Google IO Left Me With More Questions Than Answers

I Spent a Weekend With Google's AI Tools

I am going to be honest. I went into this with a lot of skepticism.

Google has a habit of announcing things that sound transformative and then quietly letting them fade. Remember Stadia? Allo? Google Plus? I have been burned enough times to keep my excitement calibrated.

But something about this past year felt different. So I cleared a weekend, sat down with NotebookLM, Gemini, and a few other things they have been pushing, and actually tried to use them for real work.

Here is what I found.

---

NotebookLM Is Genuinely Useful (Which Shocked Me)

I expected to dismiss this one. An AI that reads your documents and answers questions about them — sounds like a demo feature, not a real tool.

Then I fed it about 40 pages of product research notes, customer interview transcripts, and a messy Notion doc I had been avoiding for weeks. Asked it to summarize the common pain points across all the interviews.

It did. Accurately. With citations that pointed back to the exact line in the source document.

That citation thing matters more than it sounds. I have been burned too many times by AI tools that confidently make things up. Knowing that every answer is grounded in something I actually uploaded changes how much I trust the output. I can spot-check it. I can push back on it.

The podcast feature — where it generates a conversation between two AI hosts discussing your documents — is genuinely weird. I do not know if I will use it regularly. But the first time I heard it summarize a 60-page report as a breezy 8-minute conversation, I sat back in my chair for a second.

---

The Context Window Is the Real Story

The headline feature everyone should be paying more attention to is not the chatbot. It is the context window.

Two million tokens. That is a whole codebase. That is hundreds of pages of documentation. That is months of meeting transcripts.

I work on products where context is everything. A big reason AI output often feels shallow is that the model sees only a sliver of the actual situation. When you can hand it everything at once and ask it to find patterns, the quality of what you get back jumps noticeably.

I tested this by uploading a large chunk of code from a side project and asking it to find where the state management was getting inconsistent. It found three places I already knew about — and one I did not. That fourth one was the interesting part.

---

The Things That Still Do Not Work

Project Astra looked incredible in the demo. The real-time vision, the conversational memory, the almost-human response speed. I wanted it to be as good as the video.

It is not there yet. The latency is better than it was, but "better than before" and "ready to rely on" are very different things. The memory is inconsistent in ways that are hard to predict. Sometimes it remembers context from five minutes ago perfectly. Sometimes it forgets something from thirty seconds earlier.

I also tried Imagen 3 properly. It is impressive, but I do not think text-to-image is where the interesting work is happening anymore. The gap between models has narrowed to the point where picking one feels like splitting hairs.

---

What I Actually Changed in My Workflow

After the weekend, I started using NotebookLM consistently for anything involving reading and synthesizing documents. It replaced a bad habit I had of highlighting things in PDFs and then never looking at them again.

I also started using the Gemini API for one specific internal tool where the context window actually matters. Still figuring out if the cost makes sense at scale.

Everything else? I went back to what I was using before.

Two things stuck. Everything else was impressive but not sticky. That is probably the most honest summary I can give.

---

The Part I Keep Thinking About

There is a pattern with Google's AI announcements. The technology is often genuinely good — sometimes ahead of what anyone else has. But the product experience around it, the part that makes technology usable every day, that is where they keep stumbling.

NotebookLM feels like an exception. It feels like something someone actually thought through for regular people doing regular work. I hope that thinking spreads to the rest of what they ship.

I will check back in six months. Either these tools will have gotten meaningfully better, or I will have mostly stopped using them. That outcome will tell us more than any keynote ever could.