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

• OpenAI Launches Web Crawler with Some Controls

• Competition in Chipmaking

• AI Use Cases: Education and Rent Collection

• How to Assess the Right LLM for You

• …and more!

News of the week 🗞️ 

🕸️ OpenAI Launches Web Crawler with Some Controls

OpenAI launched GPTBot, a new web crawler that will scan website content to train future AI models. Notably, OpenAI has incorporated enhanced controls, ensuring GPTBot filters out sources behind paywalls, those gathering personally identifiable information, and those contrary to OpenAI’s policies. Website owners can also restrict GPTBot by adding some code to the robots.txt file or blocking select IP addresses.

A step towards the right direction…From a data sharing and privacy perspective, this update empowers website owners to safeguard their data from OpenAI. If you own a website, you should consider the implications and act accordingly. How the opt-out approach will impact OpenAI models' quality in the future remains to be seen.

⚒️ Competition is Heating Up among Chipmakers

Nvidia, the global chip production leader, continues to push boundaries. Its latest chip, the Grace Hopper Superchip (GH200), is claimed to have the same processing power as its most powerful chip model but triple the memory capacity.

A few companies are stepping up to the challenge. Modular, a startup that streamlines the process of training and implementing ML models, is in discussions to raise Series A funding at a $600M valuation. Modular's strategy is to build software that is compatible with multiple chip providers (likes of AMD, Intel, and Google), in contrast with Nvidia's software which is only compatible with its own chips.

In a gold rush, sell picks and shovels…You don't need to strike a gold mine to make it in a gold rush. Selling picks and shovels could be a savvy business move, especially as miners are flocking to town. Let's hope that such development will lead to a virtuous cycle where AI builders can gain cheaper access to better tools.

If you want to learn more about the building blocks of generative AI industry, take a read here.

🎒 Real-life AI Applications: Education and Rent Collection

Two noteworthy companies are making waves with real-life AI applications.

Uplimit: Founded by former Coursera employees, Uplimit is building a platform for professional education. It uses generative AI to amplify human teachers' reach, as opposed to replacing them altogether. They aim to facilitate better learning and teaching experiences for students and teachers alike.

Colleen AI: An Israeli startup, Colleen AI aims to facilitate better rent collection processes. It claims to have reduced unpaid rent by 30% and it just raised $3.5M in seed funding.

How would you apply AI?…Sometimes it's helpful to examine concrete examples to brainstorm on how you might apply AI. Did these examples spark any ideas for you? If so, share your ideas with us—we are happy to jam!

AI product highlight ✋ 

Do you want an AI assistant that just…works?

If so, Impel may scratch your itch. Impel is claiming to be an assistant on your Mac capable of diverse tasks, such as keeping a to-do list, recording and transcribing meetings, and even planning your next travel.

Disclaimer: We're not going to lie—we haven't tried it yet because it feels a bit creepy. But wanted to share for the brave souls out there.

For the AI nerds 🤓 

Assessing LLMs: Choosing Wisely

The recent launch of Claude Instant 1.2 got us thinking a bit more about how to evaluate and choose LLMs for different use-cases.

1. Output quality in your domain

Different LLMs will be better in different domains. It's important to consider what you will be asking the LLM to do and which one will be best for your specific scenario, especially if you're trying to save on cost by avoiding using the biggest models.

2. Cost

Often you will have a trade-off between higher upfront costs for lower ongoing costs. For example, a lot of research has been coming out that indicates you can create smaller and cheaper fine-tuned models that rival the performance of the biggest models when targeted at specific domains. However, the process of fine-tuning, or even prompt engineering, can add significant cost to your development.

3. Hosting

You can self-host your own model (using open source models) or you can use the APIs provided by other providers. This choice can obviously affect cost, but it can also have implications for data privacy, reliability, and more.

4. Context window

A model’s context window refers to how much content it can consider at one time. When models cannot process your entire input all at once, you will have to spend more engineering time to create work-arounds.

These are not the only considerations but this is a good list to get started with!

Before you go 💨 

🤳 Have you not been getting enough matches on Tinder? Tinder is trying to help solve that by using AI to select the best looking photos from your profile. If this works out, make sure to invite AI to your wedding!