Happy August! Welcome to the second edition of the odo newsletter. We’re here to provide a free, weekly newsletter on key AI developments for product builders. We’ll cover news from the past week, AI product highlights, and any resources we think you’ll find useful.

News of the week 🗞️ 

💊 Amazon Released Generative AI Product for Healthcare

Last Wednesday, Amazon announced AWS HealthScribe, which is an API that helps clinicians transcribe and analyze conversations with patients. HealthScribe aims to improve the quality and efficiency of medical charting, which is a time-consuming and manual process of documenting details of each patient visit. Healthcare software vendors such as 3M, ScribeEMR, and Babylon have already been leveraging this technology for their applications.

Less time writing, more time doctor-ing?...Studies have shown that doctors are spending over 4 hours a day on completing electronic health records. If doctors can use AI-based tools to complete their grunt work faster, hopefully it will free up their time to deliver better care for patients.

🖼️ Stability AI Announces Stable Diffusion XL 1.0

Stability AI, a leading open-source AI company, has announced its latest text-to-image suite of models called Stable Diffusion XL (SDXL) 1.0 –yes, AI companies are not getting any better with naming. Stability AI claims that the newest model will generate more vibrant and accurate colors, with better contrast, lighting, and shadows than the previous model. In addition, SDXL 1.0 has improved text generation capability, which is something that text-to-image models have struggled with in the past.

Not just an art project…These improvements are a huge step towards being able to leverage AI technology in context beyond just art, such as generating technical diagrams or low-fi mock-ups.

❌ OpenAI Shut Down a Product that Detects AI-generated Text

OpenAI has shut down AI classifier, a product that was supposed to distinguish AI-generated text from human-generated text, due to a low accuracy rate. According to OpenAI’s blog, the classifier correctly identified only 26% of AI-generated texts, while misidentifying 9% of human-written texts as AI-generated.

Spiderman who? Spiderman you...Not being able to distinguish whether human or AI generated an output could have serious implications in sectors such as education, arts, and entertainment. It’s not a huge vote of confidence to see OpenAI shutting down the product temporarily.

AI product highlight ✋ 

Are you someone who practices the art of Astrology but also likes new technologies? We’ve got a product for you. Meet Aistro, a personalized AI astrology powered by LLM (yes, this exists).

We see you rolling your eyes. But, when we asked for our daily motto, it said: Stay grounded and embrace change. It may be challenging, but it will lead to growth and new opportunities. Trust in yourself and be open to new experiences. 

It basically told us to start a company, so we’re on board with it.

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

For the AI nerds 🤓 

Using LLMs to Structure Data for More Traditional Processing

Background

A lot of effort has been going into getting LLMs to think more thoroughly, especially through Chain of Thought prompting as we discussed last week. However, even the most powerful models are still mediocre at reasoning. A couple recent papers we found inspired us to think a bit differently about how LLMs could be applied in more targeted ways.

What’s the Big Idea?

LLMs are particularly good at connecting similar words or ideas. They have a strong sense that “tiger” and “animal” are strongly associated while “tiger” and “tea kettle” are a lot less so. The big idea behind these papers was to specifically leverage what LLMs are best at while manually programming the reasoning for the final outcome.

This works by using the LLM to convert freeform text, which has a virtually unlimited vocabulary, into something with a much more limited vocabulary. It then becomes feasible to manually reason or program around that vocabulary.

Why Should We Care?

There are a couple of reasons these techniques could be relevant to us product builders:

1. Competitive analysis – We can leverage LLMs to automate tasks such as competitive analysis. LLMs are extremely good at parsing through large amounts of text and translating them into an output that is easily digestible.

2. Efficiency/Cost Savings – Techniques like this can be used in place of the super large models for reasoning tasks. Those models are slow and expensive.

3. Traceability – In this method we can actually answer why a decision was made and improve either the reasoning or the parsing independently (and even see when the parsing is wrong).

Before you go 💨 

It’s all fun and games until you make a major AI breakthrough. Researchers created an AI Agent to perform tasks in Minecraft. The agent plays Minecraft just like a person would (watching the screen and using the mouse and keyboard). That means many of the techniques can be applied to pretty much any task on a computer.