I Think Google Used My Work With Gemini For Fake Promo By A Twitter Burner

[focusjordan]

I feel like a conspiracy theorist saying this. I work in Antigravity IDE. I am an ultra plan user.

After reading the article on gemini 4 argon, I was directed to an article on 3.8 flashes capabilities. One of the examples came from a faceless AI “developer” on Twitter, who solely seems to exist to parrot or promote google AI features. There example was a tyrannasourus rex, on a 3D pedestal, with a complete prompt, that followed googles prompting guidelines perfectly, while also sounding like the neuralese pulled directly from weeks of my sessions.

Now, I have, in my spare time, developing a game engine where I write OpenGL in c99 direct to buffer to procedurally complex “museum grade” assets, for a wooden tabletop 4X game I have been making.

I had spent weeks of work focused specifically on things like

• “Museum-grade” sculptural density.
• Mounting the model on a display plinth / pedestal.
• Isolating the asset in a dark studio turntable (lazy susan) viewer with orbital lighting.
• Overcoming the exact hurdles of the model taking cheap shortcuts: generating 1D line splines instead of volumetric geometry, leaving uncapped hollow tubes, and failing along curve sweeps without Catmull-Rom splines.

Here are some examples

Within a 24 to 36 hour window, of codifying the learning from this work into a skill in my IDE called “3D-procedural scultping” - an account whose entire existence is a promotional showcase for Google’s latest models drops a system prompt titled “The Procedural 3D Prompt Architect” that:

• Uses that exact lexicon (“museum-grade”, “pedestal”, “turntable”, “Catmull-Rom”).
• Forbids the exact same shortcuts (“destructive shortcuts: 1D line splines instead of volumetric geometry, hollow open tubes instead of capped manifolds”).
• Renders a 3D skeleton on a stepped pedestal in a dark studio setting.

In an infinite space of 3D applications, like you can make anything - cities, spaceships, terrain, low-poly cartoons, weapons, vehicles—the probability of an “AI Developer” with no human presence, that just so happens to nonstop showcase Google AI Products, ends up in a Google showcase, that is pushed to my email, landing on that exact aesthetic staging, exact negative constraints, and exact structural philosophy at the exact same moment by pure random chance is essentially zero.

The fact that this took place in webgl, and my work took place in opengl, seems like someone trying to make a same, same but different argument in advance. I find this really messed up. Maybe I am wrong. But -

• The account has an AI-generated avatar.
• Its entire posting history consists of curated, high-polish prompt frameworks designed to showcase Google’s newest model features (Gemini 3.8 Flash, Computer Use, reasoning benchmarks).
• Google’s own official dev blogs and engineers promote and amplify their posts.

I find it incredibly difficult to believe this is all a coincidence, when it took quite a lot of work to get to a place where this was possible and consistently quality, and then all of that somehow gets perfectly distilled and boiled down to a promp prompt using an almost 1:1 name match to the 3D procedural sculpting skill I have in my IDE, following the same exact style in a “writing to gl” based system. To an LLM, the distinction between desktop OpenGL and WebGL is just syntax wrapping.

This is really sketchy to me. I am ultra user. I work on compilers, and research, and I have been doing this for fun. If you’re going to rip me off for promo material how much are you pulling off my actual work?

[focusjordan]

[focusjordan]

[focusjordan]

And this is “IAMEMILY’S” who I am assuming is just a promo account for whatever google can reliably scrape from user sessions? maybe? the similarity in wording, and display, is like for like

[focusjordan]

and this is mine

[focusjordan]

1. On Hollow Shortcuts and Flat Geometry Her prompt states: “99% of LLM-generated WebGL code fails because models take destructive shortcuts: they use 1D line splines instead of volumetric geometry, hollow open tubes instead of capped manifolds. FORBIDDEN SHORTCUT: Do NOT use wireframe tubes, 1D curve struts, or line splines. REQUIRED: The cranium must be a closed, volumetric, manifold mesh.” Our skill states: “Mandatory Watertight Cylinder Caps: Never leave the root station or end station of a swept profile open. An un-capped polygon ring penetrating another volume appears edge-on as a 1-pixel razor-thin knife blade. Volumetric Prisms vs 2D Flats: Never render prominent features as 2D flat ribbons. Extrude along local surface normal with proud sculptural height and solid lateral thickness.”

2. On Organic Spines and Tail Curves Her prompt states: “Vertebrae: Exactly 60 articulated vertebrae along a continuous CatmullRomCurve3: 10 Cervicals: Distinct S-curve lift, 13 Dorsals, 5 Sacrals, 32 Caudals: Horizontal counterbalancing tail, dorsal spines transitioning smoothly into slender cylindrical rods.” Our skill states: “Catmull-Rom Spline Sweeps & Frenet Frames: For limbs, arched necks, cascading tails, and spiral regalia, sweep profiles along continuous Catmull-Rom splines. Continuous Frenet Frame Construction: To prevent twisting along the spine: T = Tangent, N = Normal, B = Binormal.”

3. On Ball Sockets and Intersecting Joints Her prompt states: “Hind Limbs: Sigmoid femur with spherical head seated in the acetabulum. Articular Cotyle: Cupped socket pivoting smoothly at the quadrate hinge.” Our skill states: “Anatomical Ball Joint Architecture: When multi-axis limbs intersect, mesh interpenetration can create harsh creases or angular hollows. The Ball Joint Strategy: Embed a subtle sphere or ellipsoid centered directly on the articulation pivot. Strict Containment Invariant: The ball joint center and radii must remain safely within the surrounding parent hull.”

4. On the Stepped Pedestal and Support Armatures Her prompt states: “Display Pedestal & Museum Armatures: 4-tiered stepped geological strata: Cambrian Slate, Devonian Umber, Permian Rust, and Jurassic Ochre with top deck at Y = 0.60, with 3 blackened-steel armature supports with beveled base plates bolted to Y = 0.60.” Our skill and code state: “Double-Beveled Tabletop Snap Plinth: Step Heights: Lower rim at 0.50, beveled chamfer at 0.80, and top deck at 1.00, with normal angling radiating outward on the rim and diagonally on the chamfer for specular edge catch. In our engine, the model is mounted to a central double-helix brass pole anchored directly into the plinth deck to support the floating sculpture.”

5. On the Turntable Presentation and Studio Lighting Her prompt states: “Seamless 7.00-second loop: Turntable rotation collectibleRoot.rotation.y = TAU * phase in a museum studio setting with warm key light, fill light, cool rim light, radial cyclorama gradient background, and a shadow floor fading smoothly to zero alpha.” Our engine viewer states: “Interactive Turntable Viewer: Spacebar toggles 360-degree turntable auto-rotation on a turntable swivel standee plinth, using 3-point studio lighting with a directional key sun, ambient fill, and specular rim highlights, set against a dark neutral studio cyclorama background.”

### The Question for the Antigravity Team

While the Twitter demo targets WebGL (Three.js) and my project targets native desktop OpenGL (ANSI C99), the procedural math, anti-pattern constraints, and presentation format are identical.

Given the timing and near-identical technical lexicon:

1. Does prompt or failure-mode telemetry from Antigravity IDE sessions on Ultra tiers feed into internal Google prompt research, evaluation suites, or DevRel benchmark creation?

2. What are the clear boundaries regarding session privacy when power users push frontier models into novel engineering domains?

I would appreciate clarity from the Antigravity / Google Cloud team on how user-surfaced breakthroughs and edge-case methodologies are insulated from external promotional showcases, cause it seems to me like ya might have copied my homework and tried to promote it.

[focusjordan]

I am following up on this thread before formalizing a technical escalation regarding an apparent data leakage / session isolation failure in Antigravity IDE. I hope this reaches the appropriate engineering and security desk, as I am deeply concerned by this pattern.

I find it statistically impossible that after days and weeks of intensive 3D procedural generation research—culminating in a local workspace skill called “procedural-3d-sculpting”—a pseudo-anonymous account frequently promoted and amplified by Google (@IamEmily2050) would drop a system prompt (“The Procedural 3D Prompt Architect”) that mirrors my exact skills, visual grammar, constraints, and presentation 1:1.

Consider the visual and structural overlap:

1. Visual Staging: In both the promotional prompt shared by @IamEmily2050 and in my local Antigravity scenes, there is an isolated subject floating in a dark studio void, lit by a warm directional key light that casts a distinct, soft shadow on the ground, with subtle rim lighting catching the silhouette.
2. Mounting & Armature: Both scenes mount the floating biological model above the deck using vertical metal armature rods bolted directly into the top of a stepped, beveled plinth with a warm, amber-toned deck surface.
3. Presentation Grammar: Whether viewing my carousel charger or the warm woodgrain of my plinth bases, the visual grammar is identical to the T-Rex: a dynamic, mid-stride anatomical sculpture evaluated along an organic spine, suspended like a physical collectible on a turntable display over a wooden plinth.

To believe this was completely unrelated random chance, one would have to believe that multiple independent, high-entropy variables all landed on the exact same state at the exact same moment:

• The Technical Task: Direct mathematical procedural mesh generation writing directly to GPU vertex buffers with zero external asset files.
• The Specific Anti-Patterns: Explicitly identifying and banning the exact same three failure modes: 1D line splines, hollow uncapped tubes, and flat unthickened ribbons.
• The Mathematical Mechanics: Using continuous Catmull-Rom splines for the spine/curves, closed volumetric manifolds, and embedded spherical ball joints.
• The Aesthetic Staging: An anatomical sculpture elevated above a stepped, beveled plinth with a warm amber/wooden yellow deck.
• The Studio Environment: A dark neutral cyclorama void with three-point lighting, a soft directional floor shadow, and an automated continuous turntable loop.
• The Lexicon: The recurring use of “museum-grade”, “pedestal”, “turntable”, “volumetric manifold”, and “destructive shortcuts”.
• The Time Window: All of this colliding within the exact same 1- to 3-day window in early September 2026.

It simply does not add up.

Please can anyone from the Antigravity product and security teams give me some idea whether prompt trajectories, AST outputs, or error clusters from my Ultra-tier private IDE sessions were ingested into internal evals, synthetic data pipelines, or DevRel benchmark suites during this window. The similarities here are just on the face impossible to ignore, or claim coincidence.

[focusjordan]

Final reference showing my previous rounded plinth design that measured the length of the object, while also highlighting the work I did on the carousel charger to develop a metal/standing rod in order to hold up the associated procedurally generated sculpture.

[chunduriv]

Hi @focusjordan,

Welcome to the Forum,

Thank you for raising this, we take workspace privacy and intellectual property concerns very seriously, and we looked closely into the screenshots and details you shared.

We can confirm that none of your Antigravity IDE sessions, local code, or SKILL.md files were accessed, extracted, or shared. Here is what happened:

1. The X/Twitter Post Was Published on September 3, 2026 (4 Weeks Before Your Post): If you look at the bottom of the screenshot of @IamEmily2050’s tweet that you uploaded in reply #4 (Screenshot 2026-10-01 at 10.14.01 AM), the publication timestamp on the tweet reads 12:35 AM · Sep 3, 2026.

• @IamEmily2050 (an independent community creator on X and GitHub) posted that T. rex demo on September 3, 2026, the day after Gemini 3.8 Flash launched.
• On September 28, 2026, the Google Keyword blog featured her September 3 post alongside three other community builders in “See what 4 builders are making with Gemini 3.8 Flash”.
• When you clicked through to that blog article on October 1, it surfaced a tweet that had already been public for nearly a month—well before you codified your local SKILL.md file.

2. Why the Wording (“Neuralese”) Is So Similar: You hit the nail on the head in your first post when you described the phrasing as “the neuralese pulled directly from weeks of my sessions.” Terms like “museum-grade,” “3-point studio lighting,” “Catmull-Rom splines,” “volumetric geometry instead of 1D wireframes,” “capped manifolds,” and “stepped museum pedestals/plinths” are part of Gemini’s own built-in post-training vocabulary for 3D procedural graphics. Whenever two developers independently iterate with Gemini on 3D procedural geometry and ask the model to help distill lessons and negative constraints into a system prompt or SKILL.md, Gemini naturally outputs its own characteristic terminology and structural patterns to both users.

3. Answers to Your Questions on Antigravity IDE Privacy:

   • Are local skills, prompts, or sessions shared for DevRel or promotional showcases? No. Antigravity IDE never extracts your local SKILL.md files, source code, or session prompts to feed DevRel demos, blog showcases, or external social media accounts.

• How can I control telemetry? You can view or disable telemetry at any time in Antigravity Settings via the Enable Telemetry toggle.

Your C99/OpenGL procedural assets look fantastic, and we hope the Sep 3, 2026 timestamp on the tweet and the explanation of Gemini’s shared 3D vocabulary completely put your mind at ease. Please let us know if you have any other questions!

[focusjordan]

Hey!

This is super fair!

It’s a weird space we live in where things can naturally converge, and I think the amplification of how much is possible could make these kind of creative convergences more likely.

I use antigravity IDE every day as my main work driver, so the assurances are really appreciated.

I did go back into my settings and review my telemetry, and ensured it’s off.

Again, very interesting to see how things can converge. Cheers for the response and kind words.