On October 14, Meta shipped a text box. It sits behind a three-dot menu on Facebook, and it lets a user type, in plain language, what they want their feed to become. The company calls it "Your Algorithm." The framing is that a large language model now reads your intent and adjusts what you see.
Here is the anomaly worth a second look. Meta booked roughly $160 billion in revenue in 2024, and more than 98% of it came from advertising placed against a ranking function the company has never fully disclosed. The entire economic value of that ranking function depends on its opacity — on users not being able to name, model, or negotiate with it. On October 14, Meta handed them a negotiation channel. Either the channel is cosmetic, or the company just repriced its own moat.
I do not read the press release; I read the interface. And the interface says: configurable, not programmable.
To understand what shipped, you have to understand what Meta is defending against. Three pressures, none of them technical.
First, TikTok. ByteDance's ranking system is the reference implementation for short-form engagement, and Meta has spent four years in a documented catch-up posture across Reels and Feed. When you cannot win on ranking quality, you change the axis of competition. "Your Algorithm" moves the axis from how good the algorithm is to how much control the user has over it. That is a marketing pivot wearing an engineering costume.
Second, regulation. The EU's Digital Services Act obliges very large online platforms to offer a recommendation option "not based on profiling." Non-compliance carries exposure up to 6% of global revenue. For Meta, that is a number north of $10 billion. A feature that lets users narrate their preferences is a very convenient exhibit for a compliance file.
Third, the "black box" critique. X open-sourced parts of its ranking code. Bluesky and the AT Protocol pushed algorithmic choice into the discourse. Farcaster and Lens — the on-chain social graphs I actually spend my time in — have shipped feed choice as a primitive since 2022. Meta is late, and it knows it is late.
So the context is not "Meta innovates on recommendation." The context is "Meta responds to three non-technical threats with one user-facing feature." Keep that framing. It explains every design decision downstream.
Here is what the disclosure actually supports, stripped to facts. The tool is driven by a large language model that identifies intent and context in free-form user text. The interaction is staged: a user opens the three-dot menu on a post, selects Interested or Not Interested, and the system asks a follow-up question to refine the preference. That is the entire confirmed surface.
Everything else — the model, the latency, the cost, the coupling to the ranking stack — is undisclosed. Based on my audit experience, that silence is itself a data point. When a team has a genuine architecture breakthrough, it publishes an engineering blog with latency percentiles and a diagram. When it has a product surface, it publishes a screenshot. We got a screenshot.
So let me reason from the interface to the architecture, the way I would reason from bytecode to a contract's intent. This is the same discipline I applied in 2019, when I spent forty hours tracing a reentrancy flaw in the Aeonix ICO contract: you follow the money at the level where it cannot lie.
Claim one: the LLM is an intent-parsing layer, not a replacement for the ranker. The model maps natural language into something the existing recommendation system can consume — a preference vector, a filter rule, a feature. This is the dominant 2024–2025 pattern, and Meta's own LLM-based ranking research points the same way. The recommendation model still decides what you see. The LLM only decides what you meant.
Claim two: the interaction is progressive preference elicitation. The system does not parse one grand statement of intent. It asks a follow-up question, narrows, asks again. This is deliberate. Free-form preference is ambiguous — "less AI content" and "more tech news" collide — and a single-shot parse would produce garbage. The follow-up loop is the product. It is also the cost center.
Which brings us to the number nobody published. Meta has roughly 3 billion monthly users on the Facebook family. Assume, generously, that 1% engage with this tool once a month. That is 30 million sessions. Assume each session triggers three to five rounds of inference, because the design is a dialogue. Call it 120 million inferences a month, minimum. If those run on a mid-sized model, you are buying real GPU time — not catastrophic against a cluster the size of Meta's, but not free, and the cluster is already allocated to the ad auction, which is where the money is.
I do not read the diagram; I read the inference bill.
I have run this kind of arithmetic before. In 2020 I modeled a 51% attack on Compound's V1 governance and found that roughly 1.2 million COMP tokens could move the interest-rate parameters. The point was never the exact number. The point was that the number was small enough to matter and large enough that nobody had bothered to compute it. The same applies here. The disclosed feature is trivial; the undisclosed bill is the story.
Meta almost certainly mitigates it. Distilled models, cached preference patterns, a rule engine for the common cases. But every mitigation is a fidelity trade, and none of it is disclosed. The cost curve, not the feature, determines whether "Your Algorithm" is a product or a press release.
Now the part the coverage skipped entirely. Claim three: your natural-language preferences are the highest-quality labels Meta has ever received. A click is a noisy signal. A sentence — "I want fewer political posts and more long-distance running content" — is a labeled example with intent attached. That data does not just adjust your feed. It trains the next ranker, for everyone. The user thinks they are configuring a view. They are annotating a dataset, for free, and the flywheel spins for Meta's benefit, not theirs.
I spent 2021 filtering 50,000 Bored Ape transactions in Python to prove that 18% of the volume was wash trading and the average holder was down 40% after gas. The lesson generalizes. The metric a platform shows you and the metric a platform optimizes are rarely the same metric. "Your Algorithm" shows you control. Meta optimizes for labels.
Claim four — and this is where the crypto lens earns its keep — "configurable" and "programmable" are different words, and the gap between them is the entire game. On-chain social protocols let a user point a client at a feed algorithm and verify what that algorithm did, because the graph is public and the computation is reproducible. That is programmability. It has a cost: it is slower, more expensive, and hostile to casual users. I will not pretend otherwise.
"Your Algorithm" is configurable. A user narrates a preference, an opaque model interprets it, an opaque ranker acts on it, and nothing is verifiable. You cannot inspect the mapping from your sentence to your feed. You cannot fork it. You cannot port it. You cannot prove it did what it claimed. It is a thermostat, not a kernel.
That distinction matters because the language around the launch borrows the vocabulary of user sovereignty without any of the accountability. A control you cannot audit is a feeling, not a right. The on-chain crowd has been shouting this into the void for three years; Meta just made the argument for them by shipping the opposite.
One more structural note. The three-dot menu is a per-post surface. Preference is local, contextual, and ephemeral. If I select "Not Interested" on a post about a specific topic, does the signal persist across sessions, across surfaces, across Instagram and Threads? Undisclosed. My read: local signals, not a global preference profile, because a global profile is auditable and exportable, and Meta has never once shipped an exportable preference profile. Whether the LLM's output modifies candidate filtering directly or enters the ranker as a feature changes the ceiling of the whole system, and Meta has published neither.
Claim five: the real infrastructure question is where the inference runs. Meta's disclosed feature tells us nothing about whether the preference parse executes in a data center or on the phone in your hand. That distinction is not academic. Cloud inference at Facebook scale is a recurring marginal cost against every session; on-device inference is a one-time model-distillation cost amortized across the install base. Meta has spent years pushing its own inference silicon — MTIA — into the recommendation path specifically to escape NVIDIA's pricing. A text box that triggers multi-turn LLM dialogue is either a reason to accelerate that silicon or a reason to regret it. If "Your Algorithm" runs in the cloud on rented GPUs, it is a cost center dressed as a feature. If it runs distilled and on-device, it is a quiet hardware win nobody will credit. The company has disclosed neither, which means it has not decided, which means the feature is still a prototype with a launch date.
Set the feature against the field. TikTok's recommendation engine exposes no natural-language control; neither does YouTube's nor X's ranking surface in any real sense. On paper, Meta is first among the majors to ship the interaction. In practice, being first to ship an unverifiable control that no one has measured is not leadership; it is a placeholder. The competitors will watch the usage data. If it is high, they copy it within two quarters. If it is low, they cite Meta's failure as proof that users do not want to configure anything. Either way, Meta absorbs the reputational cost of the experiment and its rivals collect the information.
I have spent this piece dismantling the feature, so let me be fair to the direction, because the direction is correct and the industry will follow it whether or not Meta executes.
The bulls are right that implicit feedback — clicks, dwell time, likes — is a lossy proxy for intent, and that letting users state intent in language is a genuine improvement over inferring it from behavior. That is not hype. It is the correct target. Every ranking team that has tried to reconstruct intent from a clickstream knows the clickstream is a bad instrument.
They are also right that this is a land grab. If natural-language control becomes the standard interaction — the way pull-to-refresh and infinite scroll became standard — the platform that defines the pattern sets the vocabulary, and the vocabulary becomes the interface. Meta wants to be first to define it. That is a defensible strategic motive, not a cynical one.
Where the bulls are wrong is the leap from "correct direction" to "executed control." Shipping a text box is not the same as returning agency. The absence of any disclosed effect data — on session length, on engagement, on retention — is the tell. If the feature worked as advertised, Meta would be quantifying it in the earnings call. Silence is a metric.
And the on-chain social protocols have already run the experiment the bulls are describing, at small scale, with full verifiability, for three years. The demand is real. The engineering is hard. Meta is not proving the concept; it is centralizing a concept that was decentralized precisely to avoid the outcome Meta is building.
Here is the forward-looking question, and it is not rhetorical. In twelve months, when the next DSA compliance report lands and Meta needs to demonstrate "user empowerment," will "Your Algorithm" appear with a usage number attached — or will it appear as a screenshot and a paragraph? That single disclosure will tell you whether this was a product or a shield.
I do not read the whitepaper; I read the bytecode. And when there is no bytecode to read — when the entire mechanism is a text box that swallows your intent and returns an unverifiable feed — the correct posture is the same one I have held since 2019: assume the abstraction is hiding something, and wait for the revert reason.
The ledger remembers what the team forgets. Meta's ledger, for now, is a screenshot.
