Meta Feeds Your Facebook Posts to Strangers; Audit Says Rio Copied 60% of Its LLM

01Meta Turns Your Old Facebook Posts Into Search Answers for Strangers

When you type a query into Facebook search now, a new option sits next to "People" and "Marketplace." It is called AI Mode, and the answers it generates draw on something you may have forgotten you wrote: your own public Facebook posts, according to The Verge.

Meta started rolling out the feature Monday, part of a batch of AI additions the company is shipping to keep users active on a platform where rivals have set the pace. AI Mode does not stay inside Facebook. TechCrunch reported that it pulls public information from across Meta's platforms, stitching together whatever you have left visible into source material for someone else's search result.

That changes what a public post does. A status update from years ago, a comment left under a friend's photo, a recommendation you typed once and never thought about again can now surface as a citation when a stranger asks the system a question. You wrote it for an audience you imagined at the time. The model treats it as a fact to retrieve.

The same launch shows where Meta's attention sits. Alongside AI Mode came photo presets that swap sports jerseys onto people in pictures, the kind of consumer toy meant to drive engagement rather than answer questions. The search feature carries the heavier implication, because it repurposes content users already created under older expectations of who would see it.

Meta has not detailed how a user opts a specific post out of being used this way, and neither report describes a per-post control. The default appears to be inclusion: if a post is public, it is eligible. That places the burden on users to audit their own histories, post by post, if they want to limit what AI Mode can quote.

The feature is live now for the users in the rollout, not a preview. For a company playing catch-up, the fastest fuel was already sitting on its servers, written by the people it is trying to keep.

02The generation most fluent in AI is the one quietly opting out

The story repeated for a year — try generative AI once and you use it for everything — is failing against its own evidence. Gabriel Weinberg argues people consume AI the way they eat meat: some embrace it, some ration it, some avoid it entirely. He writes that last year's New York Times Magazine AI issue assumed two things now turning out false: that trying AI makes you a daily user, and that by now almost everyone has.

Gallup's figures say otherwise. Take Gen Z, where awareness runs highest. Adoption barely moved across the year even as the models got better. Roughly 80% report using AI at least rarely, yet about a third reach for it only monthly, and close to one in five never use it. Over the same period, the share angry about AI climbed from 22% to 31%. The cohort that knows the tools best is sampling them, not living in them.

Move up to whole countries and the spread widens. An MIT Technology Review correspondent describes landing in Seoul and walking through an unmanned checkpoint where a machine scans her face and passport, then riding a subway built around the same comfort with automation. National enthusiasm there sits at one extreme while individual American consumers scatter across the whole range.

Then the profession that should fall first. Arvind Narayanan and Sayash Kapoor examine software engineering, a field with few regulatory barriers and unusually high exposure to automation. They argue there is enough evidence to reject the claim that crossing some capability threshold triggers mass layoffs. Engineers have not been replaced. Because software faces the fewest barriers, they reason most other professions are better insulated still.

By country, by consumer, by occupation, the same gap surfaces: real use sits far below the assumption that everyone is using AI for everything. The near-universal awareness number hides a population that mostly samples, rations, or refuses.

03Rio's 397B "homegrown" LLM is 60% another company's weights, a tensor-by-tensor audit alleges

The city of Rio de Janeiro published Rio-3.5-Open-397B as an original 397-billion-parameter model trained by IplanRIO, its municipal IT agency. A startup called Nex-AGI says the weights tell a different story.

In a GitHub issue, Nex-AGI alleges that Rio's model is a direct element-wise merge of two existing systems: roughly 0.6 of its own Nex model blended with 0.4 of the official Qwen3.5-397B-A17B base. The firm says it found no evidence of any independent training.

Its proof runs two ways. Strip out Rio's hard-coded "You are Rio" system prompt, and the deployed model identifies itself as "Nex, from Nex-AGI" 79% of the time, according to the issue. It calls itself "Rio" 0% of the time. The model also recites Nex-AGI's custom backstory word for word. Separately, every weight tensor in Rio matches the same 0.6/0.4 blend to thousands of standard deviations, across all 60 layers. Nex-AGI argues ordinary finetunes cannot produce that pattern.

The post drew 394 points and 212 comments on Hacker News. IplanRIO has not publicly responded to the allegation.

Days earlier, a similar collision played out at a larger institution. KPMG pulled a report about AI usage after apparent hallucinations surfaced in the document, TechCrunch reported. The consulting firm's report on how AI gets used contained AI-generated errors of its own.

The two cases share a structure. An institution publishes an AI artifact under its own name. Outside parties check it against the underlying evidence: model weights in one case, citations in the other. The published claim does not survive the check.

What differs is the cost of being caught. KPMG retracted a document. Rio's model still sits under a government namespace, prefeitura-rio, and remains the basis of a public-sector AI claim that an independent audit disputes.