If the rumors flying around are anything to go by, the public can expect a new approach to OpenAI’s ChatGPT very soon. Reuters recently reviewed some internal documentation of the AI giant, and their findings align with the testimony of an insider that ChatGPT may soon get a new reasoning model codenamed “Strawberry.”
Microsoft-backed OpenAI is on a mission to prove that it can deliver advanced and novel reasoning capabilities which are second to none. Its teams are working round the clock on a project called Strawberry, a new research tool that ChatGPT will now feature. Based on an internal OpenAI document that Reuter saw in May, efforts are on top gear to birth the new model very soon. No date is, however, on the document for its release time, but the idea is cork valid.
Insiders at the company revealed that the working mechanism of Strawberry is a secret that is unknown even to many of the workers in the organization. However, from a few facts gleaned from the document assessed, it seems that the broad aim of this project is to empower the company’s AI tool to be proactive by conducting an autonomous and reliable internet search for deep research. In that case, it means that ChatGPT will not only generate answers to questions posed to it but also conduct deep research by combing the internet ahead.
Limitations of Current Commercial AI Models
According to AI researchers, the inability of AI models to view the world from a real-life perspective is the major setback of this technology. Users on Reddit also admit that while leading AI tools like ChatGPT can generate plausible answers to queries, it doesn’t have true reasoning and understanding.
“ChatGPT and other Large Language Models (LLMs) are neural networks trained on human written texts. They can produce incorrect answers and spread misinformation,” says NurseAI on Reddit
To date, no AI model has the capacity to offer what Strawberry is rumoured to be programmed to do. That said, it seems that everyone at OpenAI is trying to keep the giant step secret for now because when asked about Strawberry, almost everybody finds a way to deflect the question. For instance, one spokesperson to OpenAI gave this response when asked about the planned model:
“We want our AI models to see and understand the world more like we do. Continuous research into new AI capabilities is a common practice in the industry, with a shared belief that these systems will improve in reasoning over time.”
It cannot be any clearer that the spokesperson deliberately left out the Strawberry question.
However, as at last year, the project which was still known as Q* then was widely acknowledged as a significant breakthrough. At least two people privy to the inner workings of OpenAI hinted that the model could answer very tricky math and science questions, which no existing model can attempt.
Bloomberg did report that an AI research project demo by OpenAI demonstrated what it called “a human-like reasoning skills” at one of its internal all-hands meetings. However, while an OpenAI spokesperson confirmed this meeting, they declined admitting whether it was Strawberry or something else.
Casino experts are already super-excited on the positivity it portends for the industry. From helping to detect fraud to analyzing players’ gaming patterns, online casinos stand to gain a lot if Strawberry becomes a reality. Existing casino review sites like Ecasinos.ph, which gives useful information about casino sites in the Philippines, can also automate many of their internal processes using the Strawberry model. The advanced reasoning model can collate and analyze authentic information about casino sites, offers, and bonuses using its deep research capabilities. With these pieces of information, it can help rule out fake sites and rank the authentic ones.
OpenAI Hints Developers
Although OpenAI is doing its bit to keep Strawberry a secret for now, it’s on record that the company has started hinting developers and some outsiders of its plans to roll out a slightly higher reasoning model for its ChatGPT. This signalling is necessary because such updates will require developers and other relevant outsiders The people that leaked this information declined further comments on it because it wasn’t in their place to announce this big discovery,
Again from leaks, it appears that OpenAI wants its Strawberry to include a “post-training” generating AI models or adapting its earlier models to enhance the tool’s performance on things it had already been trained for. Post-training feature is not new in the tech world. Although it presents in many flavours, the basic idea is to fine-tune responses by being able to take feedback.
In some ways, Strawberry is similar to Self-Taught Reasoner (STaR) developed in 2022 at Stanford. According to source privy to the workings of this new reasoning, just like STaR enables AI models to bootstrap into a higher intelligence by iteratively coming up by its self-generated data, so also is Strawberry.
Some of the tall dreams that Strawberry is aiming to achieve based on the document Reuter reviewed is Long-Horizon Tasks (LHT). These tasks are incredibly complex and demand that AI models plan ahead using several data sets beforehand. To achieve this, OpenAI is rejigging its models to perform what it calls deep research by planning ahead. What is not yet clear is the data the AI will dwell on and what will constitute the timeframe that it will use to prepare ahead.
We also heard that OpenAI will utilize CUA (Computer-Using Agents) to help its ChatGPT perform deep research and act based on what its found. It’s an exciting time ahead if these ambitions and aims come to reality. The internal document did point to the seriousness of the AI giant to pursue the vision rigorously.
The company is, however, not the only one seeking to revamp its reasoning model. Other AI tech companies like Google, Microsoft, and Meta have not also gone to bed. They are currently carrying on experimentations, just like the numerous academic labs that focus on AI research are not also resting on their oars. However, there are no records that these other AI outlets are pursuing human-like reasoning model like the Strawberry of OpenAI.
