific templates directly from the chat interface.
The goal is for plugins like Canva to have their own LLMs that can generate content on the fly, providing more customized results.

All-In Summit 2023 recap:

The All-In Summit featured world-class speakers who delivered exceptional content and conversations.
There were discussions about improving future events, including potentially changing ticketing options and venue sizes.
Hosting large-scale events in cities like LA poses logistical challenges, but efforts were made to provide a great experience for attendees.

Google's chances in the race for multimodal LLM:

Google has access to vast image and video databases through platforms like Google Images and YouTube, giving them an advantage in multimodal capabilities.
However, OpenAI's focus on integrating different modalities into a single platform could also position them as a strong competitor.

The potential of sidekick-type applications:

Sidekicks are overlays on existing applications that leverage context and knowledge from those applications.
This approach allows for more seamless integration of multimodal prompts without requiring complete application redesigns.

Using Google Search experiments:

Google search experiments, such as generative search, provide advanced features that enhance the search experience.
These experiments may need to be enabled within the Google Labs settings or by using a personal Gmail account rather than a Google domain account.

Canva plugin integration with ChatGPT:

The Canva plugin allows users to request templates for graphics or social media posts directly from the chat interface.
Users can generate different template options and open them in Canva for further customization.

The future of multimodal UIs:

There is a need for more specialized UIs that cater to specific use cases, such as interacting with documents or code.
Improvements in UI design will allow for more natural and efficient interactions with AI models.

Sunny demos Canva’s ChatGPT plugin:

Discussion on plugins' limitations:

Code LLaMa's potential and Falcon 180B’s unique features:

Code Llama released as Meta's version of Code Interpreter, made available open source
Falcon language model has different size models, ranging from 1 billion to 30 billion parameters
Falcon 180B demo released, UAE involved in its development
Comparison between parameter sizes of chat GPT4 and Falcon 180B
Chat GPT4 likely consists of four or five 200 billion parameter models working together

Potential improvements in code-related AI tools:

Code Llama currently not as good as Code Interpreter based on industry benchmarks, but can improve through modification due to being open source
Predicted that by January 2024, it will surpass the capability of Code Interpreter
Distance between chat GPT4 and other models narrowing in certain verticals like code

Implications of reinforcement learning with human feedback:

AI-enhanced headshots:

Using AI to turn photos into professional headshots
Mentioned issues with overly sexualized or racially biased outputs
Growing trend, with multiple services offering AI-generated headshots at affordable prices
Cost comparison with traditional professional photography

Closing remarks on This Week in AI:

Mon Sep 18 2023