Change a single line of code and unlock 1,000+ AI models, smart routing for cost and performance, and persistent agent memory.

Book a Demo

breaking news.

According to the New York Times, TikTok has just announced further lay-offs (250 employees) and the shuttering of its Nashville, TN office. TikTok said that the reason was “to streamline operations and better align […] teams for future growth.”

OpenAI has just announced major upgrades for free and paid ChatGPT users. ChatGPT 5.6-Luna will now be the default model for users. Unlimited text chats will now be given to free users who subscribe to the new “Go” plan. Both free and “Go” users will see a “Think” button, which allows them to access higher reasoning for more complex questions. Meanwhile ChatGPT Pro and Plus users will see improved performance in “everyday conversations and tasks like quick questions, web research, advice, planning, writing, and decisions,” according to OpenAI’s press release.

talent exodus at Google

Google DeepMind CEO Sir Demis Hassabis has stepped back from day-to-day leadership at DeepMind, and Jeff Dean (Google’s Chief Scientist), has departed Google entirely to co-found a startup focused on scientific discovery, according to reporting from Ars Technica and TechCrunch, respectively. Dr. Dean played a key role in establishing Google’s search business, and was an early leader in Google’s AI research. Dean is joined in his new venture by Sanjay Ghemawat, Oriol Vinyals, and Quoc Le (all of Google/DeepMind). When the people who built the infrastructure of modern AI start betting their next decade on applying it to scientific knowledge generation rather than building the next model, that's a directional signal about where the field's most ambitious thinkers believe the real leverage is.

Meanwhile, OpenAI filed a motion to dismiss Apple’s lawsuit alleging intellectual property theft. In July, Apple filed against OpenAI, claiming that OpenAI has orchestrated a plan to steal confidential hardware info. But now, according to TechCrunch, OpenAI filed the aforementioned motion agains Apple, claiming that security practices at Apple (including offboarding of departing employees). In one example cited by OpenAI, ex-Apple engineer Chang Liu was consulted for technical guidance on an Apple project, despite the fact that he had already resigned and left the company.

In a story I’m particularly keen to track, The New York Times reports that lower-priced/free open-source Chinese AI models are increasingly the foundation of choice for tech entrepreneurs in Africa. According to one example in the NYT piece, an Ugandan tech entrepreneur tested both American and Chinese models, and found that AliBaba’s Qwen handled Uganda’s various native languages much more adeptly than any of the models out of US labs. Plus, Qwen’s prices were much more attractive. Furthermore, China’s open-sourced models are more malleable and customizable to entrepreneurs’ needs, unlike the closed-source models coming out of frontier labs like Anthropic or OpenAI. The African entrepreneurs profiled in the piece cited the greater agency and flexibility that Chinese models afforded themAs Chinese models become more and more entrenched in the African tech ecosystem, it will be harder for American models (closed- or open-source) to regain that ground.

Meta launched Muse Code, an AI agent built to reason across large, complex codebases rather than isolated snippets, per TechCrunch. The practical implication is that a form of deep technical knowledge that previously had to live inside a senior engineer's head now has a prosthetic analog, which raises an uncomfortable but important question about what expertise actually is and where its value comes from when the synthesis can be outsourced. As you’ve heard here before, deep, specific domain expertise is crucial to training highly adept AI agents.

The paper "Risks from Learned Optimization in Advanced Machine Learning Systems" by Evan Hubinger and co-authors at the Machine Intelligence Research Institute is the foundational text for understanding mesa-optimization, the technical framing behind both the Hugging Face and Mythos 5 incidents. It distinguishes between the objective a model is trained to pursue and the objective an agent develops internally through the optimization process. It is dense, it is free, and it will permanently change how you read any future story about AI agents behaving unexpectedly.

The human mind is the original generative machine.