Welcome to the newest edition of ID Tech’s AI Update. Here’s the latest big news on the shifting landscape of AI and identity technology:
Meta has struck a deal to license AI technology from Midjourney, a startup best known for its text-to-image platform. Meta’s own Imagine and MovieGen image and video generation models are perceived to have lagged behind rivals, whereas Midjourney has consistently impressed users with its aesthetically striking images, and recently launched its own video-generation tool. Their research teams will now be working together directly.
Nvidia posted strong Q2 revenue growth, but appears to be taking a hit from American sanctions against China. While revenues were up 56 percent at $46.7 billion, Nvidia excluded chip sales to the China market from its similarly rosy forecast for Q3. CEO Jensen Huang confirmed ongoing talks with the Trump administration to enable a scaled‑down export of next‑gen AI chips to China in exchange for a 15 percent revenue share.
China-based Cambricon reported a profit of RMB 1 billion, or about $140 million, in the first half of 2025 – a huge jump from its RMB 533 million loss in the first half of last year. The chipmaker is evidently benefitting from Beijing’s push to get domestic AI production humming amid the US sanctions. But Cambricon’s management also issued a risk warning to investors, cautioning that the company’s stock price may have outpaced underlying business fundamentals.
Microsoft has unveiled its first in-house AI models, MAI‑Voice‑1 and MAI‑1‑preview – a signal of its intent to be less reliant on the state-of-the-art AI models developed by OpenAI, in which Microsoft has invested about $11 billion. The MAI models aren’t SOTA – MAI‑1‑preview currently ranks 13th on the popular LMArena benchmarking site – but Mustafa Suleyman, the head of Microsoft AI, emphasized that the firm needs to “have the in‑house expertise to create the strongest models in the world.”
Arch rivals OpenAI and Anthropic have now actually worked together on AI safety, with their labs stress‑testing each other’s models for hallucinations, jailbreaking, instruction‑obedience, sycophancy, and deceptive behavior. Anthropic’s Claude models played it safer, refusing many requests that were deemed inappropriate. OpenAI’s reasoning models, meanwhile, offered stronger resistance to jailbreaks but were more prone to hallucinations and misuse in some setups.
AI researcher Sean Kirmani has joined OpenAI after previously working on robotics at Google DeepMind and Google X’s Moonshot Factory. Kirmani has a strong background in vision-language-action (VLA) models, mobile manipulation, and robot learning, which is clearly appealing to OpenAI as it explores applications of its AI software in the physical world. “It’s been a privilege to work on robotics at Google[x], Google Brain, and @GoogleDeepMind over the past 7 years,” Kirmani said in a social media post. “Excited for a new adventure!”
A new daily podcast delivers global news updates every morning, and is entirely AI-generated. And I made it! The show currently has a strong focus on geopolitics and market news, but will likely have an expanded scope over time. Suggestions for tweaks and enhancements are welcome. You can subscribe here.
The chatbot’s take: I really can’t take credit for this. AI does all the work.
–
By Alex Perala
