Blog

AI has made extraordinary progress in multimodal learning, particularly in aligning text and images into shared embedding spaces. Models like CLIP and GPT-4V(ision) are prime examples of this success, showing how architectures can bridge the gap between vision and language. But when we add videos to the mix, the landscape changes dramatically. Despite existing architectures and innovations, aligning text, images, and videos into a common embedding space is still imperfect and incomplete. ...

If we can’t deploy something useful to real world then what is the point of having this amazing AI technology at hand. This is my POV on Why I Believe AGI is the Unclear Goal (and Why Specialisation is the “Practical” Useful Way Forward) In the quest to create machines that think and learn like humans, Artificial General Intelligence (AGI) has often been regarded as the “holy grail” of artificial intelligence. However, I believe that AGI may not be the most practical goal and could be based on a misunderstood view of intelligence itself. Instead, I suggest a more grounded approach that mirrors the highly specialised nature of human knowledge and skills. ...

I would like to take a moment and talk about the most important skill we, as engineers, must master which is understanding the problem before designing the solution. There’s a simple but powerful truth in problem-solving: If you don’t fully understand the problem, you’ll either overcomplicate the solution or fail to solve it entirely. This applies to everything from AI and robotics to everyday challenges at work or home. I would like to write about it because it’s so critical and I often find especially during interviews/projects/brainstorming sessions that engineers don’t pay enough attention to it. As a result it leads to waste of resources and time on doing wrong things. ...

If you’re an aspiring robotics engineer, you’ve probably dreamed of building robots that seamlessly interact with humans, understanding their actions, anticipating their needs, and working safely side-by-side. Let’s talk about one critical component which is extremely essential to build such robots, and the name of that component is “Action Recognition.” What is Action Recognition? “SlowFast networks, introduced by researchers at Facebook AI, use two parallel video pathways—one operating at a lower frame rate for richer spatial context (Slow), and one at a higher frame rate for capturing rapid motion (Fast). This dual-pathway approach efficiently fuses both slow and fast temporal features, providing state-of-the-art performance in video action recognition.” ...

The Broken Code Advantage There’s a moment in every programmer’s career when they realize something counterintuitive: the people getting promoted aren’t the ones writing the most code. They’re the ones fixing it. This seems backwards. We celebrate the architects of new systems, the builders of elegant APIs, the creators of innovative features. But if you watch closely, you’ll notice that the engineers who become indispensable, the ones who get pulled into the most important meetings, who get the urgent Slack messages, who somehow always seem to know what’s really going on, aren’t the prolific feature builders. They’re the debuggers. ...