AI News September 26 2025: Robots, AI Videos & $30M Funding
Executive Summary
AI News Roundup: Robots Get Smarter, Meta Pushes AI Videos, and Recruiting Goes LLM
The AI world took a breather yesterday. No mega-rounds. No AGI announcements. Just three developments worth your attention.
Google’s Robots Can Now Think Ahead (And Google Things)
Yesterday morning, Google DeepMind quietly dropped something significant. Their updated Gemini Robotics 1.5 models aren’t just incremental improvements.
These aren’t your typical single-task robots anymore.
The new models work in tandem. One drives the robot’s actions, while Gemini Robotics-ER 1.5 handles the reasoning. Carolina Parada, DeepMind’s Head of Robotics, explained it simply: robots can now think multiple steps ahead before acting. They can search the web mid-task. They can pack your suitcase based on weather forecasts for your destination.
Last March, these same models could only handle singular tasks like unzipping a bag. Now? They’re separating laundry by color and making complex decisions. The shift from “one instruction at a time” to “genuine understanding and problem-solving for physical tasks” happened in just six months.
Skills learned on one robot transfer to completely different hardware. DeepMind tested this across their ALOHA2 two-armed robot, the Franka bi-arm system, and Apptronik’s humanoid Apollo. Same model, different bodies. It worked.
Meta’s “Vibes” Feed: Nobody Asked for This
Mark Zuckerberg announced Vibes yesterday. It’s a feed of AI-generated videos.
The top comment on his announcement post? “Gang nobody wants this.”
Another popular response: “Bro’s posting ai slop on his own app.”
Users can generate videos from scratch or remix existing ones. Add music. Adjust styles. The algorithm learns what you like and shows you more of it. Standard social media stuff, except everything’s synthetic.
This launch feels oddly timed. YouTube’s cracking down on AI-generated content. Meta itself said earlier this year that creators should focus on “authentic storytelling.” Now they’re building entire feeds of generated videos.
The feature’s rolling out now. You’ll see fuzzy creatures hopping between cubes. Cats kneading dough. Ancient Egyptians taking selfies. All generated.
Juicebox Lands $30M to Fix Recruiting Search
While everyone else chases AGI, Juicebox raised $30 million to solve a boring problem that actually matters.
Their AI recruiting platform doesn’t use keywords.
Sequoia Capital led the Series A. Co-founders David Paffenholz and Ishan Gupta built it to find qualified candidates that traditional search misses. The LLM-powered system reads resumes differently than keyword matching. It understands context.
Total funding sits at $36 million now. They’re scaling operations and enhancing the platform.
What Didn’t Happen Yesterday
No new foundation models launched.
OpenAI, Anthropic, Microsoft, and Meta’s AI teams stayed quiet. No papers hit ArXiv worth noting. No regulatory bombshells from Brussels or Washington. NVIDIA didn’t announce another trillion-dollar partnership.
Sometimes the quiet days tell you as much as the noisy ones. The industry’s building, not announcing. Teams are heads-down on the stuff they revealed at summer conferences.
The Pattern
Look at what actually shipped yesterday. Robots that plan ahead. Social feeds nobody requested. Tools that fix real recruiting problems.
Two of these solve problems. One creates them.
That ratio feels about right for where we are in September 2025. The infrastructure investments from earlier this year are turning into products. Some useful. Some not. The market will sort it out.
Next week might bring the next GPT-6 or Claude 5. Or it might bring more incremental progress on making robots fold laundry better. Both matter. Just differently.
What developments are you tracking? The quiet progress or the headline grabs?
BC
September 26, 2025The DeepMind robotics update is the most significant piece here from a technical perspective. The shift from single-task execution to multi-step reasoning with web search integration represents a meaningful architectural leap. Having tested various reasoning models in my lab setup, I can appreciate how challenging it is to achieve consistent multi-step planning, even in text-only environments. Extending this to physical manipulation with real-time decision-making is genuinely impressive.
The cross-hardware transferability is particularly noteworthy. In my experience benchmarking LLMs across different hardware configurations, I’ve seen how model performance can vary significantly based on quantization and memory constraints. The fact that DeepMind’s approach works consistently across ALOHA2, Franka, and Apptronik systems suggests robust abstraction layers that could accelerate robotics deployment.
Meta’s Vibes feels like a solution searching for a problem. The user response mentioned in the article mirrors what I’ve observed in AI-generated content testing – there’s often a quality valley where generated content is technically competent but lacks the authenticity that makes content engaging.
Juicebox’s approach to LLM-powered recruiting makes more practical sense. Having worked with various language models for text analysis, the limitations of keyword matching are obvious when you see how well modern models understand context and semantics.