AI Digest, Oct 18: Meta MuseSpark Open, PilotDeck Smart Routing, Solar and Gold

0:02PM Lead ProThis is PM Lead Pro. In this episode: Meta's new open-weight model, how to cut the cost of AI agents, and why energy and gold markets are shifting alongside the AI race.

0:14So imagine finding out that the top U.S. export right now, it isn't tech or weapons or cars.

0:20It's actually gold. And it is kind of quietly flowing straight to China.

0:26Yeah, it's wild.

0:26And at the exact same time, the U.S. AI open weight scene is basically waging this massive proxy war to undercut its own domestic rivals.

0:35Exactly.

0:35So welcome to today's deep dive. We are looking at a huge stack of reporting today from InterConnect,

0:42CEPR, Qubit AI, PV Magazine, and Numismatic News.

0:47It is a lot of ground to cover.

0:48It is. But our mission for you today is to unpack these new frontier AI models out of the U.S. and

0:54China, find out how you can slash your agentic AI costs, and explore how the massive energy

1:00demands of all this tech are just driving these tectonic shifts in global markets.

1:05So let's just jump right into the AI model race.

1:09Okay, so according to Interconnects, Meta just dropped this massive new model called MuseSpark Open.

1:14Right.

1:14And it boasts like 816 billion total parameters, but, and this is key, only 42 billion are active during generation.

1:22Wow, okay.

1:22Yeah, so this approach, it's often called a mixture of experts.

1:26It essentially means the AI isn't firing up its entire brain for every single prompt.

1:30It's just using what it needs?

1:32Exactly.

1:32Exactly. It only activates the specific pathways needed to answer your question, which saves

1:36just a tremendous amount of computing power. And it's a really big deal because it closes

1:40that performance gap we've been seeing with top tier Chinese models.

1:45Models like GLM 5.3 and Kimi K3?

1:48Yeah, exactly.

1:49So the US scene is reclaiming some ground here But I do have to push back a bit on Meta using the open label Oh totally I mean when you read the actual license for this thing it packed with revenue caps and much more

2:21Exactly. They are trying to commoditize the exact same technology that their competitors charge a premium for.

2:27But, you know, the real story here is that deploying an 816 billion parameter model, that's really meant for highly sophisticated workflows.

2:34Right.

2:35A model this heavy, it's built to act as an agent, basically executing complex, multi-step tasks over time rather than just, like, chatting with you.

2:43Which perfectly sets up our next source because agents are getting incredibly capable, but there's kind of a huge catch.

2:51Yeah, the cost.

2:52Right. So a new CEPR study showed that giving AI models agency, specifically letting them write, test, and self-revise their own econometric code for statistical modeling, it boosts their success rate from 79% to a massive 97%.

3:09Which is huge because you aren't just relying on the AI to get the math perfect on the very first try anymore.

3:15Right.

3:15You're giving it this feedback loop to spot its own errors and rewrite the code.

3:19But letting a flagship model endlessly loop and revise itself, that just drastically skyrockets your API costs.

3:26It's so expensive.

3:28PM Lead ProFor a team lead: letting a model write, test and revise its own code raised success from 79% to 97% in the CEPR study. The trade-off discussed is API cost from endless revision loops.

3:42Enter PilotDeck 2.0.

3:44So this is an open source agent OS developed by Tsinghua University And this was detailed over in Qubit AI Yeah this is fascinating It super cool It cuts those skyrocketing costs by 83 using what they call smart routing

3:58So think of it like a law firm, right?

4:01Okay.

4:01You have a senior partner, which is a massive expensive model like Claude's on it 4.6

4:05and they map out the high level legal strategy.

4:08Right, the big picture stuff.

4:09Exactly.

4:10Then the system automatically routes the tedious everyday execution tasks down to junior clerks.

4:16Which are the smaller, highly efficient models like Minimax M2.7.

4:22You got it. Because you don't pay $1,000 an hour senior partner to just make photocopies.

4:28No. You let the smaller models handle the heavy lifting of the execution.

4:31But, and this is the catch, running an endless army of these AI junior clerks, it still requires just staggering amounts of electricity and compute.

4:41Which directly drives the macro shifts we're seeing in global energy right now.

4:45PM Lead ProFor an owner adopting AI: PilotDeck routes planning to a large model and routine execution to smaller ones, and reports an 83% cost cut. Worth checking which of your tasks need the flagship model.

4:58Exactly.

4:59Because PV Magazine is reporting that global solar power just surpassed nuclear for the first time ever.

5:05They added a record 510 gigawatts in 2025.

5:09That is just a massive number.

5:10It really is.

5:11And concurrently, tying back to that hook from earlier, Numismatic News reports that gold is now the top U.S. export.

5:18Analysts are noting that it's heavily flowing to China to build non-dollar reserves.

5:22So how do these actually connect?

5:24Well, the AI arms race is basically forcing countries to completely rethink their power

5:28crits, right?

5:30Solar is highly decentralized and it's fast to deploy, which makes it perfect for these

5:35energy hungry data centers.

5:37But at the same time, funding this immense new physical infrastructure, it requires massive

5:42capital So countries like China are hoarding gold to de Wow Yeah they essentially trying to insulate their tech supply chains and their economies from the traditional centralized U financial system

5:56So looking strictly at the trade data, these two decentralization trends like in energy grids and global finance, they're just accelerating side by side.

6:05Yeah, completely in tandem.

6:06Okay, so here is the practical takeaway for you today, whether you manage software engineers or you're running a small business.

6:13You really need to adopt a smart routing mindset for your workflows.

6:17100%.

6:18You have to stop wasting your most expensive resources, whether that's flagship API tokens or, frankly, your most experienced employees, on simple formatting or execution tasks.

6:28Yeah, because efficiency is basically the only way to scale intelligence.

6:32So true.

6:32Okay, for our model radar today, keep two things on your screen. First, Meta's Mew Spark Open. It is natively multimodal, built for complex, agentic work, and designed to run locally on heavy-duty hardware.

6:45Right. And second is Pilot Deck 2.0 from Chuan. This introduces quote-unquote white-box memory, which is a huge breakthrough because it finally lets humans manually edit what the AI remembers about a project.

6:58That is so useful.

6:59It really is. It allows you to debug the AI's actual thought process while it just runs background tasks indefinitely.

7:06Okay, here's a final provocative thought to leave you with today.

7:10If agent OS platforms like PilotDeck can already autonomously route tasks to cheaper models to save money,

7:17how long until AI agents start managing their own cloud computing budgets?

7:21Oh, wow. Or like negotiating API prices directly with the providers.

7:26Exactly. Something to mull over. We'll see you next time.

7:29PM Lead ProMore at pmlead.pro: the transcript and sources are on the episode page. Question for you: which tasks in your workflow still go to your most expensive resource by default?