Lately the story goes like this: Meta is in the press for something that sounds straight out of a near-future novel — employees asked to install software that watches nearly everything they do on their machines, all in the name of feeding the beast. Whether the details match your worst reading or a more benign internal pilot almost does not matter for the shape of the conversation. The image lands: total visibility, total capture, the company that already sits on social graphs now claiming the desktop too.
In parallel you have the corporate arc everyone already half-remembers. In 2025 Meta took a roughly fourteen-billion-dollar position in Scale AI — on the order of half the company — and pulled some of Scale’s leadership closer to Meta’s own AI efforts. Layoffs followed, as they often do when balance sheets get rearranged around a new story. The strategic rhyme is easy to spot: another play at replicating human expertise, the same phrase that has been selling boardrooms and keynote stages since “data is the new oil” wore thin.
I want to sit with that bundle of facts without forcing a single clean thesis. Some of it is genuinely about capability. Some of it is theater — and I do not mean theater as in “fake,” I mean theater as in staged narrative, where controlling the headline is nearly as valuable as the underlying technical win. If you work in communications or marketing inside one of these orgs, you are not decoration; you are part of the weapon. In the current AI race, looking like the company that will own expertise may matter as much as the incremental eval improvement your last release bought you. That is not a cynical aside about spin doctors. It is a labor-market fact: narrative work around AI is its own growth industry — specialist agencies, in-house strategy roles, crisis playbooks, and the endless content machinery that translates model releases into stories investors and recruits can repeat. The boom in that layer is as real as the boom in GPUs; it is just easier to sneer at because it does not ship a tarball.
The loop
The pattern is almost rhythmic. Meta says or does something that scans as dystopian. Commentators react. Someone publishes a longer piece — sometimes critical, sometimes explanatory — and the brand still occupies the center of the frame. Even austerity moves, like layoffs in a stretch where traditional macro indicators would normally read as “uncertain times” (if you still trust those indicators to describe this economy), get folded into the same narrative gravity. The company looks decisive, unsentimental, willing to do what the mission demands. Rinse and repeat.
I am not arguing that nothing real sits underneath. Large models really do get better when you pour money, data, and talent into them. I am arguing that the interpretive layer around those investments is doing more work than we admit when we treat every announcement as a literal map of what will happen next.
What Scale-style data can and cannot be
Here is where I get skeptical — not in the sense of calling anyone a fraud, but in the sense of asking what is actually being purchased when a giant buys a labeling company and the mythos of “captured human knowledge.”
Human-in-the-loop supervision is not snake oil. Preference data, rubric-based grading, safety tuning, domain-specific review — all of that moves models in ways pure self-supervision on web scrapes does not. There is real nuance in how Scale and its peers operate, and I am not interested in flattening that into a Twitter-length scam chart.
The doubt I keep returning to is older than LLMs, and it has a sharper name than “logs versus
vibes.” People call it the two clocks problem. Every serious organization runs
two timelines in parallel. There is a state clock: what is true right now
— the closed deal value, the ticket status set to resolved, the config line that today
reads timeout=30s. And there is an event clock: what happened, in
what order, with what reasoning — the negotiation before the number, the argument before the
status change, the fact that someone tripled the timeout from five seconds and why.
git blame tells you who touched the line; it does not resurrect the architectural
debates that made the change feel obvious afterward.
We have spent decades building trillion-dollar infrastructure for the state clock. Relational databases, CRMs, ticket queues, dashboards: all optimized for “what is true now.” The event clock is starved by comparison. State is easy to store because it overwrites cleanly. Events want to append; reasoning wants to live somewhere durable — and for most of history it did not live in software at all. It lived in heads, in Slack threads, in meetings nobody recorded. The CRM can say closed lost without capturing that you were the second choice and the winner had a feature you are shipping next quarter. The chart says switched to Drug B, not that Drug A worked until insurance stopped paying. The contract shows a sixty-day termination clause, not the push for thirty that you traded for a liability cap. When humans were the default reasoning layer, you could reconstruct the event clock on demand through conversation. Now we ask models to exercise judgment from precedent the systems never bothered to keep.
That is the structural complaint behind desktop capture fantasies: screen recording still mostly feeds the state clock — pixels, clicks, finished artifacts — unless you invest heavily in turning trajectories into durable events with semantics. And the problem compounds. Real systems are only partially observable: legacy corners, third-party black boxes, emergent behavior across services. There is no universal ontology for “customer” or “risk” across companies. Everything mutates daily. Most “knowledge management” fails because it ingests frozen documents — state printed on paper — instead of capturing living process. Human-in-the-loop labeling helps at the margin, but it is still often someone grading an end state, not the full event clock of how a skilled person navigated ambiguity. A plausible partial patch is to treat rich agent trajectories — runs that investigate, decide, and act inside messy real systems — as implicit maps of ontology discovered through use rather than specified up front. That direction is promising, early, and still hemmed in by the same observability and churn constraints.
Machine learning names a slice of the same gap. Imitation from observations alone means watching state transitions without the teacher’s internal policy and inferring what must have happened in between. Process supervision widens the aperture by training on intermediate steps — closer to respecting an event clock — but only where those steps can be made legible, scored, and scaled. Many inner stories still fit the same outer trail.
So when a narrative promises to “replicate human expertise” at civilization scale, I reach for a quieter prediction before the science-fiction one. The frontier that keeps moving is still, in large part, compute and data mass — more tokens, more flops, more careful curation of what gets counted as a good answer. The threshold people gesture at when they say “superintelligence” — whatever that even means on inspection — still looks less like a solved engineering ticket and more like a horizon we redraw every year. I do not think anyone has a convincing way to rebuild the event clock for whole organizations at fidelity the state clock enjoys. The marketing department will still name the product as if they have.
Why the story still wins
Meta does not need a philosophical resolution to keep pedaling the narrative. It needs enough capability to stay in the top tier of benchmarks, enough talent to credibly claim depth, and enough drama in the public telling that investors, recruits, and regulators all believe the company is inevitable. The Scale deal fits that bundle neatly: capital, expertise, a story about proprietary human signal, and a headline that makes rivals wonder what they are missing.
If some of the workplace-surveillance reporting is softer or messier than the scare copy suggests, that almost makes the point. The version that spreads is the one that keeps Meta’s AI ambitions welded to a mental image of total capture — which is either the truth, or a useful exaggeration, or both, depending on which room you are standing in.
I end up with a conversation, not a verdict. The models will keep improving for mundane reasons we mostly understand. The stories around them will keep improving too — including the commercial ecosystem built to shape those stories — and that progress is not illegitimate just because it is hard to benchmark. The rest of us are left to separate, as best we can, the state clock of public claims from the event clock of what actually happened, knowing both will stay noisy for a long time.