ChatGPT is the generative AI chatbot I use the most, and I also happen to pay for the Plus subscription, even though the free tier would give me access to most of the chatbot’s features. I also happen to use GPT-4o, OpenAI’s latest and best ChatGPT model. But I can’t say I’m fully satisfied with what GPT-4o can do for me, and I can’t wait for OpenAI to fix it.
What I hate most about GPT-4o is the speed at which it produces responses. These responses can be wrong and need correction. I wish it could slow down a little and reason.
When I tell the chatbot it made a mistake or a few mistakes, it doesn’t always correct them immediately. That is, the corrections continue to be wrong. Getting the right information out of it might take a few nudges. And it won’t always perform internet searches when I tell it the information isn’t accurate.
I’m trying to be persistent with it, as it teaches me how to interact with AI chatbots better while I wait for OpenAI to improve GPT-4o. Now that I have heard of the purported Strawberry upgrade, I can’t wait for OpenAI to deploy it for ChatGPT. It can’t remove all hallucinations, but if the rumors are right, Strawberry might give ChatGPT the ability to reason.
The Strawberry codename comes from a Reuters report based on a purported OpenAI internal document dating back to at least May.
According to a source, Strawberry is a work in progress that’s being kept a secret even within OpenAI. The Strawberry models would let ChatGPT plan ahead and navigate the internet autonomously and reliably for “deep research” operations.
That’s music to this ChatGPT user’s ears. I want “Deep research” from my ChatGPT interactions, especially on matters that might require multiple, complex internet searches.
According to Reuters, Strawberry is the evolution of ChatGPT Q*, an internal codename that appeared online around the time Sam Altman was fired. This was described as a major AI breakthrough at the time.
The report notes that two sources saw the Q* version of ChatGPT in action earlier this year. The model answered “tricky science and math questions” that are “out of reach” for the current chatbots. A different source claims that OpenAI has tested AI that scored over 90% on a MATH dataset, but it’s unclear if this is the same Strawberry project.
Seemingly confirming the Reuters report, a story from Bloomberg detailed an internal all-hands meeting at OpenAI this week where the company demoed a project with “human-like reasoning.” It’s unclear if this was Strawberry.
OpenAI also gave employees a new classification system for AI models. Level 1 is the kind of AI you can now experience via the likes of ChatGPT. Level 5 would be AI capable of performing the work of an entire organization.
Interestingly, Level 3 would be AI that can spend several days taking action on a user’s behalf. Level 4 would be AI that can innovate on its own.
The Strawberry codename does not appear in Bloomberg’s report. But that Level 3 above does seem to appear in the Strawberry documentation that Reuters has seen:
Among the capabilities OpenAI is aiming Strawberry at is performing long-horizon tasks (LHT), the document says, referring to complex tasks that require a model to plan ahead and perform a series of actions over an extended period of time, the first source explained.
OpenAI is training these models with a mysterious “deep research” dataset. The result would be a ChatGPT version that could research the web autonomously via computer-using agents (CUA), which can then act based on the findings.
OpenAI wants to test these models on the work of software and machine learning engineers. That might actually lead to AI training and developing future variants of AI on the path to AGI or artificial general intelligence. That’s speculation from yours truly at this point. But, as the report notes, AGI is the holy grail of current AI research.
Reuters also describes how Strawberry is being trained. It’s not just about ingurgitating as much data as possible. OpenAI is also using a post-training method to improve the AI. The technique uses “fine-tuning,” or a process where humans give feedback to the AI.
Furthermore, Strawberry is similar to a method developed at Stanford in 2022. It’s called “Self-Taught Reasoner” or STaR. This allows AI to train itself into higher intelligence by creating its own data.
When will Strawberry be ready for mass consumption? Reuters says that OpenAI has been telling developers and other parties in recent months that it’s “on the cusp” of releasing AI with improved reasoning abilities. While it’s unclear whether we’ll get any ChatGPT upgrades this year, the AI industry is moving at a breakneck pace. OpenAI is forced to deploy new innovations as soon as possible.
I wouldn’t be surprised if Strawberry became ChatGPT 5, but I’m speculating again. I’m also ready to keep correcting GPT-4o while I wait for the next upgrade to drop.
\