ChatGPT will enter the "Strawberry" era in two weeks? Three hours before the news broke, another chief researcher of OpenAI "left"

avatar
36kr
09-12
This article is machine translated
Show original

Is the legendary AI model Strawberry about to be launched?

On September 10, local time, foreign media quoted two people familiar with the matter who had tested the Strawberry model as saying that OpenAI plans to release Strawberry as part of the ChatGPT service in the next two weeks. This is earlier than the "autumn release" (September to November) rumored two weeks ago.

The pricing of this latest model has also aroused interest from the outside world. It was previously reported that the subscription fee for the new model will be as high as $2,000/month, but some people in China claim that ChatGPT Pro has appeared for $200/month, and speculate that this may be paving the way for the release of Strawberry. However, the authenticity of this news cannot be confirmed.

In addition, the latest report from foreign media said that the initial version of Strawberry can only process text, which may be a hasty decision made by OpenAI under increasing competitive pressure. It is worth noting that three hours before this news was exposed, Alexis Conneau, the head of GPT-4o and GPT-5 research at OpenAI, announced on the X platform that he would leave OpenAI to start his own business.

01 Strawberries will be harvested within two weeks?

At the end of August this year, foreign media quoted people familiar with the matter as saying that OpenAI's Strawberry model is expected to be launched as part of a chatbot as early as this fall. But the latest news shows that OpenAI plans to launch it as part of ChatGPT in the next two weeks, and the release process is accelerating. Existing subscribers to ChatGPT may have the opportunity to use Strawberry before free users.

According to foreign media reports, although Strawberry will be integrated into ChatGPT, it will be an independent option. It is not clear how users can access it, but there are speculations that users only need to select Strawberry from the list of available models in the ChatGPT interface.

People familiar with the matter revealed that OpenAI began focusing on the development of Strawberry last month, and said that the difference between Strawberry and other generative AI models is that it focuses more on reasoning: unlike the almost instant response of existing models, the Strawberry model usually takes 10 to 20 seconds to consider the response to any question or prompt.

According to their explanation, Strawberry's extra thinking helps avoid mistakes and makes it more likely that it will know when it needs to ask customers follow-up questions so it can fully answer their questions. This purposeful delay allows Strawberry to process each query in a more methodical way, reducing the chance of errors. By taking the time to think, OpenAI hopes that Strawberry will be able to provide better response quality. This means that Strawberry will not only perform better on math problems and coding, but also better at completing more subjective business tasks, such as developing product marketing strategies.

02 How will strawberries be priced?

Foreign media reports also said that the pricing model of strawberries may also be different.

Unlike GPT's existing free and subscription models, Strawberry's pricing structure is expected to limit the number of messages a user can send per hour. Moreover, pricing may be tied to response speed. For those who want faster responses, Strawberry may also offer higher-priced packages.

Previous reports revealed that in OpenAI's early internal discussions, a monthly subscription price of up to $2,000 for the new model was on the table, but no final decision was made. However, most people do not believe the authenticity of this news and only regard it as a marketing hype by OpenAI. Some netizens commented that OpenAI leaked this price in order to make people feel that they have taken advantage of it when they later launched the price of $200.

Coincidentally, someone in China claimed to have seen the $200 ChatGPT Pro and tried it, but there are currently no new features or new models. The only difference is that the number of uses of GPT-4o is basically unlimited. The person speculated that ChatGPT Pro members are preparing for the release of Strawberry in advance. But it is worth mentioning that the only proof picture given by the source is a payment page, but from the comparison of the following figure, it can be found that the validity period of the bank card is not displayed on the payment management page of OpenAI.

Image source: Digital Life Kha'Zix (WeChat official account)

Image source: OpenAI's payment management page

However, some foreign media also stated that OpenAI may privately select some people to join the ChatGPT Pro subscription list, so the authenticity of the news cannot be determined.

Putting aside pricing for now, for users, high subscription fees should be matched with a good user experience.

But some people who have used the Strawberry prototype have complained that the quality of Strawberry’s responses is slightly better than OpenAI’s currently released GPT-4o, but not worth the extra 10 to 20 seconds of waiting, according to people familiar with the matter.

And, while Strawberry’s reasoning capabilities are hailed as advanced, sources said the new model still has some limitations.

Some early testers reported that while Strawberry was theoretically able to skip thinking steps when users asked simple questions, it didn’t always do so in practice: It might mistakenly identify a simple question as one that took too long to think about, while other OpenAI models could respond in a fraction of the time.

In addition, although Strawberry can remember previous conversations in order to provide more personalized responses, in actual experience, it sometimes suddenly forgets.

Not only that, foreign media reported that the initial version of Strawberry is expected to be released can only receive and generate text, but not images, which means it does not support multimodality. Most of the large language models released today are multimodal.

03 Key figures behind GPT-4o and GPT-5 leave

Foreign media analysis said that releasing Strawberry in plain text ahead of time may be a hasty decision made by OpenAI under increasing competitive pressure. Last month, Google launched an AI voice assistant Gemini Live. This new tool is flexible enough to cope with user interruptions and sudden topic changes.

In addition to pressure from rivals, OpenAI's staff turnover seems to be continuing.

Three hours before the latest news about the release of Strawberry came out, Alexis Conneau, OpenAI’s chief researcher and a key figure behind GPT-4o, announced his departure from OpenAI.

“After the amazing journey building #Her at @OpenAI, I’ve decided to start a new company,” Conneau posted on X. He also invited those interested in “building something amazing” to contact him, and the new company is hiring.

Image source: X Platform

In the acknowledgments list of GPT-4o released by OpenAI, it can be seen that Conneau specifically led the research work of the GPT-4o Audio Pre-Training, Encoders, and Interruptions teams. The research content of the Interruptions team is the key to allowing GPT-4o to be interrupted like a human.

When GPT-4o was released in May, it won the industry's favor with its "all-round" functions in text, vision and audio, especially the emergence of Her mode (GPT-4o Voice). This mode enables the AI model to respond to human language interactions in real time, just like Scarlett Johansson's famous movie "Her".

However, after its stunning debut, GPT-4o Voice has not been heard from again. Conneau's departure has added a hint of uncertainty to the future of GPT-4o Voice. What's more worth mentioning is that Conneau is also the head of GPT-5 research.

In this regard, one netizen asked: "To be honest, if GPT-5 is really groundbreaking and completely revolutionary, why are so many people leaving OpenAI?"

Image source: Reddit

In addition, another heavyweight model, Sora, also seems to be facing R&D difficulties.

On September 9, local time, CNBC broke the news that the Sora model, which was officially announced in February this year, is still in the research stage and may not be launched this year. On the one hand, it is due to political reasons, and on the other hand, it is not technically ready.

This article comes from the WeChat public account "Daily Economic News" (ID: nbdnews) , author: Daily Economic News reporter, 36Kr is authorized to publish.

Source
Disclaimer: The content above is only the author's opinion which does not represent any position of Followin, and is not intended as, and shall not be understood or construed as, investment advice from Followin.
Like
Add to Favorites
Comments