Thoughts on GPT-4
Jonathan Kressaty
Groupthink is the place for your entire team to chat with a GPT-powered AI that reads the internet to provide personalized responses. Invite your friends and family, collaborate on projects, and make use of Groups and Threads to organize conversations.
Welcome new (and returning) users! For those who may be seeing this email for the first time, we are the founders of Groupthink and we send this daily email with our progress.
Apologies in advance for nerding out in this email, but we had some questions from the team today and wanted to share some thoughts!
Some Thoughts on GPT-4
Today, OpenAI announced GPT-4 – their newest large multimodal model. GPT-4 can not only understand text, but can also “see” images. OpenAI claims that GPT-4 is better at general reasoning, being instructed in chat-like settings, and understanding more nuanced instructions than the existing models.
For some quick background if you’re new to Groupthink – Groupthink is powered by OpenAI’s language models, including GPT-3 and GPT-3.5 (which is what ChatGPT is built on). GPT-4 is new, and is what Microsoft’s new Bing search bot uses.
One of the things we learned fairly quickly when building Groupthink is that the language model is not the bottleneck when it comes to building something useful. Creating a compelling product offering that’s useful to an audience is just difficult, and continuously improving the bot’s ability to give a unique response doesn’t really fix that. That said, when specific applications are in mind, updates to these “AI” services can be especially helpful.
As we’ve been building our features to support having better meetings, we’ve noticed that our current system has difficulty “understanding” the full meeting agenda. The numbering system (“let’s look at agenda item three”) is often confusing to the current language model. The system also struggles to understand how different agenda items are related to each other, especially when it’s contemplating the agenda combined with background information like a website or research from the internet.
I think GPT-4 can help address these problems. GPT-4’s ability to answer difficult questions that require complex reasoning demonstrates a better understanding of the way humans write and think. I’m also excited to see the improvements the model has made in understanding mathematics – it appears to have nearly doubled in terms of factual accuracy when it comes to solving math problems, and that’s hugely important in any business context.
We’re excited to integrate GPT-4 into Groupthink as it becomes generally available.
Happy Pi Day!