The nickname “Strawberry” for OpenAI’s new model is an inside joke, but no one told ChatGPT. Here’s the punchline: “Finally, ChatGPT knows how many R’s are in the word Strawberry.”
OpenAI, the company behind ChatGPT, came out with their newest model last week called o1. It was nicknamed Strawberry. This is a different kind of AI model than its predecessors. Instead of just generating text based on what came before it, o1‘s goal is to come up with the right answer to a question.
This has allowed the model to become much “smarter” than previous models. It does this through a system called “chain of thought.” Essentially, the model has been trained to think step by step. Instead of just answering right away, it breaks down the process, solving each part in sequence. This method mirrors how people tackle complex problems: by taking things one piece at a time. The result? Smarter, more accurate answers—kind of like a student who can try several different things and quickly learn from their mistakes.
o1 has become quite a good student. It now gets 95% of the LSAT questions and 83% of AP Calculus questions correct. And most importantly, it can tell you how many R’s are in the word “Strawberry.”
How Many R’s are in the Word Strawberry
Why do we care how many R’s are in the word “Strawberry?” Because this is one of the most popular ChatGPT memes on social media. It’s such an easy task for us but ChatGPT fails at it constantly! Here are some examples of posts.
In this one, it highlights 3 but tells you that there are only 2.
In this one, the tool quickly corrects itself when it’s told that it’s wrong.
This is because ChatGPT doesn’t look at the word “Strawberry” as a series of individual characters, like a human would, but as parts of the word, like “straw” and “berry.”1 This is the way that ChatGPT understands the meaning of words. This lets it do some amazing things. However, as a result, it might make silly mistakes like missing the exact count of specific letters.
OpenAI clearly knows about this. It’s even in the suggested prompts for o1:
ChatGPT’s Response
But no one told ChatGPT about why its new model had this nickname. When I asked ChatGPT why o1 was nicknamed Strawberry, here’s what it said.
Me: Where does the name “Strawberry” come from as a code name for
o1?ChatGPT-
4o: In tech development, code names like “Strawberry” often follow a pattern of using simple, familiar, or whimsical words that make internal projects easy to refer to without tying them directly to their technical nature or final public identity. These code names, which can include fruits, colors, animals, or everyday objects, are meant to serve as placeholders that are easily recognizable and memorable during the development phase.
I get it, many companies have code names for things. Some of my favorites are Android2 (alphabetical candy) and MacOS3 (big cats and California landmarks). But not OpenAI. The closest that they came to clever names were the models Ada, Babbage, and Darwin. Since then, models have names like GPT-4 and GPT-4o.
But when I asked it about the meme, it said that there was no connection.
Me: Is the name “Strawberry” as the code name for
o1related to the meme, “How many R’s are in the word strawberry?”ChatGPT-
4o: While the three “R”s theory you mentioned is an interesting one, there’s no official confirmation linking the choice of “Strawberry” to that concept. Instead, it’s likely part of an informal, fun naming pattern that gives engineers and teams an easy, neutral, and memorable way to refer to an advanced but yet-to-be-released technology during development.
However, it changed its mind when it found the following article that talks about the meme as part of the name.
Me: Are you sure it’s not related to the meme? Look it up.
In essence, the meme represents a simple but telling example of the model’s ability to handle basic reasoning tasks—tasks that were previously problematic for earlier models
It took ChatGPT–4o a while to get to the right answer. I have no doubt that o1 would have come up with the answer more quickly. But right now it can’t connect to the web and its knowledge cutoff is 2023.
Footnotes
- The OpenAI tokenizer shows that in GPT-
4oStrawberry breaks down to Str(3504) aw(1134) and berry(19772) and strawberry breaks down into st(302) raw(1608) and berry(19772). ↩︎ - Android Code Names (alphabetical candy)
Android 1.5 – Cupcake
Android 1.6 – Donut
Android 2.0–2.1 – Eclair
Android 2.2–2.2.3 – Froyo (Frozen Yogurt)
Android 2.3–2.3.7 – Gingerbread
Android 3.0–3.2.6 – Honeycomb (Tablet-specific version)
Android 4.0–4.0.4 – Ice Cream Sandwich
Android 4.1–4.3.1 – Jelly Bean
Android 4.4–4.4.4 – KitKat
Android 5.0–5.1.1 – Lollipop
Android 6.0–6.0.1 – Marshmallow
Android 7.0–7.1.2 – Nougat
Android 8.0–8.1 – Oreo
Android 9.0 – Pie ↩︎ - macOS Code Names (big cats and California landmarks)
macOS 10.0 – Cheetah
macOS 10.1 – Puma
macOS 10.2 – Jaguar
macOS 10.3 – Panther
macOS 10.4 – Tiger
macOS 10.5 – Leopard
macOS 10.6 – Snow Leopard
macOS 10.7 – Lion
macOS 10.8 – Mountain Lion
macOS 10.9 – Mavericks (Transition to California landmarks)
macOS 10.10 – Yosemite
macOS 10.11 – El Capitan
macOS 10.12 – Sierra
macOS 10.13 – High Sierra
macOS 10.14 – Mojave
macOS 10.15 – Catalina
macOS 11 – Big Sur
macOS 12 – Monterey
macOS 13 – Ventura
macOS 14 – Sonoma ↩︎
You must be logged in to post a comment.