Welcome to the third edition of the odo newsletter—a free, weekly newsletter on AI for product builders. We’ll cover:

• Google DeepMind’s RT-2 Launch

• IBM and NASA Open-source an AI Model

• Google’s AI Search Gets Better

• A tip on using OpenAI’s Custom Instructions

• …and more!

News of the week 🗞️ 

🤖 Google DeepMind’s RT-2 Applies LLM Techniques to the Physical World

Google DeepMind launched RT-2, a vision-language-action model that improved the robot's ability to generalize and perform reasoning by following the LLM’s playbook–leveraging the transformer-based techniques and large amounts of internet data. The new model showed significant improvements in handling tasks that it wasn’t explicitly trained on, such as interfacing with unseen objects, backgrounds, and environments. Notable example: RT-2 was asked to pick up an extinct animal from a table full of fruits, balls, toy animals, and other objects, and successfully picked up a toy dinosaur, despite not being explicitly programmed to map “dinosaurs” to a category of “extinct animals.”

From chatbots to robots…The initial conversations about AI applications have mostly centered around how to improve productivity in desk-jobs, such as customer support or research. RT-2 highlights how combining different AI technologies can accelerate progress beyond developing specific capabilities in a silo. Maybe the world of WALL-Es and EVEs is not too distant a future.

🌐 IBM and NASA Partner to Deploy Geospatial AI Foundation Model

Last Thursday, IBM and NASA made the watsonx.ai geospatial foundation model available on Hugging Face, a leading open-source community for ML models. The model was trained on NASA’s satellite data, and then fine-tuned for flood prediction with labeled data on flood and burn scars mapping across the continental U.S. This fine-tuned model showed a 15% improvement in precision for flood prediction with only half the labeled data.

Let’s make Greta proud…It’s an understatement to say that climate change is one of the most pressing societal issues of our generation, given the huge surge in wildfires and floods in recent years. Releasing such a foundation model, especially in an open-source forum, will help unlock rapid developments in climate efforts by reducing the need to pre-train AI on large amounts of climate data. We’ve seen some interesting companies trying to apply AI to solve a variety of climate problems, such as Pano AI.

🎥 Google’s AI Search Gets More Video and Better Links

Another news from Google this week. Google’s Search Generative Experience (AI-based feature in Search Labs) is now showing multimedia results and better links. If you haven’t already, you can try the feature yourself by going to Search Labs on the search screen and enabling SGE.

No longer a sleeping giant…When OpenAI first launched ChatGPT last year, it seemed like Google’s search business had finally met its match. Since then, Google has shipped a number of improvements to its search experience. For consumers, these improvements will hopefully help them find actual answers to their questions, as opposed to links to whoever bid the highest to the search terms.

AI product highlight ✋ 

“Um…so…like…when do you think you can maybe…possibly…get this done?”

Have you ever had to ask someone for something at work and felt extremely awkward? Communication is one of the most important skills for product managers, but it’s not always easy to know how you are doing and what you can do to improve.

If this resonates with you, look no further than Yoodli AI (no, it’s not Yooni AI…). It’s a personalized AI communications coach that will listen to you silently during meetings and give you feedback.

Disclaimer: We are not getting paid to highlight the product. We just think it’s cool and want to share!

For the AI nerds 🤓 

OpenAI’s Custom Instructions

OpenAI recently released custom instructions which allows users to have more control over how ChatGPT responds. Previously, you might have given specific prompts in the chat and had to repeat yourself every time ChatGPT “forgot.” If this was your pain point, it’s worth checking out custom instructions.

There are two fields you can set:

1) “What would you like ChatGPT to know about you to provide better responses?” This is typically where you set a role, such as “I am a product manager” or “I am a 3rd grade English teacher.” We’ve used the instructions below when writing PRDs using ChatGPT.

2) “How would you like ChatGPT to respond?” This is where you can note what type of output you want to receive. If it’s text, you can suggest tone, length, or formatting you prefer. We’ve used the instructions below when writing PRDs using ChatGPT.

Before you go 💨 

Every company is becoming an AI company these days. Case in point: White Castle is planning on bringing AI to its drive-thrus. Soon, HaroldAIs and KumarAIs will be getting your slider orders. Maybe they will be better about packing you enough ketchup.