AI Digest, Oct 26: Agent Cloud, China's Cheap Compute, Old CNN Beats LLMs

0:02PM Lead ProThis is PM Lead Pro. Today's digest covers AI agents that negotiate purchases, cheaper models from China, and a decade-old network that outscored frontier models on a medical task.

0:14So imagine you're buying a pair of shoes online, but behind the scenes, your personal AI assistant is locked in this high-speed, completely invisible, haggling war with the store's dynamic pricing algorithm.

0:25Right, which sounds totally like science fiction, but it is actually happening right now.

0:30Yeah, exactly. Welcome to a deep dive into today's frontier AI research curated specifically for you, the engineers and leaders actually adopting this stuff.

0:39Today we're looking at this booming agent economy, some really radical efficiency breakthroughs out of China, and, well, why a decade-old AI just completely embarrassed today's most expensive frontier models.

0:51I mean, we are seeing a massive architectural shift right now. Just recently, a startup called Daytona raised $60 billion at a $1 billion valuation just to build what they're calling an agent cloud.

1:03Okay, wait, let's unpack that for a second. What exactly is an agent cloud, and, like, why is it worth a billion dollars?

1:09Well, think about the difference between a simple chatbot and an autonomous agent.

1:14A standard language model, it just generates text, right?

1:18Right, like ChatGPT or whatever.

1:20Exactly. But an agent actually executes tasks on your behalf, like navigating a website or, as you said, negotiating a purchase.

1:27To do that safely, it needs a persistent, secure computer environment to run on, basically an isolated sandbox.

1:35And consumer shopping agents are already out there doing this, battling vendor algorithms.

1:38and many more I mean we the consumers might end up paying the exact same price for those shoes but now AI companies are taking a huge cut just to host the negotiation Yeah and that is the billion question

2:00Because if every single online transaction suddenly requires two opposing AIs haggling in the cloud, the global compute cost just becomes astronomical.

2:08It would bankrupt everyone.

2:10Right. So for this agent economy to actually scale, the underlying models have to become drastically cheaper to run.

2:16Which naturally leads us to the hardware side of this compute problem.

2:20Looking at recent data, Chinese developers are heavily driving costs to the floor right now.

2:25Kimi K3.1 Pro just took the number four spot globally for intelligence, leading all open weights.

2:31Yeah, and DeepSeek's new V4.1 Flash model is running at just 26 cents per task.

2:3626 cents.

2:37That's wild. It is.

2:39But while the pricing is aggressive, the real breakthrough here is architectural.

2:42A Chinese team at Rock Matrix just released a new model called EdgeRock 4B.

2:48And to understand why this matters, you have to look at how Western models operate.

2:52Okay, lay it on me.

2:53They typically demand massive memory overhead because of their KV cache,

2:57which is essentially the AI's short-term memory footprint, storing all the context of a conversation.

3:03Right, so if you're running millions of agents, that memory footprint adds up incredibly fast.

3:08Precisely. And EdgeRock completely turns that on its head.

3:11AI and so on Okay that is unbelievable speed But hold on if you restrict an AI short memory that aggressively doesn it just you know lose the plot Usually yeah I mean these are known as bounded state models and they typically lose complex reasoning capabilities because they simply can remember distant context

3:48Makes sense.

3:48But somehow, EdgeRock maintained an 82.9% on the AIME benchmark, which is a notoriously difficult math and reasoning test.

3:58Wow.

3:59So for local memory-constrained devices, this proves you just don't need a massive data center to run highly intelligent tasks.

4:06Exactly.

4:07Well, speaking of that, if massive data centers aren't always necessary for high-level reasoning, do we actually need massive frontier models for specialized tasks?

4:16PM Lead ProIf agents run your workflows, hosting them in isolated sandboxes becomes a cost line. The hosts' point: the agent economy only scales if the underlying models get much cheaper to run.

4:31Right, the specialization versus scale debate.

4:33Yeah, because a researcher, Dr. Magnus Lynch, recently tested this, and the results were shocking.

4:38He asked top models GPT-6 Astra and Claude Fable 5.1 to diagnose 200 skin cancer images.

4:46They scored an abysmal 42%.

4:48And then he ran Inception V3, which is a 10-year-old convolutional neural network.

4:53He fine-tuned it for just 20 minutes, and it scored 62%.

4:58A decade-old model beat the modern titans by 20 points.

5:02How does that even happen?

5:04It comes down to architecture, really.

5:05A convolutional neural network, or a CNN, is specifically built to process spatial image data, like finding patterns in a photo of a mole.

5:14Oh, I see.

5:15Yeah, whereas modern frontier models are built on text-based transformers.

5:20They are incredible generalists.

5:22But when you force a generalist transformer to do a specialist visual job you are basically paying a premium for worse performance Plus the CNN is so efficient it runs entirely locally in a phone browser and so on

5:54PM Lead ProFor a specialized task, test a small fine-tuned model before paying for a general frontier API. In the skin cancer test, a CNN tuned for twenty minutes scored sixty-two percent against forty-two.

6:08It's optimized strictly for taking actions and making choices rather than just chatting, right?

6:12Exactly. It's built specifically for that agent economy we were discussing earlier.

6:17And then from Google, we have Gemma 4 E2B.

6:20Right, which is a tiny model built for maximum compute efficiency, really targeting mobile devices and Internet of Things hardware.

6:27So it's all about pushing intelligence to the very edge of the network.

6:31So your smart devices don't always have to phone home to a cloud server to make a simple decision.

6:36Spot on.

6:37Which gives us our practical takeaway for today.

6:40Before you default to paying for an expensive general purpose frontier API for your company's workflows, test a small, fine-tuned specialist model first.

6:50It might just be faster, cheaper, and far more accurate.

6:54Function over flash. Every single time.

6:57Definitely. And I want to leave you with a final thought, circling back to our opening.

7:01If autonomous agents soon do all of our shopping and negotiating in these invisible sandboxes,

7:08will consumer advertising become completely obsolete?

7:11Like, will it be replaced entirely by marketing aimed directly at algorithms?

7:16Ooh, that's a wild thought to end on.

7:18Right.

7:19Think about that next time you browse the web.

7:22Until next time.

7:23PM Lead ProMore at pmlead.pro: the full transcript and sources are on the episode page. If agents negotiate for us, who will advertising be aimed at: people, or algorithms?