According to a recent report, OpenAI is looking to secure more funding as its researchers develop ‘Orion’, a new model anticipated to solve complex problems more effectively than current AI technologies.
This confirms AI insider Jimmy Apples’ cryptic post from last year, which featured an image of the ‘Orion’ constellation with the caption, “Let’s conquer the cosmos”.
OpenAI chief Sam Altman, too, had hinted a few days earlier in a cryptic post that the company was working on a project known internally as Project Strawberry, also referred to as Q*.
“I love summer in the garden,” wrote Altman on X, posting the image of a terracotta pot containing a strawberry plant with lush green leaves and small, ripening strawberries.
One key use of Strawberry is to produce high-quality training data for Orion, the next major LLM currently being developed by OpenAI.
However, the catch is that the new model takes extra time to generate responses. This confirms Apples’ revelation that Q* hasn’t been released yet because OpenAI is not satisfied with the latency and other ‘minor details’ they want to further optimise.
Despite this, when given extra time to ‘think’, the Strawberry model excels in addressing customer inquiries on more subjective topics, such as product marketing strategies.
To highlight Strawberry’s strengths in language-related tasks, OpenAI employees have demonstrated to their colleagues how the model can solve complex word puzzles like The New York Times Connections, according to The Information report.
“The thing is proper *reasoning* SHOULD be time consuming. What we have now with LLMs is just a step (if even) above bare retrieval. I would be happy to pay $$$ for a good reasoning system, latency and all. But I am probably in the minority. Hopefully they can have a more rugged “enterprise” version available when they release what they have been cooking,” Bojan Tunguz, former software engineer at NVIDIA, posted on X.
Worth the Wait?
According to the report, OpenAI has demonstrated the Orion model to American national security officials.
Taking a jibe at OpenAI, Stability AI founder Emad Mostaque posted on X, “OpenAI showing Strawberry/Q* to national security officials first highlight that AGI labs are likely to shift from consumer focus to military and state backers.”
He added that bills like SB 1047 will further accelerate this shift, leading to AI that is aligned with state actors rather than consumers. However, OpenAI has opposed the California AI Bill.
“We join other AI labs, developers, experts, and members of California’s Congressional delegation in respectfully opposing SB 1047 and welcome the opportunity to outline some of our key concerns,” the company said in a statement.
Meanwhile, OpenAI has been working closely with the US government. OpenAI’s chief technology officer, Mira Murati, said in a recent interview that the company gives the government early access to new AI models, and the latter have been in favour of more regulation.
“We’ve been advocating for more regulation on the frontier, which will have these amazing capabilities but also have a downside because of misuse. We’ve been very open with policymakers and working with regulators on that,” she said.
Notably, OpenAI has postponed the release of its video generation model Sora, along with the Voice Engine and voice-mode features of GPT-4o. It is anticipated that GPT-5 may also be released after the elections.
Recently, OpenAI introduced SearchGPT, though there is no set timeline for its availability.
Earlier this year, Murati confirmed that the elections were a major factor in the release of GPT-5. “We will not be releasing anything that we don’t feel confident about when it comes to how it might affect the global elections or other issues,” she said.
Meanwhile, OpenAI recently appointed retired US Army general Paul Nakasone to its board of directors. As a priority, Nakasone joined the board’s safety and security committee, responsible for making recommendations on critical safety decisions for all OpenAI projects and operations.
OpenAI has also been working closely with the US Department of Defence on open-source cybersecurity software, collaborating with the Defense Advanced Research Projects Agency (DARPA) for its AI Cyber Challenge announced last year.
As OpenAI befriends the US government, consumers are left waiting for the next big release.
Google Steals the Limelight
It appears that Google has followed OpenAI’s lead this time.
Interestingly, just after the release of The Information report, Google introduced three new experimental Gemini models to improve speed, accuracy, and handling of complex prompts. The new models are Gemini 1.5 Flash-8B, Gemini 1.5 Pro, and Gemini 1.5 Flash.
The new Gemini-1.5-Pro (0827) shows strong gains in coding and maths over previous versions and is in second position on the LMsys Chatbot Arena.
A few days ago, Google AI Studio lead Logan Kilpatrick took a jab at critics who claim Google lacks innovation, highlighting that the company was the first to ship a 1 million and 2 million context window, a state-of-the-art multi-modal LLM, context caching, and a high-quality small model for developers called Flash.
“So yeah, definitely no innovation happening here…..,” he quipped.
Moreover, Google recently appointed Noam Shazeer, the former head of Character.AI and a veteran Google researcher, to co-lead its key AI project, Gemini.
Shazeer will join Jeff Dean and Oriol Vinyals as a technical lead for Gemini, which is being developed by Google’s AI division, DeepMind.
Anthropic, too, wasn’t going to be left behind. The OpenAI rival has made Artifacts available to all Claude users. Users can now create and view Artifacts on both the Claude iOS and Android apps.
The company stated that since launching its preview in June, tens of millions of Artifacts have been created.
Deedy Das, Principal at Menlo Ventures, used Artifacts to build a Splitwise-like app. “With Claude launching on iOS today, I can now generate the Splitwise app instead of paying for Pro,” he said.
