One company, several technologies, each with its own site. Take your pick below.
Phones, tablets and laptops with three engines sharing one display: the device’s own processor, a Secure Processor, and an Infinity Machine.
The Infinity Machine is a Raspberry Pi RP2350 embedded in the phone, tablet or laptop, yours to program in MicroPython.
Security you can see.
STATUS: Patent-pending. Play with a Virtual THWIXEL at thwixel.com ›
Other technology from Thort Werx
Structural privacy in an ad network that works like a stock exchange. Bids are matched openly and everyone knows precisely where the money is, while the pipeline is split so that no party ever holds both who you are and what you were shown.
STATUS: Software-implemented; advertiser console in beta. Now at thwango.com.
An image arrives as a program that builds it, not a grid of pixels, so its structure is in the stream: components with names, reusable, and readable by a machine without being guessed at.
Detailed interactive walk-thrus are under way. Details soon.
Conventional codecs treat an image as a grid of pixels and compress it. THWITRS treats an image as a program that builds it. The bitstream carries executable instructions that operate directly on pixel space, and the decoder runs them to reconstruct the picture.
This is a taster for a large piece of technology. It will take several walk-throughs of its own to do it justice, and those are on the way.
The instruction streams are called Active Entangled Streams. One stream entangles image data, control information, memory management, parallel-processing control and the decoding method itself, and may carry entropy-coded and encrypted material in the same flow.
What executes it is a Virtual Processor, which together with its Worksheet forms a virtual machine. The preferred implementation runs that machine across a plurality of GPU cores, with the engine written in shader code. The instruction format is fixed at four bytes for exactly this reason: it meshes with GPU cores, and it is position independent, so the encoder may reorder instructions to suit the entropy coder that follows.
Memory management is embedded in the bitstream rather than left to the system, which is what makes a GPU implementation practical at all: GPU cores are not suited to malloc and garbage collection. The encoder knows the state of the Worksheet at every instant and allocates and reuses it by instruction. The only transient memory the machine uses is a push stack, and that manages itself. Specialised work is handed to Special Purpose Engines, fed parameters through that stack and returning finished components to the Worksheet.
The Worksheet is one large texture in graphics memory, of arbitrary size: current GPUs, the small ARM ones included, carry a 16k by 16k texture, which at 32 bits a pixel is a gigabyte. Buffers, scratchpads and frame stacks are not separate allocations, they are simply regions of that texture, controlled by the encoder. Viewports decide which part of it is translated to the rendered image, so scrolling is a translation of the viewport rather than the re-layout, re-render and re-fetch a conventional pipeline performs, and the encoder may build ahead with spare bandwidth.
Components are built once and held in the target compute, and frames are assembled by reusing them. The encoder may retain patches, whole frames, or anything else sent since playback began, and may send a patch from a future frame to build the current one. It knows exactly what the decoder is holding.
So for many applications the group of pictures is not needed at all. The keyframe-and-group structure conventional codecs are built around exists to give the decoder whole reference frames and to provide random access points; here the encoder keeps whatever it judges useful in decoder memory, and a Ramp-On method provides random access for server-based distribution in place of the GOP. A GOP remains available where the application calls for it, such as through-the-air transmission with no return path.
The instructions describe the components of the image, and components are modules with names, reusable across a frame, a scene or a library. So the structure of the picture is carried in the stream rather than being something a machine must infer from a grid of pixels after the fact. What the image is made of, and what those parts are called, is already there. AI and other machine processes can devise their own encoding strategies and have them converge on the same consistent environment. The published application is titled Image and Texture Rendering System for Artificial Intelligence, which is the point of it. For a reader that was never human, that is a different starting point altogether.
Compression is split in two. The first stage represents the image as a computer program built from drawing primitives, which is itself a level of compression and, unusually, editable in that compressed form. The second stage entropy compresses the program. An image can therefore be edited, or compressed further, without being expanded back to pixels first.
Later filings take it further. A generative model may be targeted at an arbitrary patch rather than a whole frame, one of several methods the architecture admits, and the result is made reference-capable: stable enough to serve as the reference for subsequent frames instead of being produced afresh each time.
If images are built from components, the components become assets in their own right. Later filings set out a component repository holding Controlled Image Modules: position-independent, relocatable instruction sequences carrying semantic metadata, each able to be dropped into a bitstream and executed to build part of a picture. Alongside them sit image fingerprints, property vectors characterising a patch, so a region can be matched to a component that already exists.
Because a component is a discrete, identifiable thing rather than an anonymous region of a raster, it can be discovered, licensed, usage-tracked and paid for. Conventional formats offer creators no way to do any of that for an element inside a composite work. Humans, AI systems and convolutional networks draw on the same repository.
The filings develop the applications at length. A sample of what the architecture makes possible, and what conventional systems cannot do:
Conferencing. Member streams are composited at the recipient rather than at a server, so every viewer gets their own layout, changed frame by frame with no re-encoding of anybody's stream. More unusually, one transmitted stream can yield different authorised reconstructions at different decoders: the parts a given recipient may not see are never constructed at all, rather than decoded and then hidden by trusted client software. A member can be invisible to one person and visible to the rest. Private side-channels run inside the main conference. A small viewport asks for a smaller variant, so a phone limits how many participants it shows instead of failing.
Games and 3D. Textures are generated by instruction directly into GPU memory and sampled by the rendering pipeline; the 3D engine need not know they were not loaded from bitmap storage. Server-based gaming streams instructions instead of video: textures transmit once and persist rather than being re-encoded every frame, there is no video encode and decode in the loop, and frame rate is set by the client's rendering rather than by the stream, which puts modern games on devices that could not run them. In-game advertising is generated the same way, as the texture on a billboard, a vehicle livery, a shirt, a stadium board, moving and deforming with the object. Two players in the same game can see different advertisements on the same object, and because the advertising is built by the same instruction mechanism as everything else in the scene, an ad blocker cannot distinguish it from the game without breaking the game.
Creation. A paint program that emits instructions directly instead of rasterising to pixels and compressing afterwards, so the work stays resolution independent and structurally editable. The encoder works alongside the artist, substituting cheaper representations as the picture is drawn and showing the encoded size as it goes. Animation is generated from static source material by displacing the vertices of a quad mesh, so any frame can be produced directly without reference to any other, and scaling the displacements turns a light breeze into a strong wind. AI generation becomes something you can interrupt, correct and resume, with sight of which components and techniques the model used rather than a text box and a finished image.
Rights. Every component an AI selects is logged with its ownership, and billing can run on actual execution or display. A film abandoned halfway compensates the components that appeared in the scenes watched and not the rest. Performer likeness rights travel with a component through its whole life. Content can be signed by the camera at the moment of capture, with timestamp and device attestation, so alteration invalidates the signature: the specification puts that against the crisis in photographic evidence.
AR and VR. An environment is cut into segments whose shape follows the content instead of being forced into an equirectangular projection or six cube faces, and the segments persist in decoder memory rather than being retransmitted because a frame boundary arrived. One worked example replaces 72 uniform tiles with 11 content-shaped segments, for an estimated 60 to 75 per cent less bandwidth. Quality follows the eye, with segments enhanced before they reach the centre of vision so no transition is visible during a head turn. Return to a place you have already been and the virtual content is there at once, with no round trip to a server.
And further out. A microcontroller behind an industrial panel that is driven by nothing but a stream. A three-dimensional Worksheet for printing, CT and MRI. Streams that destroy themselves as they are processed, for sensitive material and remote simulators. A frame attribute that fires a haptic seat, a scent or a fog machine.
A page of any source may be delivered as an instruction stream and rendered in place of markup and script, with the text of a block held once and its formatting applied at the decoder. That has a security consequence: the decoder executes instructions of defined semantics and does not interpret arbitrary code in a general scripting language, so a large family of browser attack surfaces, script injection and cross-site scripting among them, is removed or much reduced.
Because the stream is instructions rather than pixels, arbitrary data can travel within it. Payloads marked for a secure processor are routed there by the decoder instead of to display memory, still encrypted, and the decoder never holds them in the clear. The stream is effectively two streams in one: visible content, and an independent encrypted stream for the secure processor, interleaved. An observer, even inside the system, cannot tell a banking transaction from a key update from a rights-management request.
Since the image is constructed by instructions rather than stored as pixels, encrypting the small fraction of instructions essential to reconstruction renders the whole frame unrecoverable without the keys. Strong protection for very little encryption. Reusable components can be protected the same way and bound to a context, a place, a time window or an authentication, so a licensed effect runs only where and when it is licensed to.
Published PCT application WO 2026/015933, published 22 January 2026, with a family of provisional applications covering selective synchronisation, decoder asset reuse, multistream processing, generative patch reconstruction, multiparty conferencing and real-time texture generation. They are grouped as the THWITRS codec family on the IP page. The full IP list ›
A separate product line: an open-architecture, MicroPython-programmable modular test instrument built on two Raspberry Pi Pico 2 modules. Dual-channel scope, logic analyser, frequency meters and a growing family of plug-in picoPods.
STATUS: Production ready. Kickstarter shortly.
Rotate your phone for a better look