If people look back at the mid‑2020s and ask, “When did we have to get serious about the singularity?”, the answer might well be: right about now. AI systems are racing ahead in research labs, regulators are stumbling to keep up, and the tech giants are quietly tuning the economic engines that decide what billions of us see and buy every day.
This piece digs into two connected stories:
- Why fears of AI-driven mass unemployment are missing something fundamental about humans.
- How Facebook’s Kunlun recommender reveals a new, predictable way to pour compute into AI systems that directly control attention and ad money at global scale.
The Human Edge: Why Some Work Stays Human Even When Machines Can Do It
Economist Adam Ozimek recently resurfaced a basic but uncomfortable fact for AI doomers: we’ve had the technology to automate a surprising number of tasks for years—and yet a lot of those tasks are still done by people.
Think about it:
- We’ve had high-quality recorded music for decades, but live shows still sell out.
- We can order food via apps, but people still pay extra to be waited on in restaurants.
- Online travel sites exist, yet some travelers still seek out human travel planners for complex trips.
These aren’t failures of automation. They’re reminders that some services are bought not just for the outcome, but for the experience of dealing with another human.
The “Human Touch” as a Product People Pay For
Ozimek frames this as demand for what he calls the human touch: the added value that comes from live interaction, nuance, presence, and the feeling that “someone like me” is on the other side.
Economists have a dry term for it: a normal good. That’s something people demand more of as their income rises. When people get richer, they don’t just buy more stuff—they buy better, more curated, more personal versions of stuff.
Applied to work, that means:
- As societies get wealthier, demand for human-delivered experiences tends to grow.
- Automation doesn’t just erase jobs; it can shift demand toward more human-intensive services.
Examples that fit this pattern:
- High-end dining: same calories as takeout, totally different demand curve because of service, atmosphere, and status.
- Personal coaching and therapy: you could read a book or use an app, but many still pay for a human presence.
- Luxury retail and concierge services: the product is partly the feeling of being personally looked after.
A Different Path Through the AI Revolution
If we project forward to more capable AIs—systems that can draft legal memos, generate code, negotiate contracts, or even prove math theorems—it’s tempting to imagine a straight-line story: machines get better, humans get sidelined.
But there’s another plausible trajectory:
- AI wipes out or compresses the value of routine, standardized tasks.
- Economic growth from those gains, plus sane policy, boosts overall spending power.
- Richer societies start buying more human-centric experiences: customized, intimate, live, or status-laden work that is about being human in the loop.
In that world, we don’t just get AI baristas and AI customer support; we also get:
- A surge of human artisans who lean into authenticity and craftsmanship.
- Upgraded versions of existing professions—chefs, coaches, performers, guides—who use AI as a backstage tool while selling the front-stage human presence.
- Entirely new kinds of work that are hard to name today, the same way “social media manager” would have sounded absurd in 1995.
The catch? For this to not turn into a two-tier dystopia, governments and firms would need to consciously push for:
- Labor protections that stop human-touch jobs from becoming gig-economy serfdom.
- Policy that shares AI-driven productivity gains broadly enough that more people can actually afford high-touch services.
AI may change what we do, but Ozimek’s point is that humans don’t just disappear from the economic map. They cluster where other humans still want them.
Inside Kunlun: How Facebook Is Industrializing Recommendation AI
While economists argue over the “end of work”, Meta (Facebook’s parent company) is focused on something more immediate: how to more efficiently decide what you see next.
The company recently detailed Kunlun, a new recommendation system designed to be dramatically more compute-efficient than its predecessors—and, crucially, to scale in a predictable way as Meta dumps more GPU power into it.
Recommendation engines like Kunlun are not side projects. They are:
- The cash machines behind targeted ads.
- The attention routers deciding which posts, videos, and products billions of people encounter.
Why Recommenders Are Not Just “LLMs with Likes”
We’ve gotten used to hearing about scaling laws for large language models (LLMs) like ChatGPT and Claude: pour in more data and compute, and you can roughly predict how much your loss will drop or your accuracy will climb.
Recommender systems live in a different universe:
- They have to juggle sequential behaviors (what you clicked over time) with non-sequential context (who you are, what device you use, what time it is).
- Their inputs are highly heterogeneous: tiny embedding vectors, irregular data layouts, and operations that slam repeatedly into memory bandwidth limits.
The net result: they often waste hardware. A key metric here is Model FLOPs Utilization (MFU)—how much of your theoretical floating-point horsepower the model actually uses.
Typical numbers:
- LLMs: often reach 40–60% MFU on modern GPUs.
- Legacy recommenders: sometimes limp along at 3–15% MFU, leaving vast amounts of compute on the table.
Kunlun is Meta’s attempt to close that gap.
What Kunlun Actually Does Differently
Under the hood, Kunlun is a heavily optimized recommendation stack designed to make GPU hardware do real work instead of stalling on memory and awkward tensor shapes.
Two major architectural pieces stand out:
1. Kunlun Transformer Block
This component handles context-aware sequence modeling. Instead of treating your past behavior like a flat list of clicks, the system processes it with a transformer-style block adapted to recommendation:
- Multi-head self-attention to track relationships across a user’s interaction history.
- GDPA-enhanced personalized feed-forward networks—think of these as tuned MLP layers that adapt to individual users or contexts rather than being purely global.
The goal is to squeeze richer signal out of the same behavioral trail: what you did, when you did it, and in what kind of session.
2. Kunlun Interaction Block
If the transformer block mines temporal patterns, the interaction block handles information exchange between different kinds of features:
- Personalized weight generation: the model can adjust how heavily it leans on certain features for certain users.
- Hierarchical sequence summarization: instead of pushing an entire clickstream through every layer, it builds progressively higher-level summaries.
- Global feature interaction: user-level, item-level, and context-level data can talk to each other in a structured way.
In plain language: Kunlun is engineered to stop wasting GPU time on awkward, memory-bound plumbing, and to spend more of it doing the actual math that improves predictions.
The payoff Meta reports: MFU jumps from around 17% to 37% on NVIDIA B200 GPUs—a massive efficiency gain in a world where GPU time is scarce and expensive.
The Real Prize: A Scaling Law for Attention and Revenue
The most important part of Meta’s Kunlun work isn’t just raw speed; it’s predictability.
With LLMs, researchers discovered that if you double or triple compute and data in the right ratio, model quality improves along fairly smooth power-law curves. That lets companies budget billions in GPU spend with some confidence about what they’ll get back.
Meta is claiming something similar for Kunlun: a scaling law not over text loss, but over a metric called normalized entropy (NE), which tracks how well the system is modeling the uncertainty in user behavior.
They report that:
- As you feed Kunlun more training compute (measured in gigaflops), NE improvements follow a predictable pattern.
- You can also map improvements in NE to other business-relevant metrics, effectively building a bridge from GPU hours to ad performance.
Once you have that bridge, you’re no longer just making educated guesses about model size. You can treat compute as an investment with a rough expected rate of return in clicks, dwell time, and revenue.
Why This Matters Beyond Meta’s Balance Sheet
Put these threads together and you get a cleaner picture of the real frontier we’re crossing:
- On one side, AI skepticism about mass unemployment reminds us that humans are not purely cost-minimizing robots. People will pay for other people.
- On the other, hyper-optimized recommendation systems like Kunlun give firms a precise dial for shaping what those people see, want, and buy.
Two broad implications:
1. The Rise of Human-Heavy Work That Is Still Algorithmically Mediated
Even if we see a boom in human-to-human services—live performance, bespoke coaching, high-touch hospitality—discovery and demand will still be channeled through AI recommenders.
The human touch might be the product, but algorithmic curation becomes the gatekeeper. If systems like Kunlun can be tuned like financial instruments, then attention allocation becomes as engineered as high-frequency trading.
2. We’re Quietly Locking In Technical Decisions About the Future
Highly capable models proving math theorems or generating code grab headlines, but industrial systems like Kunlun are where AI’s social impact is being operationalized day by day.
Questions that follow naturally:
- How transparent should scaling laws for recommender systems be, given they map near-directly to monetization and influence?
- Should regulators treat ad-targeting compute and frontier model compute differently, or are they converging on the same core capability: learning to model and shape human behavior at scale?
- As we debate singularity timelines and superintelligence, are we paying enough attention to the mundane superpowers already in production—systems that don’t pass philosophy exams but quietly steer billions of micro-decisions?
2026 might not be the year we cross a visible singularity line. But it is clearly a year where companies like Meta start treating AI scaling as a well-understood economic lever, not a research gamble—and where economists remind us that even in a world of powerful machines, humans themselves remain a premium, not just a cost.
The next decade will be defined by how we negotiate that tension: humans as irreplaceable experiences versus humans as data points to be modeled and nudged. The technology is moving fast. Our institutions and norms need to catch up just as quickly.