top of page

7 OCTOBER | LONDON 2024

SEPTEMBER 12TH - 14TH
The O2, LONDON

The CogX Blog

Thought leadership on the most pressing issues of our time

OpenAI seeks new path to smarter AI as current development hits a wall



 By the CogX R&I team

November 13, 2024



The AI industry's powerhouse OpenAI is grappling with an unexpected challenge: diminishing returns in AI model improvement.


Credit: CogX


According to recent reports by the Information, the company's next-generation model, code-named Orion, isn't delivering the dramatic performance leaps seen in previous iterations. While the latest model boasts some improvements over existing systems, company testers reportedly observed a less pronounced advancement compared to the jump from GPT-3 to GPT-4.

 

Perhaps most concerning for the AI giant is Orion's inconsistent performance in crucial areas like coding, where it sometimes falls short of its predecessors.

 

In response to this slowdown, OpenAI has established a dedicated foundations team to tackle these challenges head-on. The team is reportedly exploring innovative approaches, including a controversial strategy to train the models on AI-generated synthetic data.

 

This development arrives at a critical juncture for the AI industry, where the dwindling supply of quality training data is emerging as a significant bottleneck.



More from OpenAI:


  • OpenAI loses another key player in AI safety: Lilian Weng, the former VP of Research and Safety at OpenAI, has announced her departure.


 

Now read the rest of the CogX Newsletter


Google's next-gen AI model, Gemini 2.0, might be closer than we think


Google appears to be quietly preparing its next major advancement. Recent findings by TestingCatalog uncovered what could be Gemini’s next major AI release hidden in one of its recent UI updates.


Thanks to the meticulous reverse engineering work by @testingcatalog, the intriguing new model, labelled as " 2.0-Pro-Exp-0111", was spotted in Gemini's model selector.


The new model reportedly delivers lightning-fast responses, though it still appears to be in development as evidenced by its performance on certain benchmark tests, including the "strawberry test," which it reportedly struggled to pass.


Credit: @testingcatalog


While these developments are exciting, it's worth noting that the current responses attributed to version 2.0 might be generated by existing models. According to the source of the leak, until Google officially confirms and releases Gemini 2.0, "we can't definitively say for sure what we're seeing". However, the early signs are promising.

 

 

Also in the news


AlphaFold3, the Nobel Prize-winning protein-structure prediction tool, is now open source. After initially keeping the code private, DeepMind has now made the model's code and weights accessible to scientists worldwide. This groundbreaking tool can predict the 3D shapes of proteins, crucial for understanding their function and developing new drugs.

 

Microsoft introduced Magentic-One: A new multi-agent AI system, aimed at helping enterprises automate complex tasks. This innovative system can handle a wide range of functions, from web browsing and file management to coding and command-line operations. By breaking down complex tasks into smaller subtasks and assigning them to specialised AI agents, Magnetic-One can work autonomously, reducing the need for human intervention. 

 

It's getting harder for Chinese companies to get hold of AI chips: TSMC, a major Taiwanese chipmaker, has been ordered by the US to halt shipments of advanced AI chips to Chinese companies, including Huawei. This move comes after US authorities discovered a TSMC-manufactured chip inside a Huawei processor, raising concerns about potential circumvention of existing export controls.

 

OpenAI acquired Chat.com: Previously purchased by HubSpot's CTO for a hefty $15.5 million back in 2023, the domain name ‘Chat.com’ is now the property of OpenAI. Now, typing "Chat.com" will take you straight to OpenAI's AI chatbot. But while some experts say this move could make ChatGPT even more popular worldwide, OpenAI hasn't revealed the exact cost of this splashy new domain purchase.

 

X is testing a free version of the AI chatbot Grok. Initially available only to premium users, Grok is now being rolled out to free users in some regions. While there are usage limits, this move could help xAI gather more user feedback and compete with other popular AI chatbots like ChatGPT and Claude.

 

Ever wondered if a robot could create million-dollar artwork?


AI art has reached a new milestone. Ai-Da Robot, a humanoid artist, recently sold a portrait of Alan Turing for a record-breaking $1.08 million at Sotheby's, far exceeding its estimated value of $180,000.


Credits: BBC


The sale marks more than just a price tag. As the first artwork by a humanoid robot to sell at auction, it represents a pivotal moment in the global art market. The portrait, part of a series titled "A.I. God", took up to eight hours to complete and was one of 15 paintings of the computing pioneer.

 

The breakthrough has sparked discussions in the art world. While some celebrate this as launching a new frontier in the global art market, others see it as a moment for reflection on AI's growing influence. As Ai-Da Robot itself put it, the work serves as "a catalyst for dialogue about emerging technologies".

 

In case you missed it


Watch XPENG's new Iron Robot working in a Factory:




Previous
Next
bottom of page