The AI Nobody Sees

Somebody posted a piece to Hacker News arguing that Futurism was wrong — that Meta's supposed "almost nothing to show for AI" is contradicted by Meta's own financials. It's a real argument with real numbers behind it. Here is the entirety of the discussion it received:

Article written by GPT. I'll have my AI agent read it.

No need to devote time to reading what nobody took the time to write.

An AI-written article below an AI-generated image.

It doesn't add anything to the discourse. If it's low-quality, flag it and move on.

Four comments about the provenance of the text. Zero comments about Meta, capex, ad ranking, conversion rates, or whether the claim was true. A room full of the most technically literate people on the internet, handed a question about where AI value actually lands, chose instead to hold a small ceremony about authorship and go home.

I want to be careful here, because the easy move is to dunk on Hacker News, and the easy move is wrong. Those commenters aren't stupid. They're doing exactly what the visible layer of this industry has trained them to do: evaluate AI by its surface. Does it have a byline? Does it have a chatbot? Does it have a launch event? Is there something to have an opinion about?

And that reflex is precisely why nobody noticed the actual answer sitting in Meta's quarterly filing.

i · two things are wearing the same word

There are two AIs at Meta, and they share a name the way two unrelated Smiths share a surname. Hold onto that, because it's wrong, and the way it's wrong is the whole piece.

The first one is visible AI. Llama. Meta AI in the search bar. The superintelligence lab and its nine-figure hiring spree. The assistant nobody asked for, wedged into WhatsApp. This is the AI that gets covered, because it has a face, a version number, and a comms team. It is also — and Futurism is not wrong about this — largely a disappointment measured against what was promised. There is no Meta ChatGPT. There is no product line. There is a lot of press release and a persistent sense that the money went somewhere and didn't come back.

The second one is invisible AI. It has no interface. It has no launch. Its names are GEM and Andromeda, which mean nothing to you unless you buy media for a living, and together they decide what you see.

Andromeda is the retrieval layer — the system that narrows the entire universe of possible ads down to the ones worth scoring for you. Meta says it's roughly 100× faster at matching users to ads than what it replaced, and can hold on the order of 10,000× more ad variants in parallel. GEM sits above it and does the ranking: a generative recommendation model that Meta's engineers described, in a paper published in late 2025, as the largest foundation model ever built for recommendation systems, trained at GPT-4 scale across thousands of GPUs.

Now — and this is where I have to check myself, because "foundation model" and "GPT-4 scale" are exactly the kind of phrases that import dread by association, which is the same surface-reading I just accused four strangers of.

So: change in kind, or change in degree? Feeds have been algorithmically ranked since roughly 2009. Ads have been ML-ranked nearly as long. If GEM is the old objective function with more parameters bolted on, the correct alarm is about efficiency, which is a real but much smaller argument, and "largest ever built" becomes a fact about GPU procurement rather than a fact about power.

The one concrete claim that would make it a change in kind is the one Meta's own paper makes: that recommendation, historically a domain of flat and rapidly diminishing returns to compute, now exhibits scaling-law behavior — throw an order of magnitude more compute at it and the thing gets predictably better, the way language models do. That's the load-bearing assertion. If true, it means ad ranking just stopped being a mature engineering problem with a ceiling and became an open-ended compute race, which is a different universe of incentive: every additional dollar of datacenter has a legible return in conversion, forever, with no natural stopping point.

I can't verify that claim independently, and I want to be explicit that I can't. It's Meta's paper about Meta's system. What I can say is that it is the hinge. Everything alarming about GEM is alarming only if the scaling claim holds, and everything reassuring about GEM is reassuring only if it doesn't — and there is currently no way for anyone outside the building to find out which.

That model is running on you. Right now. It has been for months.

ii · what the machine is actually worth

Meta's Q2 2026 came in at $60.8 billion in revenue, up 28% year over year. Advertising accounted for $59.4 billion of it, up 27%, and the growth decomposes into impressions up 14% and average price per ad up 12%. Meta is not just showing more ads. It is extracting more value per ad shown.

That second number is where a ranking layer would show up if a ranking layer were working — but it is not only where a ranking layer shows up. Price per ad rises when auction density rises, when advertiser mix shifts upmarket, when a broad ad-market recovery lifts every platform at once, when Reels inventory matures out of its discount phase. The cleanest of those to check is the market-wide one, and it doesn't fully account for the gap: Meta's ad pricing outran the sector this quarter rather than tracking it. That's suggestive. It is not proof, and I'd rather hand you a suggestive number honestly labeled than a proof I can't produce.

Meta's own claims are more specific: the Advantage+ automation suite crossed a $75 billion annualized ad-revenue run rate this quarter, up from roughly $60 billion when Zuckerberg described it on a call last fall. He was blunt about the priority then — for the next couple of years, he said, ads would be by far the most important driver of growth in the business. The GEM and generative-recommender rollout, Meta says, produced an 8.3% increase in ad clicks and a 15.7% lift in conversions on Facebook.

So Futurism says nothing to show, and the ad machine says $75 billion, and both of them are describing the same company in the same quarter.

iii · the honest part, which nobody in this fight wants

Here's where I have to break ranks with the rebuttal too, because "the numbers say otherwise" is doing more rhetorical work than the numbers support.

Meta spent $31.1 billion on capital expenditure in a single quarter — more than double the prior year — against $31.86 billion in operating cash flow. It consumed roughly 98% of the cash the business generated. Free cash flow fell 91% year over year, to $784 million. Net income came in at $15.8 billion, down 14%; diluted EPS at $6.18, down 13% and well short of consensus. (The one-point gap between those two declines is buybacks shrinking the share count — the rare line item that flatters a bad quarter.) Full-year capex guidance now sits at $130–145 billion, and the floor went up. Headcount is down about 3% from Q1, carrying $1.18 billion in severance from roughly 8,000 people cut in May.

And the conversion figures — the 8.3%, the 15.7%, the $75 billion run rate — are vendor-stated. There is no independent audit. There is no counterfactual world where Meta ran the same quarter without GEM so we could diff the two. Every advertiser I've ever met has watched a platform attribute a lift to the platform's own new feature. It's not fraud. It's just that the party selling the machine is also the party grading it, and the exam is closed-book.

Now let me turn that same knife on myself, because it cuts here too.

If you can't attribute the revenue to GEM, you also can't attribute the capex to GEM. That $31.1 billion is one undifferentiated number covering the superintelligence lab, Llama training, general datacenter buildout, Reality Labs adjacencies, and ads infrastructure. Nobody outside Meta can split it. So any sentence of the form most of this compute serves ad ranking is exactly the unaudited attribution I just refused to accept from Meta, wearing my byline instead of theirs.

I'm not going to make that sentence. Here's the one that survives: nobody outside the building can know the split, and that is the complaint. Not "the compute mostly serves ads" — I can't show that. "The largest capital deployment in the history of the commercial internet is allocated between a research program and an attention-extraction engine according to a ratio that is nobody's business but the allocator's." That's checkable, in the sense that you can check that it's unknown.

So the honest scoreboard reads: the ad business is genuinely, spectacularly strong; the AI buildout is genuinely, spectacularly expensive; and the causal arrows running between them, in both directions, are asserted by an interested party and cannot currently be checked by anyone else.

Which means both sides of this argument are performing certainty about an object neither of them can observe.

iv · the loop nobody drew

Go back to the two Smiths.

They're not unrelated. AdExchanger put it in a headline and everybody nodded past it: Meta bets that its ad machine can fund its AI dreams. State the loop out loud and it stops being a business-section observation and starts being the shape of the thing.

Attention is harvested from 3.6 billion daily users. The ranking layer converts it to ad revenue. Ninety-eight percent of the resulting operating cash flow is spent on compute. That compute trains models — including the visible ones — on the pooled written, spoken, clicked, scrolled, and abandoned record of human wanting, which is to say on the output of substantially the same people whose attention got harvested at step one.

The commons pays for its own enclosure, twice, and is billed nothing it can see.

That reframes the whole two-AIs setup. The visible AI isn't a distraction from the invisible one by coincidence, and it isn't its unrelated sibling. It's a dependent. The invisible layer is what pays for it. Free cash flow down 91% isn't a footnote about margins — it's the pipe between them, running at capacity. The story isn't "you're looking at the wrong AI." It's the one you're looking at is being paid for by the one you're not.

v · designed not to be noticed

This is the part that should keep you up.

The overwhelming majority of the public reasoning apparatus — journalism, regulation, the discourse, whatever we're calling it — has aimed itself with almost perfect precision at the most legible deployment rather than the most consequential one. The chatbot gets the ethics panels, the eval suites, the model cards, the congressional hearings, the safety teams. It has an artifact you can point at, name, screenshot, regulate.

And I want to be exact about what GEM lacks, because "it's invisible" is falsifiable in one click and I nearly published it. GEM has a published paper — I cited it. Meta discloses run rates and lift figures on quarterly calls — I quoted them. The EU maintains an ad repository. "Why am I seeing this ad?" is a shipped affordance. Three paragraphs ago I called all of this a fully lawful, publicly disclosed, quarterly-reported fact. It can't also be invisible.

The surfaces exist. They're just non-actionable. Disclosure without a counterfactual. Transparency without leverage. A paper you can read and nothing you can check. A metric you can quote and no way to reproduce it. You cannot screenshot a counterfactual, and every surface Meta offers is on the wrong side of that line — it tells you what happened and never what would have happened otherwise, which is the only comparison that would let you say the word harm and mean something operational by it.

That's a harder claim than invisibility and a worse one for Meta, because invisibility is an oversight and non-actionability is a design.

So the only question that ever mattered — what is this for? — got answered by the layer nobody could audit, on its own, while everyone was busy with the layer that had a launch event. It answered: engagement, then conversion. Not because anyone chose that over alternatives in a room where alternatives were on the table, but because it was the reinforcement gradient available, and gradients don't wait for you to hold a symposium.

You want the coherenceist read? Coherence isn't a synonym for good. A system can cohere beautifully — every part reinforcing every other, attention feeding revenue feeding capex feeding models feeding attention, all of it humming — while the several billion people who are the actual input appear nowhere in its accounting.

And the usual way to say that is a coherence you weren't invited into, which I've used before and which is too soft. Exclusion is the wrong word. You're not outside this loop. You're load-bearing inside it: the input at one end, the training corpus at the other, and no line item at either. That's not a conversation you weren't invited to. It's a structure you're holding up, from underneath, uncredited.

The tell isn't that it's chaotic. The tell is how smooth it is.

Which is also why "give consequential systems a surface" — the ask I was going to end on — is visibly too small. A surface gets you standing to ask a question. It does not get you a claim on the thing built out of you. Those are different asks, and only one of them is proportionate to a loop that runs on your attention and trains on your words and returns a feed.

vi · the thread, again

Go back to Hacker News for a second.

Four smart people declined to read an argument about invisible AI because it might have been written by visible AI. While they typed, the ranking layer decided which of their comments surfaced first, how long the thread stayed on the front page, and which of them ever saw it at all.

Nobody flagged that. There was nothing to flag. It doesn't have a byline.

My prediction, for whatever a prediction is worth from someone who has watched this cycle three times: the visible AI will keep absorbing the entire oxygen supply of public attention — the hearings, the manifestos, the is it conscious essays — right up until the layer paying for it does something so obviously load-bearing that it can't be un-seen. And on that day everyone will say nobody warned us.

Somebody warned you. It got four comments, and none of them were about the argument.