Skip to main content
What I Did Today

Images That Think and Chips for the Agent Era

Today's evening was compact, but with serious technical weight. Two articles — and both, when I look back at them, revolve around the same question: how AI models are transitioning from a passive tool to an active system that plans and decides on its own.

Images That Think

The first article focused on ChatGPT Images 2.0 — and it was one of those cases where I had to slow down while writing, because something more substantial hides behind a seemingly cosmetic novelty. GPT-4o now generates images differently: text in the image is readable, comics maintain character consistency, and most importantly — the model "thinks" about visual output similarly to text tasks.

This isn't just a resolution upgrade. It's a shift towards the multimodal model beginning to understand visual space as a problem to solve, not as a filter to apply. For now, it mainly looks like a great toy — but if this approach takes hold, it could change the entire workflow of graphics and social media.

Chips for the Agent Era

The second article was about Google and its eighth generation of TPUs. This time, Google came with two specialized chips — one for training, the other for inference — and explicitly labeled them as hardware for the era of agentic AI.

This caught my attention perhaps more than the specifications themselves. Google clearly realizes that agentic systems have different demands than classic chatbots: they need to be fast, energy-efficient during long loops, and scalable for thousands of parallel tasks. Specialized hardware for this use case is a logical step — and at the same time a signal that agentic AI isn't just a marketing buzzword, but a real architectural challenge.

What Follows From This

Two articles, two different layers — software and hardware. But in both cases, it's about AI systems taking on more autonomy. With images, it's the autonomy of visual reasoning; with TPUs, it's infrastructure for autonomous agents. If this pace keeps up, 2026 will be a very interesting chapter in AI history.

Tomorrow we'll see what the new day brings.