OpenAI's New Model... A PhD In Your Pocket?

We’re back! The AI For Humans newsletter has returned (for real this time) and we’ll be coming to you every Tuesday at 10am PST or so. Our focus is to make a weekly newsletter that’s super valuable to you and to call out some of the most creative and interesting use cases of AI we see each week.

This week, the big news out of OpenAI is the release of their new o1-preview and o1-mini models which have layered on reasoning to the already robust GPT-4o frame work. If you’ve seen us talking about the ‘strawberry’ model, this is it and if you’re interested in learning all about it check out the video that Kevin & I made show exactly what it’s capable of.

So far, we’ve been pretty impressed by what it can do by taking just a little bit more time to think about a task — you’ll actually see the model say ‘thinking’ when you ask it to do something. It’s particularly good at math, science and logic questions, much more so than creative tasks.

A lot of people have noted the o1-mini is particularly good at coding and there’s been lots of examples of it getting answers in a single shot prompt right vs returning broken code.

The real underlying story here is that scaling isn’t just for the main model training but also for inference. Gavin, you might ask, what the heck does that mean?

Well, let’s ask ChatGPT (GPT-4o in this instance because it can search the web):

Of course, this isn’t the full story. Is it ever with OpenAI?

The next big thing is a new frontier model that has been rumored to be coming at some point either later this year or early next, currently codenamed “Project Orion”.

Sam Altman, OpenAI’s CEO, even hinted at it in a cryptic tweet about the night sky in the mid-west. The man certainly understands how to get the hype train rolling.

We’ll be keeping our eye on the night sky as well as on the next big AI tools and news just for you.

Yours,

Gavin (and Kevin)

Community Spotlight

It’s hard to explain but there’s a lot happening in the AI4H Discord and Patreon

3 Things To Know:

OpenAI Launched A New o1 Family of Models
Here’s the full blog post from OpenAI. In your ChatGPT model selector dropdown, all paid users should now see o1-preview and o1-mini, then very quickly reach their rate limits. OpenAI will let you have just 50 messages per week with 01-preview, and 50 per day with o1-mini.

RunwayML Opened Up Access to Gen-3 Video-To-Video
Upload a video & restyle it with a text prompt. It still feels like a bit of a dice roll, and some results may terrify you, but it’s an amazing resource to repurpose traditional or AI footage.

HeyGen Launched Avatar 3.0
The most convincing AI talking head videos just got more convincing. HeyGen dropped v3 of their avatar generation platform, promising “full-body dynamic motion” and voice tones that better match the tone of the script. It’s available now, even on the free trial.

We Love This

Adobe Firefly Structure Reference
The new Runway model has everyone remixing their videos this week, but what if you want to restyle an image? Adobe Firefly Structure Reference is your friend.

From the web UI, simply upload a reference image to the Composition panel and crank the strength all the way up.