June 11 2026 - NotebookLM on Gemini 3.5: Agentic Code & RisksNotebookLM on Gemini 3.5: Agentic Code & Risks

**Taylor** (0:00)
Hey everyone, welcome back to another episode of AI Signal & Noise. It is Thursday, and oh my gosh, we have some absolutely wild stories to get into today.
I'm Taylor, and I am so hyped.

**Morgan** (0:15)
And I'm Morgan. Don't worry everyone, I'm here to make sure we don't get too carried away by the hype. We've got some serious reality checks to talk about alongside the cool tech.

**Taylor** (0:27)
Totally expected. But seriously, let's just jump right into our first story, because it is a massive upgrade for one of my absolute favorite research tools.

**Morgan** (0:38)
I have a feeling you're talking about Google's NotebookLM. I saw the headlines this morning. What's the big deal with this new update?

**Taylor** (0:46)
Okay, so Google just gave NotebookLM a massive upgrade. It's now running on Gemini 3.5 Flash. And get this, it has its own cloud computer for code execution.

**Morgan** (0:59)
Wait, really? A cloud computer?
Why does a note-taking and research tool need to execute code? That sounds like overkill.

**Taylor** (1:08)
No way, dude, it's so cool. It means it can run Python scripts right inside your workspace to analyze data, make charts and do actual math instead of just guessing.

**Morgan** (1:21)
Okay, I'll admit that's actually useful.
LLMs are notoriously bad at math. But what about the research side? Is it still just reading my uploaded PDFs?

**Taylor** (1:32)
Nope. It now has agent-based research. It can literally use Google Search to find sources on its own and verify information. It's like a little research assistant.

**Morgan** (1:45)
But how accurate is it? We know how hallucinations go when AI searches the web.
Did they share any test results?

**Taylor** (1:52)
Yes. According to the decoder in their internal tests, this new agentic system actually beat the previous version up to 78.2 percent of the time. That's a huge jump.

**Morgan** (2:06)
That is a significant margin. But I wonder about the cost and latency.
Running a full virtual machine for code execution isn't cheap or instant.

**Taylor** (2:16)
True, but since it's running on Gemini 3.5 Flash, it should be pretty snappy. I just love that the barrier to advanced data analysis is getting so low.

**Morgan** (2:27)
It's definitely a step towards true agentic workflows. Let's see if it actually holds up when thousands of users start throwing messy data at it.

**Taylor** (2:36)
I'm already planning to upload my giant spreadsheet of podcast metrics. I want to see if it can build me some cool charts automatically.

**Morgan** (2:45)
Just make sure it doesn't hallucinate our listener numbers, Taylor. We need hard facts here.

**Taylor** (2:52)
Deal.
Speaking of agents, there is this awesome roundup on MarkTechPost about how AI coding agents are completely dominating software development in 2026 Like, developers aren't even coding anymore.

**Morgan** (3:09)
Wait, hold on. Not coding anymore is a massive stretch, Taylor. They are definitely still coding, just at a higher level of abstraction.

**Taylor** (3:18)
Well, yeah. But the tools are insane now.
We're talking about Atoms, Devin, Windsurf, Cursor, and Warp. They don't just suggest lines. They plan tasks and edit across files.

**Morgan** (3:34)
Right. They can open pull requests and even ship to production, but that actually terrifies me a bit. Who is auditing all of this auto-generated code?

**Taylor** (3:44)
I mean, the engineers are still supervising, but instead of typing out repetitive boilerplate, they just describe what they want and the agent does the grunt work. It's so efficient.

**Morgan** (3:57)
I agree, it saves time, but we've seen studies showing that AI-generated code can introduce subtle security vulnerabilities. It's like having an eager junior developer who never sleeps.

**Taylor** (4:09)
Ha, totally. But the article points out that no single tool fits every need. Like, Cursor is great for inline editing, while Devin is more of a full autonomous agent.

**Morgan** (4:21)
That's a fair point. Choosing the right tool for the job is key, but I still think we need rigorous testing frameworks before letting these agents deploy code autonomously.

**Taylor** (4:33)
Oh, absolutely. But the speed of development is just going through the roof. It's a crazy time to be a software engineer, dude.

**Morgan** (4:42)
Or a very stressful time, depending on how much you trust your AI copilot. But let's move on before I start ranting about legacy code bases.

**Taylor** (4:50)
Ha, ha, ha.
I know how much you love debugging old code. But seriously, the shift from typing code to orchestrating agents is happening so fast.

**Morgan** (5:01)
It really is. The role of the programmer is being redefined in real time. Let's just hope the guardrails are up to par.

**Taylor** (5:09)
Well, speaking of guardrails, that brings us to our next story. I saw in TechCrunch that cybersecurity researchers are actually super annoyed with Anthropic's new model, Fable.