Weekend links: Comments on OpenAI's DevDay, AI hardware, and a harmful domain

Almost an Agent: What GPTs can do

Ethan Mollick frames the “GPTs” OpenAI launched at the company’s DevDay as the best way yet to share prompts. I think that’s a really good way to think about them. Removing the prompt engineering from people who are not interested in that craft, but still letting them leverage the power of generative AI.

The OpenAI Keynote

Ben Thompson's thoughts on OpenAi’s DevDay.

With its New M3 Chips, Apple joins the AI party.

On the user side, AI has so far been about the software. The hardware aspect has been more apparent on the model building end of AI.

This is probably about to change. There seems to be a wave of end user products with AI features built in around the corner. Om Malik’s comments on Apple’s M3 chips are one example of this, where new hardware gives consumer products general AI capabilities. Products like the Rayban/Meta glasses are also examples of this trend, but in more purpose made hardware.

Working with AI: Two paths to prompting - by Ethan Mollick

Ethan Mollick categories LLM prompts into conversational and structured, where the latter are prompts with a clear outcome and are created to be reusable. There are a couple of principles for the structured ones in his post, but the most important takeaway is this: Start experimenting to get a feel for what works for your use cases.

De tar fram en tillförlitlig ChatGPT för europeiska språk

European research project TrustLLM aims at trustworthy language models in Swedish, German, and a couple of other languages. Correctness, transparency, energy are some of the core aspects – together with openness. (In Swedish.)

Telling GPT-4 you're scared or under pressure improves performance

Prompts that stresses that the task is important to the user gives better results from LLMs.

ChipNeMo: Domain-Adapted LLMs for Chip Design

Nvidia fine-tuned Llama to build a custom LLM that excels at tasks related to chip design. I guess we will see a lot more like this in the months to come. The general purpose language models are a good start, but adding proprietary information will give an extra edge.

.io considered harmful

An .io domain might not be what you want in your domain portfolio.