AI: OpenAI's AI plans beyond 'Strawberry' RTZ #476
Looks like we may have an inkling on OpenAI and founder/CEO Sam Altman’s ‘Strawberry Summer’ I discussed recently, a reference to their ‘Level 2’ AI Reasoning project formerly known as ‘Q*’.
Code names aside, this is their self-hinted path to AGI (Artificial General Intelligence) across 5 steps, from ChatGPT like chatbots today to AI Reasoning, Agents, Innovator AIs, and Organizational AIs. Of course, Google with its AI Gemini, has its own 5 level roadmap as well.
The Information in “New Details on OpenAI’s Strawberry” explains:
“Strawberry, OpenAI’s reasoning-focused artificial intelligence, is coming sooner than we thought.”
“OpenAI plans to release Strawberry as part of its ChatGPT service in the next two weeks, earlier than the original fall timeline we had recently reported, said two people who have tested out the model. Release timelines are always subject to change, of course, but we have a few other new details about the product.”
“We should explain that while Strawberry is part of ChatGPT, it’s a standalone offering. Exactly how it will be offered is unclear: one option is for Strawberry to be included in the dropdown menu of AI models customers can pick from to power ChatGPT, the people said. And it’s quite different to the regular service, with some advantages and shortcomings.”
“Of course, what most differentiates Strawberry from other conversational AI is its ability to “think” before responding, rather than immediately answering a query, said the two people who have tested the model. That thinking stage usually lasts 10 to 20 seconds, they said.”
“But there are other key differences. For one thing, the initial version will only be able to take in and produce text—and not images—which means it isn’t yet multimodal the way other OpenAI models are. As most large language models released today are multimodal, this seems to be a noticeable shortcoming. The decision to release it as text-only could reflect the pressure OpenAI is feeling to release products as it faces more competition.”
And of course, there’s the issue of how to price this new service, especially given the variable cost aspects of customer driven AI services I’ve discussed.
“Then there’s pricing. Strawberry is likely to be priced differently to OpenAI’s chatbot, which has free and subscription-pricing tiers. We’re not sure exactly how Strawberry will be priced, but it will likely have rate limits restricting users to some maximum number of messages per hour, with the potential for a higher-priced tier that’s faster to respond, according to another person with knowledge of the product. Such a cost-saving move could prompt more people to pay up for the new model, similar to the reason OpenAI caps messages for free users of ChatGPT.”
“We also would expect paying ChatGPT customers to have access to the first Strawberry model before it’s released to the bigger, free tier of users. Whether OpenAI would charge prices significantly higher than ChatGPT today for customers to use a bigger version of Strawberry remains to be seen. (A spokesperson didn’t have anything else to add on these topics when we reached out.)”
It’s important to emphasize that ‘Strawberry’ is but one step in how AI systems will ‘reason’ and ultimately do ‘agentic tasks:
“Strawberry also is expected to be easier to use than GPT-4o for complex or multistep queries. Currently, customers have to type all kinds of additional words into ChatGPT to get the answer they want, such as telling the chatbot to walk through its intermediate reasoning steps to arrive at its final answer, otherwise known as “chain-of-thought prompting.” Strawberry’s capabilities are supposed to help customers avoid doing that or other hacks to achieve smarter results.”
“This means that not only will Strawberry be better at math problems and coding, but also at more “subjective” business tasks, like brainstorming product marketing strategies, as we’ve previously reported. In these sorts of tasks, the model will provide suggestions that are more specific to a user’s company and more detailed, like generating a week-by-week execution plan.”
“Strawberry’s thinking stage helps it avoid making errors, one of the people said. The extra time also makes Strawberry more likely to know when it needs to ask the customer follow-up questions so it knows how to fully answer their question.”
“But OpenAI may have some kinks to iron out before or after launch.”
But as always with AI services in these early days of the AI Tech Wave, the products are still a ‘work in progress’. We’ve seen that being true for companies ranging from Microsoft’s AI Copilot to Apple’s ‘Apple Intelligence’.
OpenAI’s ‘Strawberry’ is no different:
“For instance, even though Strawberry theoretically is able to skip its thinking step when people ask it simpler questions, the model doesn’t always do that in practice, said one of the people who have tested the model. As a result, it’s possible it might mistakenly think too long to answer queries that OpenAI’s other models can answer in a jiffy.”
“Some people who’ve used a Strawberry prototype have complained that its slightly better responses compared to OpenAI’s currently released GPT-4o aren’t worth the extra 10 to 20 seconds of waiting, the person said.”
And AI memory is also something that’s part of the roadmap, but also a ‘work in progress’:
“And while Strawberry also aims to remember and incorporate previous chats it’s had with a customer before answering new questions—an important detail when users have specific preferences, like a certain format they want their software code written in—the prototype has sometimes struggled with that too, this person said.”
“OpenAI may be the runaway leader in products powered by large language models, but it faces growing competition. Last month, for instance, Google beat OpenAI by broadly launching an AI-powered voice assistant that’s flexible enough to handle interruptions and sudden topic changes from users. OpenAI first announced its own voice assistant, GPT-4o Voice, in May but then delayed it to improve its safety measures, such as making sure it would refuse inappropriate content, the company said.”
“Strawberry could help OpenAI get back the momentum it’s had for most of the last two years (but that’s assuming the launch goes well).”
Strawberry of course is a small part of OpenAI’s rich product line expected this year and next, not the least of it is it’s long-awaited GPT-5. With iterations like OpenAI’s text to video Sora and others in between.
That of course is the successor to its LLM AI class leading GPT-4o for ‘Omni’, and goes by code-names ranging from ‘Orion’ to ‘GPT-Next’.
That last code name we get from an OpenAI executive in Japan, as reported by TechRound in “GPT-NeXT: The AI Model “100 Times More Powerful” Than GPT-4”:
“About a year and a half after the release of GPT-4 in March of 2023, users have been told that they can expect the launch of a new and improved version of Chat-GPT in the near future. In the very, very near future, in fact – before the end of 2024.”
“Indeed, Tadao Nagasaki, the CEO of OpenAI in Japan, revealed information about the much-anticipated successor of GPT-4, a model that’s been named GPT-NeXT.”
“This isn’t the first time we’ve heard of GPT-NeXT, however. The project was first unveiled in May this year at a tech event in Paris, but at that point, OpenAI was tight-lipped about how it would be different from previous models and exactly what it would be capable of doing.”
“Although Nagasaki hasn’t exactly spilt all the (metaphorical) beans, he certainly has given us a whole lot more than we knew before, and from the information he shared at the KDDI Summit a few days ago, it seems like GPT-NeXT is expected to do great things.”
“What Did Tadao Nagasaki Reveal About GPT-NeXT?”
“It should come as no surprise to those familiar with the tech industry that this “big news” about OpenAI’s exciting new model wasn’t exactly all revealing. In a similar fashion to the way most tech giants seem to enjoy sharing news on new products, it was more of a teaser. Just a little more information about the impending release that we didn’t have before to get us hooked.”
“Nagasaski stated very clearly that not only is GPT-NeXT expected to be better than GPT-4, it’s anticipated to be a whopping 100 times more powerful.”
“So, what does that mean? What does 100 times more power mean for AI and how would that even work?”
“What Will Make GPT-NeXT So Powerful?”
“A lot of what was revealed at KDDI regarding NeXT came in the form of a slideshow (or, to be specific, a single slide), but even so, that provided a whole lot more than anyone knew before.”
“GPT-NeXT, which is expected to exist separately from other future models including GPT-40 and GPT-5 Strawberry, is set to make use of a smaller version of Project Strawberry.”
“Project Strawberry (other than having a name comparable to something that may have been thought up by a hungry toddler) is a project that was launched by OpenAI and has gained much attention since. It makes use of a specialised method that aims to improve the performance of AI models to become more specific after they’ve been trained on highly generalised data.”
“On top of that, the seriously enhanced performance of NeXT is set to be a result of improvements in the architecture of the platform rather than a mere increase in computing resources.”
“Previously known as Project Q (or Q-Star), Project Strawberry is believed to represent OpenAI’s attempt to move towards creating Artificial General Intelligence (AGI).”
“And Artificial General Intelligence is a doozy.”
“In theory, it’s a subset of AI that has the potential to surpass human intelligence and general capabilities.”
“Naturally, since Project Strawberry is thought to be making a move towards achieving AGI and GPT-NeXT is set to make use of Project Strawberry in one way or another, it certainly seems like OpenAI is quite ambitious. But, if they’re successful, then multiplying the power of GPT-4 by 100 may very well be on the cards.”
So OpenAI has quite the AI Word Salad for its AI roadmap a couple of months from the two year anniversary of its seminal ‘ChatGPT moment’ in November 2022. And it underlines how early things are in this AI Tech Wave for even the leading LLM AI company in terms of AI definitions and roadmaps.
It’s important to emphasize that nothing above, from ‘Strawberry’ to ‘Orion’ to ‘GPT-NeXT’ has been OFFICIALLY RELEASED by OpenAI.
It’s all a series of general ‘trial balloons’ by OpenAI executives either directly, or indirectly through media sources. A way to gauge industry and media reactions, and generally convey forward momentum in their AI product iteration plans.
And provide some comfort that there’s an eventual ‘Syncup’ in AI investments and Returns.
It’s the way to signal AI Scaling progress and momentum, as the industry itself continues to expend tens and soon hundreds of billions in AI Infrastructure Capex.
We should expect Google, Anthropic, Meta, Nvidia, Apple, and of course Elon Musk with xAI/Grok/Tesla, to add their own AI code-names into the AI word salads to come. And that’s even before we get to LLM AI driven humanoid robots, especially in our homes to make actual salads.
Consider these AI Word Salads a part of the increasing ‘AI Table Stakes’ I’ve started to discuss. So far, OpenAI’s leading this game. Stay tuned.
(NOTE: The discussions here are for information purposes only, and not meant as investment advice at any time. Thanks for joining us here)
