ChatGPT has just dropped OpenAI Strawberry. They say it’s a SIGNIFICANT step up in reasoning and output. Should you believe the hype? Check out the info we’ve rounded up and decide for yourself!
Table of Contents
What is OpenAI Strawberry?
OpenAI Strawberry is the newest large language model for ChatGPT as of September 2024. You can already select it as an option on ChatGPT right now, and it will also power the not-yet-released ChatGPT 5.0. Strawberry comes in three different versions: o1, o1-preview, and o1-mini.
o1, o1-preview, o1-mini... o my! What is the deal with these three confusingly similar sounding versions? Let me break it down for you.
What is ChatGPT o1?
ChatGPT o1 is the most advanced version of ChatGPT right now. Currently, it's only available to OpenAI’s developers. If the o1s are the family, think of this one as the wise parent who stays in the house to do business.
What is ChatGPT o1-preview?
ChatGPT o1-preview is a more basic version of the full ChatGPT o1. It’s available to paid ChatGPT subscribers, and can be accessed on some free accounts. Think of it as the older, more mature sibling of the family... the one that thinks carefully and acts deliberately.
What is ChatGPT o1-mini?
ChatGPT o1-mini is also available to the public. True to its name, o1-mini is like the younger sibling of the Strawberry family. It doesn't know as much as its older sibling o1-preview, because o1-mini is trained on far less factual data. This makes mini "younger" and more spry-- it can answer questions and go thorugh logical steps more quickly without that extra knowledge slowing it down. But it's less capable when it comes ot fact-based tasks.
Roundup of Official Open AI Strawberry Announcements
OpenAI has rolled out a number of key resources on its website to help you get to know Strawberry a little better. Below is a link roundup for their top resources. I’ve provided a brief description of each link to help you decide which links will be most useful to you.
OpenAI o1 Hub: This page links out to resources on the development, , use and safety features of OpenAI Strawberry (AKA OpenAI o1). On-page, the hub opens up with a 20 minute video discussion of o1’s development that’s feel-good but low on info. There are also short video demos of different OpenAI Strawberry use cases and prompts.
Learning to Reason with LLMs: This article shows internal research indicating ChatGPT Strawberry outperforms ChatGPT 4o in STEM, and in other professional fields such as public relations, economics, coding, and encryption. Includes side-by-side threads showing 4o and o1 take on the same decryption task.
OpenAI o1-mini introduction and report: Research indicating that o1-mini is faster than the other version of o1, performs well with logical reasoning, but may do more poorly on fact-based tasks due to its more limited access to data.
OpenAI o1-preview Introduction: A brief description of ChatGPT o1-preview. Importantly, this page is regularly updated to give the current usage limits for preview.
Can we believe OpenAI?
Yes and no. OpenAI is unlikely to lie outright– they have a track record of honesty, and they know that Strawberry is highly scrutinized and readily available to the public. But they may be putting their thumb on the scales just a little. Certainly, they have the means and motivation to test their AI under the most optimal conditions.
In the next sections, we’ll explore what news organizations and individual users are saying as they explore OpenAI Strawberry and at times provide reality checks on OpenAI’s claims and promises.
Open AI Strawberry: News Highlights
OpenAI Hardwires ChatGPT o1-preview to Discourage Transparency
Wired and Futurism report that ChatGPT is threatening to ban users who ask Strawberry about its own reasoning. Chris Lele, Elevate AI Coaching’s own CEO, personally encountered a warning for this and caught it on video (see the second video in the “Experiments” section below).
Is ChatGPT o1 Free? Sometimes!
This Tech Radar article highlights the good news: OpenAI is running an A/B test where some free users can choose o1-preview from the dropdown menu.
Only some free users are seeing o1 as available, but there are still some interesting implications here: so far, anytime a version of ChatGPT has been made free to select users, it becomes free for everyone shortly thereafter. This is the first time OpenAI has made a brand-new version free to select users, so that’s even more intriguing.
OpenAI o1 is Also Available Through Perplexity and You.com
According to ZDNet, OpenAI o1, AKA OpenAI Strawberry, is now helping to power the Pro (subscription-based) version of Perplexity AI, an AI platform designed to help with research . OpenAI Strawberry can also be accessed through the paid version of You.Com, a service that bundles multiple AI platforms into a single app.
What Can OpeanAI Strawberry Do: Some Illuminating User Experiments
Finally, we get to the good stuff! Let’s take a look at how real users have found OpenAI Strawberry works in practice.
Partnering with OpenAI Strawberry to Create Python Code
I found a great Medium article by Lucas Soares highlighting his experimentation with ChatGPT o1’s coding abilities; the contents of the article are also summed up in an accompanying YouTube video.
I love this experiment because author Lucas Soares is engaging in many of the powerful “thought partnership” techniques that Elevate AI Coaching recommends to its clients.
Mr. Soares is also very even-handed. He points out that ChatGPT o1 doesn’t know everything– he had to give it some updated info on Python in order to get its help (bringing your own expert knowledge into the thread is definitely something we recommend!).
What this experiment lacks, however, is a side-by-side comparison to ChatGPT 01’s immediate predecessor, ChatGPT 4o. Fortunately, the next two experiments I’m going to show you offer great side-by-side comparison.
ChatGPT o1-preview vs ChatGPT 4o: Which is better at math?
Over on Skill Leap AI’s YouTube channel, the host of the video did something really interesting– he looked at OpenAI’s own description of how ChatGPT o1 uses chain of thought and logic, and he instructed ChatGPT 4o to take that exact same approach! He then had 4o and o1 go head to head on some challenging math problems.
Interesting takeaways from the video:
When ChatGPT 4o was told to behave like ChatGPT o1, it performed similarly to it.
Contrary to the results of OpenAI’s own internal research, ChatGPT Strawberry (AKA ChatGPT o1) has not made great leaps forward in math. OpenAI o1-preview made a number of mistakes here, and largely missed the same questions that 4o missed!
Who is a Better Thought Partner: ChatGPT o1 or ChatGPT 4o?
Elevate AI Coaching CEO Chris Lele used ChatGPT o1 and ChatGPT 4o side-by-side for thought partnership. Ultimately, 4o performed better as a thought partner -- it provided better, more insightful thoughts, and was more responsive to input.
In the process, Chris got an onscreen warning, saying he was probing too deeply into the way o1’s logic works. Recall from our news roundup that users can get banned for that! Thankfully, Chris didn’t.
Watch the video for yourself below.
Final Thoughts on OpenAI Strawberry (for now):
OpenAI Strawberry goes by many names, and has many potential uses. Currently though, this new model does not seem to live up to the hype its creators have given it.
However, in all the reporting and experimenting, there is one hidden variable: how fast OpenAI Strawberry will be able to learn. Remember, large language models learn and grow in the course of their interaction with users. It’s possible that ChatGPT o1 could improve rapidly as more and more people continue to train it by using it. (In other words, the secret ingredient to any large language model's success is... you!)
Given the minimal transparency from OpenAI, it’s also quite plausible that ChatGPT Strawberry has some “hidden talents” that aren’t widely known just yet. As always, the limit to how a given large language model might be used is the human imagination. So I would encourage all of you to dive in and test the limits of this newest iteration of ChatGPT for yourselves.As I write this, ChatGPT o1-preview has been available to me for less than two weeks, and the full o1 model isn’t available to me just yet. The team and I at Elevate AI Coaching are definitely going to keep playing with it. Watch this blog– we’ll give you more write-ups and videos and we discover new things! BACK TO TOP
Comentarios