Meta AI's Hidden Features Reveal a Platform

Meta AI ships pre-installed on billions of devices, embedded in WhatsApp, Instagram, Facebook, and the Ray-Ban smart glasses, and most of its users are probably treating it like a slightly better search bar. A recent tutorial walkthrough by TheAIGRID — running nearly 16 minutes through the platform's lesser-known capabilities — is a useful corrective to that assumption. What the video surfaces, feature by feature, is a product that is doing considerably more than its casual reputation suggests. Whether that's welcome news depends heavily on what questions you're asking about it.

The tutorial starts with the basics: three chat modes called instant, thinking, and contemplating. Instant is self-explanatory — fast responses for quick queries. Thinking engages chain-of-thought reasoning for multi-step problems. But the third mode is where the tutorial's creator, TheAIGRID, slows down. "There is a feature called contemplating," he explains, "where you can summon up to 16 agents. You can spin parallel sub-agents that reason independently and then combine answers."

The demonstration is genuinely interesting: he prompts Meta AI to "spin up four different agents to reason about this problem," and the system responds by spawning labeled sub-agents — an online business analyst, a bootstrap entrepreneur strategist, an AI-first small business strategist, and a creator monetization analyst — which reason independently before synthesizing conclusions. The stated logic is that a single LLM tends to defend whatever position you've implicitly set up for it, while competing agents can surface genuine contradictions.

Here's the caveat that matters: this feature appears in no official Meta documentation I can locate. TheAIGRID acknowledges it himself — "there's no guarantee that Meta will keep this feature in because it does require a large amount of compute on their side." What he describes sounds less like a documented product feature and more like an emergent behavior triggered by specific prompting, or possibly an experimental capability in limited rollout. Readers who want to replicate it should treat the walkthrough as a starting point, not a user manual. The feature may work, may not be available in your region, or may disappear before you try it.

The Search Question Nobody Asks

Meta AI's search function is where the platform's structural position becomes hard to ignore. The system pulls from "the open web plus Meta's social graph," which the tutorial demonstrates concretely: a query about men's running shoes in 2026 returns results that include direct links to Instagram posts alongside conventional web sources.

When TheAIGRID checks the sourcing on a shoe query, the results include "literally just a running website and some posts on Instagram." The tutorial frames this as a transparency note — hover over the citations and you'll see where the information originates. That's accurate and worth knowing. But the more consequential point is structural: Meta is the sole AI assistant in the market that can route commercial queries through its own social content graph, where years of sponsored posts, influencer endorsements, and brand-driven user content live. The platform's influence on consumer behavior in categories like footwear, food, and fitness derives directly from this accumulated social content — which is now also the retrieval corpus for its AI search results. Whether that constitutes a useful feature or a conflict-of-interest worth regulatory attention is a question the EU's Digital Markets Act enforcement teams are, in related contexts, actively working through.

The product search demonstration sharpens this further. Asking Meta AI to "show product cards for shoes I can buy" produces clickable product listings. The tutorial presents this as convenient. The mechanism — a closed-loop system where Meta's social graph informs what surfaces, and Meta's commerce infrastructure handles the transaction — is something the FTC's ongoing scrutiny of vertical integration in digital markets was designed to examine.

Visual Grounding: The Feature That Actually Earns Attention

Set aside the strategic questions for a moment, because the image analysis capabilities demonstrated in the tutorial are technically sophisticated in ways that go beyond standard multimodal benchmarks.

The tutorial demonstrates what Meta calls "visual grounding" — the system's ability to not just describe an image but anchor analytical outputs to specific spatial coordinates within it. TheAIGRID shows a prompt asking Meta AI to analyze a fridge photo for a pescetarian with high cholesterol: green dots on recommended foods, red dots on items to avoid, health scores visible without hovering, macronutrient data appearing on hover, all localized precisely to each food item in the image.

The system produces an interactive annotated image that works largely as described. "The more detail you do have in an image, the more difficult it's going to be," he acknowledges — the accuracy isn't perfect — but the architecture here is notable. This is multi-step reasoning (dietary constraints plus visual object identification plus spatial coordinate mapping plus interactive display generation) executed in a single prompt exchange. The output format — an interactive artifact rather than a text description — reflects a design philosophy that treats vision as a primary interface rather than a bolted-on capability.

A simpler demonstration follows: a meal photo prompted with "estimate calories and macros, make it so that when I hover over the item, it says exactly that" produces a labeled interactive image with per-item nutritional breakdowns. It works. For anyone tracking dietary intake without wanting to manually log every item, the practical utility is evident.

The Image Generation Caveat

Meta Vibes, the platform's image and video generation component, is where I'd urge the most caution about claims circulating in tutorial content. TheAIGRID states that Meta AI is "currently running Midjourney" as its underlying image generation model — a claim that would be significant if true, given that Meta has invested substantially in its own models (Emu, and subsequently Imagine with Meta AI). I cannot independently verify this claim, and it conflicts with Meta's public positioning on its proprietary image generation stack. The tutorial offers no primary source for this assertion. Treat it as unconfirmed until Meta addresses it directly.

What the tutorial does demonstrate reliably: image generation works, aspect ratio control is functional (specifying "16:9" produces widescreen outputs), and the animate feature converts still images to short video clips with reasonable fidelity to prompted motion. The tutorial creator animates a tennis player image into a clip of the figure falling over — crude by premium video generation standards, but operationally functional for basic content creation.

Meta Vibes as Social Infrastructure

The Vibes tab deserves a separate beat, because it's not really a creative tool — it's a social platform for AI-generated content. Users can share creations, follow other accounts, remix others' work, and build follower audiences around their AI-generated output. On mobile, the remix function allows users to insert their own face (or others') into existing creations, add AI voiceovers across a library of voices (the tutorial mentions approximately 12 options at time of recording, though feature counts in rapidly-iterating products should be treated as snapshots rather than specifications), and layer in music tracks.

The face-insertion feature carries a disclosure in the tutorial that is worth quoting directly: "I'm guessing that you've of course accepted Meta's TOS for your face to be there." That "guessing" is doing considerable work. The consent architecture around biometric data in AI-generated social content — who has consented, to what, under which terms, and whether those terms are genuinely informed — is the subject of active litigation and pending legislation in multiple jurisdictions. Illinois' Biometric Information Privacy Act cases and the EU AI Act's provisions on biometric categorization systems are both directly relevant here. Meta's terms of service may be legally sufficient. That is not the same as being structurally sufficient for the kind of seamless, casual face-swapping the Vibes feature enables.

What's Actually Being Assembled

Look at the full feature map from the tutorial and a coherent picture emerges: reasoning interface, social search with proprietary graph, commerce integration, vision analysis, AI image and video generation, and a social sharing platform for AI content — all unified under a single product with existing distribution across Meta's family of apps.

The policy questions that picture generates are specific: Does routing commercial queries through an owned social content graph require disclosure under the FTC's endorsement guidelines? Do the AI Act's transparency requirements apply to the contemplating mode's multi-agent outputs when users may not understand they're receiving synthesized multi-model reasoning rather than a single response? Does the Vibes face-insertion feature meet the consent standards that several state-level biometric privacy laws demand?

Tutorial walkthroughs don't ask these questions, and they shouldn't — that's not their job. But a product this deeply embedded in daily digital life, accumulating this breadth of capability this quietly, is exactly the kind of thing that lands on regulators' desks after the fact, when unwinding the design choices is no longer straightforward.

Meta AI is not waiting for the regulatory framework to catch up. It's building distribution.

Samira Barnes is Buzzrag's tech policy and regulation correspondent.