GPT-4 is Out - A Roundup of Writeups

From some of my favorite authors

Generated with Midjourney V5

OpenAI has (partially) released GPT-4. You can get access by paying for ChatGPT Pro (although ChatGPT Pro is currently limited to text, while the full GPT-4 can deal with more mediums than text), or you can join the waitlist for the API.

There are a lot of really good writeups on what’s in GPT-4, so rather than write another one, I’m going to link to the great work that other “first on the scene” authors have done.

I recommend subscribing to all of these authors. I read everything they write.

Alberto Romero

at the Algorithmic Bridge writes GPT-4 in 10 Keys. Some highlights:

• Multi-modal: Edit an image based on a prompt. Build a web site off a napkin sketch (not publicly available yet)

Rowan Cheung @rowancheung

I just watched GPT-4 turn a hand-drawn sketch into a functional website. This is insane.

• How to get access: Buy ChatGPT Pro or put your name on the API waitlist

• Enlarged context window: You can paste in more text and get longer outputs.

Linus Ekenstam

at Inside My Head writes GPT-4 Just Released. Some highlights:

• Better reasoning. It can figure out what you’re talking about better than GPT-3.5

• The OpenAI Developer Livestream (now available here)

Linus also has a great Twitter thread on some amazing scenarios GPT-4 is being used for, including drug discovery and 1 click lawsuits for people receiving robocalls.

Josh Dance

at Mythical AI writes Mini Episode - GPT4 was just released, here's what you need to know. Some highlights:

• In internal evaluations, GPT-4 is 82% less likely to respond to requests for disallowed content, and 40% more likely to produce accurate, factual responses than GPT-3.5.

• GPT-4 still has many known limitations. Open AI acknowledges that social biases, hallucinations, and adversarial prompts are all still a problem.

Anthony Castrio

at Bot Eats Brain writes GPT-4 released, here's what it can do. Highlights:

• It does better on tests (AP, SAT, LSAT) than GPT-3

David Rozado

at Rozado’s Visual Analytics digs into The Political Biases of GPT-4.

• On the surface, it comes out completely neutral based on the political compass test.

• However, when jailbroken, it shows to be Left / Libertarian.

Gary Marcus

at The Road to AI We Can Trust throws some cold water on GPT-4 in GPT-4’s successes, and GPT-4’s failures. Writing, “It doesn’t actually solve any of the core problems of truthfulness and reliability that I laid out in my infamous March 2022 essay Deep Learning is Hitting a Wall. Alignment is still shaky; you still wouldn’t be able to use it reliably to guide robots or scientific discovery, the kinds of things that made me excited about A(G)I in the first place. Outliers remain a problem, too.”