Things have once more gone pretty quiet around OpenAI and the much-anticipated voice chat features have been introduced rather more slowly than expected.
But there have been rumors of recent projects within the works, including SearchGPTwhich mixes generative AI and web browsing, and the more mysterious “Project Strawberry”.
The origins of Strawberry return to November 2023, when a model (more of a training technique) called Q* appeared in Leaks from Reuters.
There was even speculation that Q* was potentially dangerous and played a job within the hiring and firing of CEO Sam Altman last yr.
Q* was intended to mix a complicated reasoning model with an AI agent that may explore the Internet.
Despite dramatic headlines claiming that “OpenAI sits on an apocalyptically powerful model,” its legitimacy was highly contested on the time.
More details concerning the Q* project got here to light in May and June of this yr, after which it was renamed Project Strawberry, or just Strawberry.
According to ReutersStrawberry features a special method for training AI models to autonomously explore the Internet and conduct “thorough research.”
The Q probably refers to Q-learning, a long-established reinforcement learning (RL) technique. As for the star
there’s more uncertainty.
According to Reuters, it is analogous to a technique developed at Stanford called the “Self-Taught Reasoner,” or “STaR.” Others say it refers to a search algorithm called A*.
According to sources, OpenAI wants the model to conduct research by browsing the online autonomously, assisted by a “computational agent” (CUA), which can also be a key component of SearchGPT.
According to those sources, OpenAI wants Strawberry to perform “long-horizon tasks” (LHT), which require complex planning and execution over longer periods of time.
Stanford professor Noah Goodman, one in every of the founders of STaR, commented on the potential of this technology:
“I find this each exciting and terrifying… if things proceed on this direction, we as humans have some serious things to take into consideration.”
When asked about Strawberry, an OpenAI spokesperson gave a general statement concerning the company’s AI development goals:
“We want our AI models to see and understand the world more like we do. Continuous research into latest AI capabilities is common practice within the industry, with the expectation that these systems will improve their reasoning over time.”
Social media causes unrest
Not long after the Reuters report, in early August, Altman posted a photograph of strawberries with the caption “I like summer within the garden,” reigniting speculation concerning the Strawberry Project. I like summer within the garden
pic.twitter.com/Ter5Z5nFMc – Sam Altman (@sama)
7 August 2024
Then user iruletheworldmo, a kind of AI-focused meme/satire account (with a profile picture of Theodore Twombly, played by Joaquin Phoenix, from the AI movie “Her,” which is related to Altman), began posting strawberry-related content, hinting at a possible “stage two” breakthrough in AI.
The user posted: “Welcome to level two. How are you feeling? Did I make you’re feeling that way?” Altman, CEO of OpenAI, responded with “amazing to be honest.”
This exchange sparked a sequence response of strawberry posts and mass speculation on X and Reddit.
Welcome to level two.
how do you’re feeling?
Did I make you’re feeling? — 🍓🍓🍓 (@iruletheworldmo)
7 August 2024
Strawberry takes a brand new turn Only recently The revealed information
that OpenAI is preparing to launch a version of Strawberry as a part of a chatbot and possibly integrate it into ChatGPT as early as fall.
OpenAI can also be said to have demonstrated Strawberry’s capabilities to US security officials.
- Interestingly, in response to The Information, OpenAI is developing two different versions of Strawberry:
- This smaller, simplified version is meant for integration into chat-based applications resembling ChatGPT and goals to enhance reasoning skills in scenarios where users need more thoughtful, detailed responses moderately than quick reactions.
This larger, more powerful version might be used to generate high-quality “synthetic” training data for OpenAI’s next flagship language model, codenamed “Orion.”
Synthetic data generated by Strawberry could reduce the dependence on text and pictures collected from the Internet for training. This could potentially result in more accurate and reliable AI models and solve persistent problems resembling AI “hallucinations” orModel collapse
.
Strangely, nonetheless, these characterizations of Strawberry don’t match up well with earlier descriptions of Q*.
Perhaps we could speculate that Strawberry, the autonomous agent, surfs the Internet autonomously and uses its “extensive research” to synthesize synthetic data.
Maybe that is more computationally efficient and more useful for model training than simply scraping the raw data yourself??
AI doesn’t understand how many R’s are in strawberries
Here the story takes a bizarre and ironic turn.
“Strawberry” could be named after a word that current AI models, including among the most advanced, often have difficulty spelling accurately.
Ask an AI what number of “r’s” are in “strawberry,” and chances are high it can confidently answer “two” as an alternative of the proper “three.” Oh
pic.twitter.com/K2Lr9iVkjQ — Rob DenBleyker (@RobDenBleyker)
26 August 2024
Sounds ridiculous, right? I didn't consider it until I attempted it with Claude.
Some claim that this is a few type of Easter egg or joke inside OpenAI's systems, but provided that Claude reacts in the identical way as ChatGPT, that's unlikely, unless AI corporations are colluding behind the scenes to create area of interest strawberry jokes.
The explanation behind it’s elegant in its simplicity.
Language models, despite their name, are math-based systems. Text is translated into code, risking the lack of context and meaning on the word level. Why Strawberry reliably detects this can be a more complex query.
So it stays unclear whether OpenAI selected the name “Strawberry” as a playful reference to this common obstacle in AI or by pure coincidence.
It actually looks like something Altman could do, no matter whether Strawberry is real or not.
How this bizarre but berry interesting (…) strawberry story continues is anyone's guess. Frankly, at this point, I feel that not one of the speculative “evidence” we get from the main news outlets is fully representative of what is happening at OpenAI.
We may have to attend for SearchGPT and/or GPT-5 to see how OpenAI's products have evolved based on Strawberry and their other projects.
