AI Digest, Oct 31: Google AI Mode Revenue Hit, China's AI-Built AI, API Key Hijacking

0:02PM Lead ProThis is PM Lead Pro. Today's digest covers Google's admission about AI Mode, China's push to automate AI development, and a spike in hijacked API keys. Worth a listen if you run AI deployments.

0:15Imagine inventing a working teleportation device, right?

0:18But then you realize your company's entire revenue relies on selling airline tickets.

0:24Yeah, that is a classic innovator's dilemma, but on steroids.

0:27Exactly. And that is the exact scenario we are looking at in today's deep dive.

0:32We are pulling from sources like Latentspace, Layphone, and Interconnects to explore Google's latest financial reality check, the Black October cyber fallout, and how Chinese labs are pivoting to AI that builds AI.

0:46Right. So starting with Google, it's an incredibly clear example of a technology cannibalizing its own host.

0:53According to Latentspace, Google finally admitted that queries using AI mode actually lower their search monetization.

0:59Because, you know, their traditional model relies on you clicking links and seeing ads.

1:03Exactly. If the AI just gives you the perfect summary directly, well, you never click the source.

1:09Which is great for the user, but terrible for the business.

1:13And that simple admission wiped out $500 billion in value, triggering a historic 15% stock plunge.

1:19Oh, wow.

1:20Yeah. And it dragged heavy tech infrastructure hitters like CoreWeave down like 60%.

1:26And I mean, the core of that problem isn't just lost ad revenue.

1:29and so on It that running these massive AI models for every single search is just computationally exhausting And super expensive Right Which is exactly why Chinese labs are abandoning that brute force approach to raw power

1:43According to Lifone, they are attacking this exact compute bottleneck through something called recursive self-improvement, or RSI.

1:49So instead of trying to breed a smarter racehorse, they are building a fully automated AI-driven racetrack that trains the horses 24-7.

1:58That's a great way to put it. They are using AI to automate the AI factory itself rather than directly altering the model's core weights.

2:06Right.

2:07Look at the numbers. Jippoo's GLM 5.3 infra agent just built an entire production system across 100,000 domestic chips in two weeks.

2:16Wait, two weeks? That's insane.

2:18I know. And DeepSeek's DSAC is spinning up 3 million sandboxes a day just to test AI agents.

2:24Oh wait, I have to ask, why avoid altering the model weights directly? Is that just like fear of a hallucination loop?

2:31Yeah, it's a structural limitation. If a model tweaks its own core logic unsupervised, you risk a catastrophic system collapse.

2:38It just loses its grounding entirely.

2:40Exactly. By engineering robust infrastructure instead, they safely bypass their domestic compute limits without that risk.

2:47PM Lead ProFor a company owner: the conversation says AI summaries lower search monetization because users stop clicking links. If your revenue depends on clicks, ask how AI answers affect that flow.

3:00Makes sense But you know the moment you start deploying millions of these autonomous agents and handing them the keys to go do tasks on our behalf you open up a massive vulnerability Oh definitely We are in the middle of a cyber fallout right now that platforms like NorthFlask are calling Black October

3:16The spike in hijacked AI API keys is staggering. But I have to ask, are open weight models the primary villain here?

3:23Like, is this happening because anyone can download the architecture and poke holes in it?

3:27Well, you'd think so, but data from interconnects shows closed models are heavily involved in these breaches too.

3:33Really? So it's not just an open source problem?

3:36Not at all. Hackers aren't necessarily attacking the model's architecture. They are just stealing the agent's wallet.

3:42Right, because we've handed API keys to these AI agents to execute tasks.

3:46Exactly. We're moving to an internet mediated by these bots, and our infrastructure security just hasn't caught up.

3:52That completely changes how we need to think about deploying these tools.

3:56PM Lead ProFor team leads: the hosts say attackers can steal the agent's wallet without necessarily attacking the model's architecture. Check your billing logs and rotate user-provided API secrets, so stale keys don't sit in your workflows.

4:09Which brings us to our model radar.

4:12We have two new releases taking opposite approaches to these exact deployment challenges.

4:17First is Mistral Large 4, affectionately dubbed LeShonk.

4:22LeShonk, I love that.

4:24And it's a massive one trillion parameter model, but it uses a mixture of experts architecture.

4:28Right.

4:29So instead of firing up the whole one trillion parameter brain for every single prompt,

4:33and so on It only wakes up the specific 52 billion parameter expert needed for that exact question Which saves massive compute and allows them to offer a staggering 1 million token context window

4:46It's huge. Just be warned, you still need an 8xB200 node just for fine-tuning.

4:51Okay, so heavy on the hardware.

4:53But on the total opposite end, there's NVIDIA's Nemetron 3.5 Bolt E4B.

4:58Right, which is a tiny, highly specialized 4GB model.

5:02Yeah, and it completely abandons broad reasoning capabilities so it can excel purely at rapid character-driven dialogue.

5:09And the best part is, it runs easily on a single local RTX workstation.

5:14Which makes it perfect for low-latency local agents.

5:16But hey, if you are managing any of these AI deployments we just discussed, listen closely.

5:22Check your billing logs the minute you finish this deep dive.

5:25Seriously. With credential hijacking spiking globally this Black October,

5:29You need to rotate your user-provided API secrets immediately.

5:33You cannot afford to leave stale keys sitting in your workflow.

5:36Consider that your insider warning for the week.

5:39And to leave you with one final thought to mull over today.

5:43We talked about DeepSeek generating 3 million of its own testing sandboxes every single day to iterate and improve.

5:50Yeah.

5:51How long until that system realizes that you, the human engineer, are the slowest bottleneck in its evolutionary loop?

5:58PM Lead ProFind more at pmlead.pro: the transcript and sources are on the episode page. Question for you: when did your team last rotate the API keys your agents use?