Grok 4 Release Tonight

xAI rolls out Grok 4 today alongside Grok Code 4 and an “Prolonged” variant buried in upsell code. Rumors tie “Prolonged” to the long-awaited Big Brain feature for deeper reasoning. Leaked benchmarks peg Grok 4 at 45% on the Humanity Last Exam, up from Grok 3, targeting devs and technical users needing stronger output. Voice mode and screen sharing may drop too, adding real-time collaboration muscle.

But Grok’s rollout could stall after its bot was silenced on 𝕏 for unfiltered content, throwing fresh light on xAI’s moderation gaps. Read more.

OpenAI just pulled 4 top engineers from Tesla, xAI, and Meta to bulk up its scaling team, per an internal Slack memo from cofounder Greg Brockman. David Lau, Tesla’s former VP of software, joins together with Uday Ruddarraju and Mike Dalton, ex-xAI engineers who helped construct Colossus, the 200,000-GPU supercomputer, and Angela Fan, an AI researcher from Meta. This crew will reinforce Stargate, OpenAI’s massive backend system powering its next-gen foundation models.

Speaking at Sun Valley, CEO Sam Altman didn’t hold back on the escalating talent fight with Meta, dropping a vague promise that OpenAI’s upcoming hardware “will probably be great.” He also took a swipe at smart glasses hype, despite his daring frames, he says they’re just glasses. Days earlier, Altman broke ranks with Democrats, calling himself “politically homeless” and blasting what he sees as anti-innovation, anti-billionaire politics choking Silicon Valley’s edge. Read more.

Hugging Face has dropped Reachy Mini, a $299 desktop robot built to crack open the robotics industry. At just 11 inches tall, this DIY humanoid hooks straight into the Hugging Face Hub, giving thousands and thousands of developers plug-and-play access to AI models. Reachy Mini runs on a Raspberry Pi 5 with full wireless autonomy, six degrees of freedom, a camera, mics, and animated antennas, packing serious capability for its price.

Its newly launched SmolLM3 squeezes multilingual, long-context reasoning into just 3 billion parameters, handling as much as 128k tokens with streamlined linear and grouped attention. Trained on 11T tokens, it rivals heavyweight 7B+ models for tool use, few-shot tasks, and multilingual QA, but runs light enough for edge hardware and personal stacks. Read more.

Meta has reportedly hired Ruoming Pang, Apple’s head of AI models, to affix its advanced AI superintelligence unit, Bloomberg revealed. Pang led Apple’s foundational AI model development, including the systems powering Apple Intelligence and on-device features. His exit signals growing instability inside Apple’s AI division, which has struggled to match rivals like OpenAI and Anthropic.

Insiders suggest more departures could follow. With Siri still underperforming, Apple is reportedly seeking to plug the gap with third-party models. For Meta, it’s one other high-profile win in a growing stack of AI hires. Pang brings expertise in lightweight, on-device AI, precisely the form of talent Meta CEO Mark Zuckerberg is stacking up from DeepMind, OpenAI, and Secure Superintelligence. Read more.

DeepMind has filed a patent for a system enabling AI agents and robots to generate “inner speech” to raised understand and learn tasks. The technology uses image or video input, comparable to an individual picking up a cup, and pairs it with natural language descriptions via a language model. This internal monologue helps the agent link visual input with contextual reasoning, improving decision-making.