Scale Without Sluggishness

A couple of months ago, WSJ reported that Mark Zuckerberg has been wondering whether Meta should start building a plethora of apps. From WSJ:

“So, like, Chris and I have been talking about ‘all right, well can we build 50 new apps?’ Like, yeah probably. But we probably should start by doing a few before we just, like, ramp up trying to do 50 all at once.”

Since roughly the time of that quote, Meta has shipped or started building a cluster of standalone apps: Forum, a Reddit-style app built on top of Facebook Groups with an AI tab called "Ask"; Instants; Facebook Creator Studio (a stand-alone AI companion app for creators); and now potentially a new prediction market app named Arena.

Meta has been a fast-follower forever. For most of the company's history, "try lots of apps" ran into two constraints: the engineering cost of building and the opportunity cost of pointing scarce talent at long shots. Once building got cheap in the age of AI, the binding constraints become distribution and knowing what to build. Of course, Meta has more distribution than almost anyone else, and looking at their recent cadence of releasing new apps, I wonder if they are just trying to gun at anything that appears to have high engagement. Once you have distribution, the call options on taking such shots can be pretty cheap.

I think Meta also got particularly encouraged by the success of Threads. While fintwit has never quite moved to Threads and hence most investors can be particularly prone to have a negative bias to Threads, the reality is they reached 500 million Monthly Active Users (MAU) just three years after the launch. I have noticed many people decry the fact that Meta may “game” these numbers by pushing Threads content onto Facebook or Instagram, but that’s more of a feature than bug in Meta’s case. Imagine, for example, the potential prediction market app. If it is indeed launched, I do expect Meta to leverage their existing distribution to show the most popular questions/polls directly on Facebook or Instagram feed. If a user engages with such content frequently, they will simply see those questions more and if a user dislikes such content, my guess is they will have the option to hide those polls from their feed. And if someone really, really likes prediction markets, they will just download the standalone app. I’m just giving prediction market as an example to make a point; I don’t have a particular view about the probability of Meta’s success there. In fact, I think Meta doesn’t have a rigid view either; what they see is prediction markets are inherently quite engaging to a lot of users and they want to offer something that elicits incremental engagement on their apps.

Meta’s plethora of bets are particularly interesting in the context of open web potentially decaying as a place to find things. With the rise of AI slop, zero-click search, and publisher economics in freefall, it is not hard to see why walled gardens who have much better understanding of your interest and engagement can offer a better, personalized “internet” for you. Of course, Meta is historically fairly bad at standalone apps. The graveyard is long: Poke, Slingshot, Paper, Rooms, Moments, Lifestage, Bonfire, Lasso, Bulletin, Hobbi. You may not have heard about any of these apps, but these are actually some of the standalone apps Meta tried and failed to gain much traction. In fact, Arena itself is a rebuild of “Forecast”, the prediction-market app that Arena is now a rebuild of. The real wins (Stories, Reels, and even Threads) were features grafted onto graphs that already existed. So the broader strategy of building a new “internet” always existed at Meta, but executing parts of that strategy just became lot cheaper. Anything that will get some traction will inevitably get lot more resources to scale it further. I would also highlight that building standalone apps is not just cheaper than it ever was before, but the way these efforts are currently undertaken have gone through a tangible and massive change post-AI.

I was intellectually aware of such a rapid evolution, but as a solo entrepreneur who doesn’t have to manage hundreds of people, admittedly I perhaps underestimated the extent of this transformation in some of these tech companies. Listening to this podcast on “How Meta Is Reinventing Product Management” helped me appreciate from a distance that “work” has fundamentally changed forever in a tech company such as Meta. Jagjit Chawla, who is the VP of Product Management- FB Feed, Reels & Search at Meta, explained how things have evolved over the last couple of years (emphasis mine):

“If you’re an IC (individual contributor) PM on the team, your prior... two years ago... would be you would sit with your engineers, sit with your UXR folks, come up with the ideas, write a detailed PRD which basically describes the problem that you’re trying to solve, the feature set you would want to focus on, what hypothesis you’re testing, and so on and so forth. These days that has changed to a one-paragraph description of what you’re trying to solve coupled with a prototype to show what you’re actually going to do. And for ML AI heavy areas, coupled with an eval set as well, because ML systems are unpredictable. You want to make sure that if you try 10 times, what type of outputs it gives us, you have to figure out precision, recall, all of those good stuff.

But previously, the product management was you sit on your desk and figure out proactively what could be the things that you should be building. And of course, even at that time, your engineers start implementing and you will be in a state where, ‘Oh, we missed thinking about this stuff,’ and you will fill the holes of what product management looked like. Now you sit next to your engineer on the eval and you agree or disagree whether this is a good outcome or a bad outcome. That is what real product management looks like right now, and the pace at which you can do that with a small set of people—a pod of five people let’s say sitting together and doing this—is very different than what it was before.

As a product executive, especially in areas like mine which are really robust, where your business’s attention to detail is important, your life as an executive is to read through a ton of material as you run product reviews. And that material is produced by your team. The team is spending a bunch of time producing those materials. Now all of that is flipped to the head. We are at a point where the ground truth—source of truth for software companies—is code written. And these companies... when you have a large 1,000-person org, you run what I call a compression algorithm. The engineer knows what they are checking in, their manager summarizes it, and by the time it comes to me, I get a small sliver of information for that particular project that I need to make a decision on. And that is still fine, that is old school, but I have 50 projects. So my context window still fills up with that small piece of information. That is the old two-year method.

Now, my AI agents scour through all the diffs that are checked in last night, all the emails and chats that I have received or responded to, all the docs that I need to review or have reviewed, and summarizes for me what exactly is the team making progress on and where are the gaps. And that is the new way to be a product executive where the compression algorithm is thrown out and there is a large amount of transparency in the system.”

As you can imagine, you can get lot more work done but still the “bottleneck” these days can be the managers themselves if they do not build a system to “scale” themselves to respond to the increased productivity of the team he/she is managing. Chawla explained:

“…culturally Meta is a very written culture. It helps that we have a version of internal Facebook, if I can call it—it’s called Workplace—where every detail in terms of experiments we ran, what ended up happening, what did we learn, is all documented really well. So that allows the AI tools to scour through all the information, let’s say on a nightly run or you can do it on an hourly run as well, to extract the right information.

Now, to your question on how this is built... the culture enables because a lot of your AI output is as good as the data that you provide it. And Meta has a lot of rich data available both in terms of written information and dashboards and access to analytics tools and so on and so forth that can actually collate things really well. There is basic infrastructure inside the company, and a bunch of AI tools are recommended and collate together, but a bunch of this is personalized built by an N of 1, which is me, for the needs that suit my style of reviews. Like what I do in product reviews is very different from just even my peer on engineering does. And we now have a bot for example where you could query what I’m going to say for a given document, and what is my end partner going to say on a given document. So it’s customized, it’s built by me using tools that are accessible.

The other question you had was how do you scale yourself? The flip side of this thing is lots of data is now visible to me, and sifting through the signal from the noise is another thing. And that’s where AI also helps, where iteratively working through the tools is the right answer here. Many a times it would hallucinate, in some cases, and many a times it will not get the right priority. So iterating and fine-tuning your approach.

So for example, on a daily basis, I would review maybe five documents. And my AI system every night will then scour through the comments that I actually left on that document, compare that to the comments it would have suggested, and fine-tunes its approach. Now that I’ve done it for over six months, it is quite accurate, but it was not to begin with. So it takes effort to fine-tune your approach, but it has gotten to a point where it’s pretty good.”

Chawla also shared a specific example to explain how the pace of change can directly have an impact on how Meta is making their product decisions (slightly edited for clarity:

“Take a question — and I'm making this one up — say it's about Reels distribution: of all those views, how many come from your connected audience versus your unconnected audience, people who follow you versus people who don't? Previously, if that came up in a live review and I thought, "our answer here might change if we knew this number," the team would take the ask to the data science team, they'd run an analysis, and it would take at least 24 hours to get a reasonable answer back. Now the tooling is at a place where I can fire that question at an analytics agent — a bunch of stuff our data science teams have built — and in five minutes it comes back with an answer. And it's grounded in truth, because I can say, "Here's the data scientist who works in this area; look at the queries they've authored and use those as ground truth before making up your own data sources." So the agent knows that the person who owns this actually queries these tables, joins them this way, has written these queries before — and it crafts the right answer from that. Five minutes, and I have it. We get the answer, make the decision live in the review, and move on. These things are happening now that were almost unimaginable two years ago.”

It can be tempting to look at the list of Meta’s past failures and dismiss this latest flurry of apps as a desperate grab for engagement. But that perhaps misses the fundamental reinvention of the organization itself. The traditional corporate “compression algorithm” where layers of management summarize and dilute the ground truth may no longer be the case in the age of AI. When a product executive can run a thousands-person org with the agility of a five-person startup pod, the very nature of corporate scale can change. By turning managers from bureaucratic bottlenecks into highly leveraged, agent-guided orchestrators of rapid experimentation, companies such as Meta can potentially scale multiple products without the sluggishness that typically beleaguer large organizations.

---

Subscribers get the daily journal and five+ years of Deep Dives, i.e. full-length analyses with financial models on 65+ companies. The daily is just how I think out loud between the Deep Dives!

---

Disclaimer: All posts on “MBI Deep Dives” are for informational purposes only. This is NOT a recommendation to buy or sell securities discussed. Please do your own work before investing your money.