AI Digest, Oct 14: SaaS Bond Bet, Agent Governance, Edge Models

0:02PM Lead ProThis is PM Lead Pro. Today's digest follows the money behind AI agents, from $108 billion in bonds to on-device models and robots. It's a short listen if you run a team or a company adopting AI.

0:19So Wall Street just quietly bet $108 billion that the software industry, as you know it, is about to die.

0:28Welcome to today's Deep Dive. We are tracking that money today.

0:31Yeah, we're looking at this massive capital shift.

0:35It's pulling AI down from those distant cloud servers and, you know, injecting it straight into your pocket and even into physical robots.

0:42Right. It's basically a SaaS-pocalypse.

0:44Let's unpack the numbers from the Center for Economic Policy Research.

0:47Well, the numbers are staggering. Hyperscalers, so the tech giants building our AI infrastructure,

0:53they issued $108 billion in corporate bonds for 2025.

0:58And investors just devoured that debt, right?

1:00They really did. It actually caused hyperscaler bond yields to drop 42 basis points compared to

1:05traditional software firms. The market is terrified that generative AI will just obliterate

1:11legacy SaaS companies.

1:12I mean, it's like a historic infrastructure boom. Investors are throwing cheap capital

1:17are the companies laying the fiber optic cables and aggressively shorting the ones selling walkie-talkies.

1:24That is a great way to look at it.

1:25And that infrastructure money is instantly funding the next phase, which is advanced autonomous agents.

1:31But if you are leading a team or running a company, you know, you can't just hand the keys to an AI without some serious trust.

1:38Which perfectly explains the latest data in the agent report.

1:40So, Naus Research just raised $90 million, but their core product isn't actually their open-source Hermes agent.

1:47Oh right, they're selling the governance and privacy tools wrapped around it.

1:50Exactly Because at the same time Google DeepMind Gemini for Argon model is proving that agents can actually defend enterprise systems It just hit 68 on the CWE bench Let clarify CWE bench for a second That a major cyber defense benchmark and so on

2:25aren't buying the AI engine, they are paying for the seatbelt and the insurance policy.

2:30Oh, I see. You can't just unleash a free, unmonitored agent into a proprietary corporate network.

2:36Right. The real product being monetized here is the auditability and data privacy.

2:41You know, the stuff that prevents a catastrophic leak.

2:44Okay, so if you're paying for five different SaaS tools right now, the agent replaces the tools,

2:50but you need a new layer of security to manage the agent.

2:53PM Lead ProPM Lead Pro note: Nous Research raised $90 million, and the hosts say the product is the governance and privacy tooling around its free agent. When you adopt agents, budget for auditability and data privacy, not just the model.

3:09Precisely. And these agents aren't just getting safer, they're shrinking enough to leave the cloud entirely. Like, Liquid AI just dropped their LFM 2.5 model.

3:19Which hits, what, 184 tokens a second on a standard Apple laptop?

3:23Yeah, the technical achievement there is how it manages its processing power. It has a large capacity, but it uses just 2 billion active parameters at any given moment.

3:33Wait, so it only wakes up the specific parts of the neural network it needs for that exact millisecond?

3:40Exactly. That's how you get blistering speed on local hardware without instantly melting your battery.

3:46And looking at a recent brief on Chinese developments, this push for edge compute is already triggering a massive hardware shift.

3:53There a startup that just launched the StepX Neo Oh right They marketing it as an agent phone which is fascinating when you compare it to Western agents Yeah Western agents are kind of like a backseat driver poking a stick at the steering wheel

4:05They awkwardly try to click your UI by simulating mouse movements.

4:09But the Chinese StepAOS integrates their agent, Amu, directly into the operating system.

4:15It's like giving the AI direct access to the car's transmission.

4:18It doesn't hover over your apps. It literally is the device.

4:21And giving an AI native control over an OS is the crucial stepping stone to having AI operate in the physical world.

4:28PM Lead ProPM Lead Pro note: agents are leaving the cloud. One model runs at 184 tokens a second on a laptop, and a Chinese phone builds the agent into the operating system. Ask where your agents run and who controls them.

4:44Latent Space has been tracking this shift toward world action models.

4:47Like Dynatube, right?

4:48Yeah. Instead of generating text, Dynatube predicts future video frames.

4:53Then it outputs the physical joint movements a robot needs to navigate that space.

4:57But humans rely almost entirely on vision to navigate.

5:01Are pixels really not enough for a robot?

5:03I mean, I'm looking at this Chinese startup, PoXiao.

5:06The ones building the C-Phys Hub, yeah.

5:08Right. They are going way beyond video.

5:12They use wearables to capture human tactile data and full-body physics.

5:17Why go through the massive engineering headache of building tactile wearables just to teach a robot?

5:22Because vision is actually incredibly redundant.

5:25The pixels in a video don't change much from one frame to the next.

5:29Oh, but tactile feedback is densely packed with crucial data.

5:32Exactly.

5:32Exactly. If you want genuine autonomy in the physical world, a robot needs to know how things feel. The friction of a cup, the weight distribution of a box.

5:43So touch provides that missing physics layer that video simply cannot see.

5:48Yeah like the physical resistance of a door handle for instance That need for physical interaction is driving a totally new breed of models Which brings us to our model radar for today We have three models dropping that are explicitly designed for these complex action

6:03tasks.

6:04First up is MetaMuse Wonderful, stepping straight into that elite Clawed 5.5 tier of frontier

6:10models.

6:11And then we have JevGPT.

6:13It introduces this thinking fast and slow capability for autonomous research.

6:17Which is wild because it actually mimics human cognition.

6:20It uses a low-parameter, fast neural pathway for immediate, simple reactions.

6:25But then it routes complex multi-step research tasks to a secondary, high-compute reasoning

6:30loop.

6:31Exactly.

6:32And finally, there's Decision 3.0.

6:35These are edge-optimized models ranging from 0.6 to 27 billion parameters.

6:40The kicker is how they are trained.

6:42They are trained exclusively to select options and trigger actions.

6:46They skip text generation entirely to save massive amounts of compute for local devices.

6:51Which really brings us to the most practical takeaway for you today.

6:54Audit your company's software stack right now.

6:57Yeah, if your daily operations rely heavily on traditional SaaS platforms,

7:01you need to prepare for massive disruption.

7:04These agentic workflows are moving fast.

7:07They really are coming for those legacy moats.

7:09So as you audit that stack, I want to leave you with a final provocative thought.

7:13If an AI agent lives natively in your phone's operating system, making decisions and taking actions on your behalf in the real world,

7:23Who legally owns the consequences of those digital actions?

7:27Exactly. If the train crashes, do you blame the conductor or the company that laid the tracks?

7:32Thanks for joining us, and we will catch you on the next Deep Dive.

7:36PM Lead ProThis was PM Lead Pro. Find more at pmlead.pro, with the transcript and sources on the episode page. If an agent acts on your behalf, who should own the consequences?