Google’s $270 Billion Gemini 3.5 Failure

**SPEAKER_1** (0:00)
The new LinkedIn Hiring Pro can't undo your last hire, the no show. Who set you back because they didn't show up on day one, or day two, or ever again? And now your product launch is delayed because you're talking to alien abduction podcasters to track them down.
But LinkedIn Hiring Pro can help make sure your next hire sticks. In fact, businesses who use LinkedIn are 24% less likely to reopen a role in the next 12 months. Hire right the first time with LinkedIn Hiring Pro. Post a free job today at linkedin.com/quality.

**SPEAKER_2** (0:30)
Alphabet lost roughly 270 billion dollars in market value over just two trading sessions following a Bloomberg report that Google is delaying its flagship Gemini 3.5 Pro AI model while OpenAI has already shipped their GPT 5.6 family and XAI has Grok 4.5 out.

**SPEAKER_3** (0:51)
I mean, a 270 billion dollar haircut is just difficult to contextualize.

**SPEAKER_2** (0:55)
Yeah, it's a massive number.

**SPEAKER_3** (0:56)
You're looking at the equivalent of erasing the entire market capitalization of a major global corporation in 48 hours. And internally, a financial hit of that magnitude creates this immediate pressure cooker for the engineering teams for upper management.

**SPEAKER_2** (1:09)
And externally, it severely weakens their leverage.

**SPEAKER_3** (1:12)
Exactly, because the stock drop is a direct reflection of the market's anxiety regarding those AI coding capabilities.

**SPEAKER_2** (1:17)
Because enterprise buyers are looking for systems that can write, test and deploy code autonomously.

**SPEAKER_3** (1:23)
Right, and if the market senses a delay in that specific function, they punish the stock immediately.

**SPEAKER_2** (1:27)
So if the company that practically invented modern AI architecture is struggling to ship its flagship model, what does that reveal about the invisible walls everyone else in the industry is about to hit?
If you're not subscribed yet, take a second and hit follow on whatever podcast app you're using. It helps us keep making this. We appreciate you being here.

**SPEAKER_3** (1:45)
Because we know the specific technical failures causing this delay, and they are not minor.

**SPEAKER_2** (1:52)
No, Google reportedly scrapped the original Gemini 3.5 Pro base model entirely.

**SPEAKER_3** (1:58)
Which is a huge decision.

**SPEAKER_2** (2:00)
It is. During early testing, engineers found structural failures in recursive tool calling. They found broken SVG generation, severe token efficiency issues, and just overarching problems, executing long horizon tasks.

**SPEAKER_3** (2:14)
And treating these issues as standard software bugs is a fundamental misunderstanding of how these models operate.

**SPEAKER_2** (2:21)
They did try to update the model data specifically to improve the AI coding, but the results just remained completely disappointing.

**SPEAKER_3** (2:27)
Because recursive tool calling is the foundational mechanic for autonomous agents.

**SPEAKER_2** (2:31)
Right. It's not an optional feature.

**SPEAKER_3** (2:32)
No. If you ask an AI to research a competitor, find their pricing, and put it in a spreadsheet, the AI has to use a search tool.

**SPEAKER_2** (2:40)
Yeah. And then evaluate the results, realize it needs to use a different search term based on those results, and then cull a spreadsheet tool to format the data.

**SPEAKER_3** (2:48)
That is recursive tool calling.

**SPEAKER_2** (2:50)
Exactly. If a model cannot reliably call a tool, evaluate the output, and recursively call another tool based on that new information.

**SPEAKER_3** (3:00)
It fails at being useful for any complex enterprise workflow.

**SPEAKER_2** (3:04)
It's useless.

**SPEAKER_3** (3:05)
And that is exactly the friction point for a developer. I mean, if you are building an application for a law firm, you cannot have an agent that successfully searches for a legal precedent, but then forgets how to format it into a brief.

**SPEAKER_2** (3:18)
No, you can't. And the broken SVG generation is another structural red flag.

**SPEAKER_3** (3:22)
Generating a scalable vector graphic sounds like a minor graphics glitch.

**SPEAKER_2** (3:26)
It sounds like it, yeah. But SVGs are not made of pixels, they are made of math. Right. To draw a simple circle as an SVG, the model has to write the mathematical coordinates and define the radius in code.

**SPEAKER_3** (3:37)
So it requires precise spatial reasoning.

**SPEAKER_2** (3:39)
Translated through syntax, exactly. If a model fails at SVG generation, its underlying reasoning architecture is flawed. You cannot patch that with a quick update.

**SPEAKER_3** (3:48)
No, these are structural limits of the current architecture they are running.

**SPEAKER_2** (3:52)
And the long-horizon task failures also speak directly to those structural limits. Yeah. A model has to maintain state and context over hundreds of sequential steps to complete a long-horizon task.

**SPEAKER_3** (4:04)
Think of it like baking a highly complex pastry.

**SPEAKER_2** (4:09)
Okay.

**SPEAKER_3** (4:09)
If the AI forgets at step 45 that it already added the eggs in step 3, the entire output is ruined.