Tell the computer what you want. Essence works out how.
This is the longer, visual version. Each chapter starts with one key idea, the diagrams add detail, and the word list below explains the terms in plain language. For the idea in five minutes without diagrams, start with the short version below.
Today, a computer does a job by following exact steps that people wrote in advance. Essence works the other way: a person says what outcome they want and what limits apply, and the system works out the steps at the moment the job runs, for the machine in front of it.
Only one part runs publicly today: a tool that makes graphics and compute jobs run faster on GPUs, measured in pilots on cloud GPUs. The rest of the platform returns when the next funded milestones are complete. Read the five-minute version or see the plain-language word list.
Words used on this page, in plain language
- Essence
- The platform. It takes a statement of what someone wants and works out how to do it when the job runs.
- AI and Essence
- Essence is not AI and does not use AI to decide or to execute what runs. AI systems can connect to it and propose, and people and policy decide.
- Wantware
- The name for this approach. Software says how; wantware says what. It is not about writing or generating code as applications.
- Meaning Coordinates
- A fixed, public set of 256 building blocks for describing what someone wants and the limits that apply, so the same request always means the same thing.
- Aptiv
- A governed building block that carries its own meaning, rules and history, and combines with others without extra connecting software.
- Composite Job Design
- The working design Essence generates for one job on the specific hardware that is present.
- Chameleon
- The GPU optimization tool at adaptwithchameleon.com. It is the only part running today.
- SecuriSync
- The part that decides whether a request may run, by checking who is asking, that it is intact, and that they are allowed.
- StreamWeave
- Protection for data in motion. It splits data into pieces and protects each piece differently.
- Wantverse [.wv]
- The package that carries composed Aptivs and intent. It is a stream, not a conventional file.
- Synergy
- The plain-language layer, where people describe an outcome and the system asks if anything is unclear. An earlier version is paused until the milestones are complete.
- illumin8
- The signal-processing layer. It is shown in recordings today and returns when the milestones are complete.
- Milestones
- Two pieces of funded work, Buffer Feeds and Synergy Expansion. When they are complete, the parts that do not run today can be deployed again.
Three ways to get a computer to do something
Every system answers one question: how does a wish become an action? There are three approaches, and they differ in how the route is chosen.
Software gives a computer exact steps, AI guesses likely steps, and Essence names the destination and works out the steps itself.
- Programmers write exact instructions for one situation. It is predictable, but it is only one way to do the thing. When the situation changes, people rewrite it.
- Generative AI picks among many likely routes. It is flexible, but it can arrive somewhere nobody intended, and the path is hard to explain afterward.
- A Meaning Coordinate names where you want to end up, the way a map coordinate names a place. It can be reached from many starting points, and the result is deterministic. You choose what matters most, such as time, cost, energy or quality, without writing code.
Go deeperSoftware vs Wantware layers
Meaning Coordinates: a periodic table for intent
Chemistry works because a small, published set of elements combines into everything. Essence applies the same idea to meaning.
A small, fixed set of building blocks for meaning lets the same request always resolve to the same thing.
- Meaning Coordinates are not a programming language and not natural language. They are a way of writing down computable meaning: what someone wants, within what limits. There are 256 primitives, organized as 4 realms of 64 and 32 groups. The set is fixed and public. The value lies in the machinery that applies and combines them.
- Operations: comparing, collecting, counting, and measuring, the acts that underpin all computation. Its eight groups: Terminator, Create, Condition, First, Operator, Essence, Numbers 0 to 7, and Numbers 8 to 15.
- Cognition: values, logic, information, thought, instances, situations, possibilities, and work. Its eight groups: Combo, Imagine, Domain, Topic, Thing, Goal, Think, and Govern.
- Physical: spacetime, physical properties, scale, shape, energy, matter, body, and thing types. Its eight groups: Heading, Work, Reflect, Point, Nuclear, Solid, Head, and Object.
- Relational: sharing, perceiving, judging, feeling, moving, doing, using, and communicating. Its eight groups: Sense, Detail, Compare, Zeal, Inside, Mark, Make, and Session.
- A request made in plain language resolves to coordinates drawn from across the realms. How many depends on the intent. The picture is illustrative and shows one coordinate in each realm. The same request always resolves to the same coordinates. Say it loosely and the system has freedom to optimize. Say it tightly and every step stays traceable.
Go deeperLevel-of-detail and Meaning CoordinatesSemantic encoding appendix
One source, any machine
Software is normally written for the kind of machine it will run on. Essence decides at the moment of running.
Essence looks at the machine when the work runs, so one description can serve many different machines.
- The original can be small. It might be 1K or less.
- Each machine’s architecture is assessed in real time. That is why no hypervisor is needed. A hypervisor exists to paper over the gap between code written for one machine and the machine actually present.
- A 4K display receives 4K. An 8K display receives 8K, from the same source. Quality no longer depends on the size of the original.
Go deeperQcode and Morpheus
The life of a request
Follow one request from a person’s words to a recorded result. Each box is a named part of the platform.
A request is stated, translated, assembled, checked, run and recorded, in that order.
- Plain language goes in. People describe the outcome and the limits, not the code. If something is unclear, the system asks instead of guessing. AI systems that connect to Essence can only propose intent. They do not execute anything themselves.
- Fill-in-the-blank templates (called Grok Units) that match common phrasing and resolve it to Meaning Coordinates. No model inference is involved, so the same words land on the same coordinate. New phrasings can be added in context, and one user’s addition does not change another user’s meaning without approval.
- These building blocks, called Aptivs, carry their own meaning, rules and history. They combine without middleware, and they do not attempt to expand their own capability or authority.
- Before anything runs, the request declares who, what, when, where, how and why. A built-in check compares identity, integrity and authority with that declaration.
- Instructions are generated at the moment of running for the hardware that is actually present, with work spread across the available processors.
- The outcome arrives with a record of what ran, where, and under whose authority. identity and data are stored without files or databases.
Parts you will hear about
Plain-language layer (Synergy)
The conversation layer. People state intent in plain language and operate and govern the system through it.
Building block (Aptiv)
A governed building block that carries its own meaning, rules and history. They are sorted into 64 categories, combine without middleware, and cannot expand their own authority.
Package (Wantverse [.wv])
The package that carries composed building blocks and intent for deployment. It is a stream, not a conventional file.
Three ways to run it
By design, the same source can run three ways: one produces ordinary files for existing pipelines, one runs live processes on an operating system, and one runs bare-metal as a unikernel for high-security appliances. None of these runs today. Only the GPU optimization build runs, and the rest return with the funded milestones.
Permission check (SecuriSync)
Decides whether something may run, checking identity, integrity and authority, and revalidates on timelines the customer sets.
Data protection (StreamWeave)
Post-quantum encryption that splits data into pieces, protects each differently and sends them by different paths. It is itself a building block.
Nebulo
Identity, data and memory without files or databases. Access rules travel with the data through the Guard Meaning Coordinate.
Elevate
Brings existing code, data and AI models inside the trust boundary by wrapping them as building blocks. Nothing unwrapped can run inside a .wv stream.
Supercell and xSpot
Supercell orchestrates cloud and on-premises work. xSpot pools edge and IoT resources without virtualization.
Adopt at your own pace
Nothing forces a full replacement. The design lets a company adopt Essence in stages, and the stages after the first depend on the funded milestones.
A company can start with ordinary outputs and move one step at a time. The later steps depend on the funded milestones.
- By design, Essence can hand standard outputs to your existing pipeline, so scanners and signing tools work as they do today. Essence is not a code generator, and these outputs are an interface to existing tooling, not its purpose. File-based output is re-established during the first 30 days of funded milestone work. Regulated teams usually start here.
- Fixed builds stay where compliance needs them, while a controlled runtime is introduced for specific workloads. Different workloads in the same company can sit at different phases.
- Work runs with purpose and policy enforcement, audit telemetry and selective export. Trust and verification become continuous instead of a one-time build check.
Three starting paths
Optimize
Minimal change. Add meaning-driven layers where they reduce friction and risk.
Modernize
Lower operating cost, higher quality. Repackage systems with governed packaging and evidence.
Futureproof
Replace legacy where full replacement pays off most.
Choose a path by risk tolerance, compliance limits and how much legacy you want to keep.
A first step
Timelines are set per deployment, because adoption depends on infrastructure complexity, scope and integration. A practical first step is a bounded evaluation that your own team runs on a few of your own workloads, with success criteria set in advance. The Technical Evaluation Brief sets out how.
Go deeperSliding scale of useAdoption phasesAdoption maturity
Nothing runs until it says why
Traditional security guards the perimeter. Essence checks purpose, before and during every run.
Before and while work runs, the system checks why it is running and who is allowed to run it.
- It states who is asking, what it will do, when, where, how and why, before it can start. Nothing runs unless purpose and authority are declared and validated.
- The built-in check decides whether the request may run, based on identity, integrity and authority. Sensitive workloads can require policy approval first.
- While it runs, behavior is compared with what was declared. Revalidation repeats on timelines the customer sets, so trust is not a one-time event.
- Policy decides the action: alert, block, halt or remediate, including rolling back to the last validated state. The effect resembles an immune system built into the artifact. Overriding a plan is subject to contextual governance.
Go deeperValidation flowPolicy and purposeSecuriSync and StreamWeave
Not AI, and built to work with it
Essence does not use AI to decide or to execute what runs. It can work alongside AI systems, and it keeps people and policy in charge.
AI can propose. It cannot decide what runs. People and policy do.
Not AI
Essence does not use AI to decide or to execute what runs. The same request resolves to the same precise description every time, so results can be explained and repeated.
AI can propose
An AI system can connect to Essence and suggest what someone wants. It can only propose. It does not run anything itself, and its proposal is checked like any other request.
People and policy decide
Before anything runs, a built-in check confirms who is asking, that the request is intact and that they are allowed. Each run is designed to leave a record.
A common worry about AI is that it can act on its own, reach outcomes nobody intended, and leave no clear account of why. Essence is designed to answer that by separating two jobs. AI may suggest. Rules that people set decide what is allowed, and the record shows who asked for what and why. The result is AI that is useful without being in charge.
This describes the design. Today only the GPU optimization build runs. The built-in checks and the other parts return with the funded milestones.
Streams with nothing to aim at
Most files have a predictable shape, and attackers study that shape. Essence removes it.
Data is designed to travel in a form with no fixed shape, so there is little for an attacker to aim at.
- Headers and fixed structures give attackers the same known places to aim, over and over.
- The runtime generates binary streams, not binary blobs: no headers, no fixed structures, no stable injection targets. Code and data travel together inside the stream. The design aims to remove classic code-injection and man-in-the-middle patterns and to reduce exposure to many known and unknown vulnerabilities.
- The protection layer splits protected data into pieces, encrypts each piece with a different algorithm, and sends the pieces by different network paths. The scheme changes on every read and write, and the strength can be raised or lowered to match the threat. It is post-quantum by design. New algorithms can be added.
Go deeperThe .wv formatSecuriSync and StreamWeaveAdaptive vs scan-ready
Evidence for auditors
Proving what happened matters as much as preventing problems.
Each run is designed to leave a record that auditors can check.
- Every run records who started it, what executed, when, where, and under whose authority.
- Version history, ownership and change approvals attach to the run.
- Records bundle into signed, portable packages for customers, auditors and regulators.
- Telemetry feeds tools such as Splunk, Sentinel and QRadar. Packages map to the control language of SOC 2, ISO 27001, FedRAMP, HIPAA, PCI-DSS and NIST 800-53. This is evidence support for audits, not a substitute for certification, and certification pathways are in progress.
Go deeperEvidence export and SIEMRegulatory alignmentTrust certification
What has been shown
Validated results and earlier internal research are kept separate here, the same way the benchmarks page separates them.
GPU speedups are measured against the first run of the same job. Older internal results are kept apart and labeled.
Independently validated
Faster than the first run of the same job, across compute, rendering and data-processing workloads on a single GPU. Validated by AWS and the Rowan University Digital Engineering Hub, and replicated on OCI and GCP.
Energy reduction on compute workloads on a single GPU. Confirmed by AWS.
What we guarantee. Faster than the first run of the same job, measured on one GPU. It is not a promise for every workload. What we guarantee is real time optimization of data ordering, scheduling, memory, networking and hardware parallelization, which is what highly skilled engineers do by hand and cannot do in real time.
The gain comes from generating hardware-tuned GPU instructions at run time, not from new hardware. Results are workload-specific. Workloads with more parallelism land toward the top of the range, and serialized or I/O-bound ones toward the bottom.
Foundational research
Edge and physical AI
Edge computing and physical AI, such as robots and vehicles, run on many small machines with tight limits. Essence is designed to work out the steps on the machine that is actually there. The company sees both as massive opportunities for businesses. Results on edge hardware have not yet been measured.
Where things stand
Running today: the GPU optimization build at adaptwithchameleon.com, which generates GPU instructions at run time. File-based output is re-established during the first 30 days of funded milestone work. The OS-hosted runtime, the signal and video processing layer and the other capabilities return when the milestones are complete. Plain-language creation of every kind of building block depends on the second funded milestone. Validation programs are active with AWS, Dell and OCI, focused on efficiency, operational simplification and hybrid orchestration. Production-scale deployments and public case studies are forthcoming.
A sensible way to evaluate is to pick one bounded workload, measure cycle time, defect rate, evidence completeness and run cost, then expand once the numbers hold.
Live, recorded, or earlier deployment
The company shows Essence through one live tool, a library of recorded demonstrations, and earlier deployments. Each carries a different label, and the label matters.
Only one tool runs live today. Everything else is a recording or an earlier deployment, and each one is labeled.
- The one live build is adaptwithchameleon.com, a GPU evaluation tool. It generates a working design for each job in real time from Meaning Coordinates, not from code. Demos 01 and 02 show GPU instructions re-synthesized as workloads run. Pilot readiness is limited to this tool until the two milestones are complete.
- Demos 03, 11 and 12 (demonstrated, earlier version) show an app or a calculator made from spoken or typed words. These recordings show the earlier natural-language version, which is paused while the two funded milestones are completed. The homepage product creator is a plain-language matching demo that runs without the plain-language layer connected.
- Demo 04 shows live video over a very low-bandwidth connection. The measured result is HD video at 28 kbps on the Milan to Denver route. The benchmarks page lists it as the company’s own measurement, so it is kept apart from the independently validated GPU figures.
- This is the signal-processing layer of Essence, called illumin8. Demo 05 is a recording of real-time video enhancement, and the benchmarks page reports a 4K source reduced to 1K producing output that measurably exceeds the original. It is not running as a build today and returns when the milestones are complete. It is the company’s own demonstration; no independent validation is recorded.
- Demo 06 shows a browser where meaning drives interaction, so it carries the demonstrated label. The company describes the browser as its governed browser surface. The video is a demonstration, not an offer of a product.
- This is an earlier deployment, called UnCloak, for object identification, reported by the company. No demonstration is published for it, and it appears on the site as a supply-chain scenario.
- Three labels cover everything: live today, demonstrated, or earlier deployment. The natural-language recordings are marked earlier version. Recorded demonstrations of the earlier natural-language version are not a statement about what is available now.
The video library holds twelve demonstrations in all. Demos 07 to 10 cover TimeWarp, visual editing by words, personalization and conventional output, shown in earlier recordings, and are not diagrammed here.
Go deeperDemos and pilotsDemo videosLive GPU evaluation tool
Essence lets people state what they want, resolves it to a fixed set of meaning coordinates, checks purpose before anything runs, and keeps evidence of what happened. It can be adopted gradually, starting with ordinary outputs. Only the GPU optimization build runs today, and the later stages depend on the funded milestones.