Why Senua AI is x10

The first pure causation AI engine

Senua AI is a globally unique, deep tech AI engine built on causation. It is the first pure causation AI engine in commercial use today.

Pure is a precise word here. There is no trained model inside Senua AI: no large language model, no small one, no neural network, no weights and no training run. The engine learns the causal structure of each stream it reads, from the stream itself, on ordinary low end CPUs, and it keeps learning on the job. A physicist would call what it learns a model of the process, the smallest one that predicts it, as the next section explains. It is not a model in the sense the AI industry sells: nothing is trained in advance on someone else's data, and nothing is stored as weights. Causal AI as an idea is not new. Gartner placed it at the very start of its 2024 Hype Cycle for AI, as the way to go beyond predictive models built on correlation [1], and several vendors sell causal tools built beside machine learning and, more and more, beside language models. Senua AI is causation all the way down.

This essay makes a bold claim and keeps it narrow. In the two places we work, physical AI and cyber security, a pure causation engine is ten times better than a correlation engine on the measures a business actually pays for: cost, speed, hardware, energy, data security and learning. Not on everything. On these. What follows is why, with the evidence: Stanford, MIT and the peer reviewed literature on what correlation costs and where it breaks, and our own published measurements on what causation does instead, each with its method and its limits, and most with the data to rerun it.

We do not use models. No LLMs. No SLMs. Of any kind.

No GPU. No data centre. No AI token costs, ever. Senua AI runs on low end CPUs.

Two ways to build an AI engine

Almost all of what is sold as AI today learns correlation. A large language model is trained to predict the next word of a text. To do that well it reads a large part of the public internet and stores what it has learned in billions of numbers, its weights. Judea Pearl, who won the Turing Award for his work on causal reasoning, puts association, seeing what goes with what, at the first level of his causal hierarchy, and calls it "the hallmark of current machine learning methods" [2]. It is a remarkable level. It writes fluent prose, summarises, translates and answers questions on almost any subject.

Senua AI learns causation in a precise sense that comes from physics. James Crutchfield and Karl Young introduced the causal states of a process in 1989 [3]: two histories of a system are the same state when they predict the same future. Cosma Shalizi and Crutchfield proved that causal states are the smallest summary of the past that predicts the future as well as anything can [4]. A causal state is a property of the system that produced the data, not of one collection of data. That is the whole difference, and every number in this essay follows from it.

The last row matters as much as the others. A causation engine is not a better chatbot, and we never sell it as one. If your product writes prose, keep whatever writes it. The ten times lives where the job is a stream and a decision.

What the correlation bill looks like

The cost of correlation is built into its architecture, and the industry's own figures show it.

• Training is expensive, and getting more so. Stanford's AI Index estimated that Google's Gemini Ultra cost about US$191 million of compute to train and OpenAI's GPT-4 about US$78 million. The original Transformer of 2017, the architecture underneath them, cost around US$900 [5]. Epoch AI found that the cost of the largest training runs has grown 2.4 times a year since 2016, and that the largest will cost more than a billion dollars by 2027 if the trend holds [6]. Meta's own model card for Llama 3.1 405B records 30.84 million GPU hours on 700 watt H100 processors and 8,930 tonnes of CO2e, location based [7].
• Running it takes data centres. The International Energy Agency puts data centres at about 415 terawatt hours in 2024, 1.5% of the world's electricity, and expects that to more than double to around 945 terawatt hours by 2030, with AI "the most important driver" [8]. In a peer reviewed study, Alexandra Sasha Luccioni, Yacine Jernite and Emma Strubell found that general purpose generative systems cost orders of magnitude more to run than task specific systems doing the same task, even at the same size [9].
• Tokens are cheaper, but they are still metered. Stanford records that the price of querying a model as capable as GPT-3.5 fell from US$20 to US$0.07 per million tokens in about eighteen months, a 280 fold drop [10]. That is real progress. But every token is still billed, and the work that matters in physical AI and cyber security is volume: every reading, every log line, every request.
• The returns are not arriving. MIT's Project NANDA studied over 300 public initiatives and found that despite US$30 to 40 billion of enterprise investment in generative AI, "95% of organizations are getting zero return" [11]. Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027, because of escalating costs, unclear business value or inadequate risk controls [12]. Meanwhile Stanford reports that 78% of organisations now use AI [10].

The MIT report names the cause in one line, and it is worth quoting in full: "The core barrier to scaling is not infrastructure, regulation, or talent. It is learning. Most GenAI systems do not retain feedback, adapt to context, or improve over time" [11]. A correlation engine is fixed when its training ends. A causation engine learns the system it serves, in place, every day. That is not a feature we added. It is what the architecture is.

Why correlation breaks when the world changes

A business does not need an engine that is right about yesterday's data. It needs one that is right about the system it is running today. That is where correlation is weakest, and some of the clearest evidence comes from MIT.

A team from MIT, Harvard and Cornell trained transformer models on New York taxi rides. The models gave turn by turn directions in New York City with near perfect accuracy. But when the researchers recovered the street map each model had learned, it bore little resemblance to the real one, and when they closed just 1% of the streets, accuracy fell from nearly 100% to 67% [13] [14]. The model had learned the routes, not the city. Another MIT team tested language models on counterfactual versions of eleven familiar tasks, the same task with one default assumption changed, and found that performance "substantially and consistently degrades compared to the default conditions" [15]. Robert Geirhos and colleagues gave the pattern its name, shortcut learning: decision rules that score well on the test and break under real conditions [16].

The same literature points to the remedy. Bernhard Schölkopf, Yoshua Bengio and colleagues argue that learning causal structure is the way to transfer and generalisation, the ability to keep working when conditions change [17]. MIT CSAIL demonstrated it in flight. Their drones navigated forests and cities they had never seen, through noise, rotation and occlusion, and the paper states the reason plainly: good performance outside the training conditions "requires that trained models learn representations that are causally associated with the task" and independent of the environmental context [18]. Their networks are small neural networks built to capture causal structure, so their result is evidence for the principle, not for our method. Senua AI applies the same principle in its pure form: no network at all, only the causal states themselves.

Ten times in physical AI

First, what we are not comparing with. A robot's control loops, a PID controller or a Kalman filter, run in microseconds, and they should; nothing here replaces them. The comparison is with the foundation models now sold as physical AI: vision, language and action models such as NVIDIA's GR00T N1, Physical Intelligence's π0 and Google DeepMind's Gemini Robotics, transformer networks that read camera images and instructions and output a robot's actions [34] [35] [36]. They carry the correlation bill above with them, all the way to the edge.

A robot does not meet a dataset. It meets a system: one motor, one gripper, one wing, one stretch of ground, each with its own wear and its own history. A Stanford led review of foundation models in robotics lists real time execution among the field's open challenges, and asks "how to deal with the high inference time of some foundation models which could hinder their deployment on robots" [19]. A machine that has to wait for a data centre is a machine that cannot act in time, and cannot act at all when the link drops.

Here is what we have measured, each figure on a published record with its limits:

• The whole engine fits on a fifteen dollar board. On a Raspberry Pi Zero 2 W, a board with 416 MB of usable memory, the complete Senua AI engine is 6.8 MB and peaks at 207 MB while holding 651 MB of knowledge on disk. It turned a plain English instruction into a drone mission, in valid MAVLink v2, in 286 ms, and sent nothing off the board [20]. This is a footprint result: the engine needs no GPU and no link. It is not a control loop timing.
• 469 times less processor time. On the same laptop, Senua AI and a one billion parameter language model answered the same 49 questions. Senua AI used 0.97 processor seconds in total; the language model used 454.65. The median answer took 0.017 seconds against 1.788, and neither fabricated an answer [21]. This is a question answering test, not a robot task. It measures what each architecture costs to run, and that cost goes wherever the architecture goes.
• A rounding error on a drone's battery. In simulation, carrying the engine on a drone's companion board costs 0.6% of the flight pack per flight, a figure three independent implementations agree on and which is still to be metered on hardware [22].

Two honest limits sit beside these numbers, and both are on the datasheet. Processor seconds stand in for energy; they are not joules. And the two systems do different jobs: the language model can chat about anything, while Senua AI answers from verified knowledge or declines [21]. For a robot, a wearable or a machine on a factory floor, the second job is the one that matters.

A head to head on a physical task against a robot foundation model is the next measurement we will publish. We expect the same order of difference, because the cost sits in the architecture, not in the task.

Ten times in cyber security

Cyber security is a stream problem. Logs, requests, sessions and processes arrive continuously, and the job is to notice, in time, the moment one of them stops looking like the system's normal behaviour. Correlation has a long and difficult history here. Robin Sommer and Vern Paxson showed in 2010 that intrusion detection is fundamentally harder for machine learning than other fields, not least because errors are so costly [25]. In 2022 Daniel Arp and colleagues reviewed 30 papers from the top security conferences and found that every one suffered from at least three of ten common pitfalls. One of them is spurious correlation: an intrusion detector that learns the IP range most attacks happened to come from, "instead of solving the actual task" [26].

So there are two things to compare with. The incumbents run locally and do real work: web application rule sets, the correlation rules in a SIEM, endpoint agents. The newer arrivals are language model copilots, and they are metered. Microsoft prices Security Copilot by the security compute unit, provisioned by the hour; its own worked example uses US$4 an hour for each provisioned unit and US$6 for overage [37]. Read raw logs through a language model and the arithmetic is unforgiving. OpenAI's rule of thumb is about four characters of English text per token [27], so one gigabyte of logs is about 250 million tokens. At US$0.07 per million, the price Stanford recorded for a GPT-3.5 level model in late 2024 [10], every gigabyte read costs US$17.50, every time it is read. That is an illustration, not a quote, but the shape holds: the bill grows with the traffic, which is the one thing a security team cannot control.

Senua AI learns each system's normal behaviour from the system itself and judges every event against it, on a CPU, inside your network:

• Against the industry's own rule set. The OWASP Core Rule Set is the standard defence for web applications. On the same 30,532 requests, with nothing tuned on either side, Senua AI stopped 66.9% of attacking sources against the rule set's 64.7%, and wrongly blocked half as many genuine visitors: 1 in 18,000 against 2. After learning from confirmed attacks, at the same false block rate, it blocked 67.4% of attacking requests it had never been given. The rule set blocked 28.0%. That is 2.4 times [38]. The attacks in that dataset are generated from templates, so the learning result shows new instances of known attack types.
• Web attacks stopped inline. On the same public dataset, CSIC 2010, 91.69% of attacking sources were stopped by their third request and every one by the tenth, while none of 6,000 genuine sources was blocked [28]. CSIC 2010 is generated traffic from 2010: the field's long standing baseline, not a portrait of today's attackers.
• At line speed on one core. Judging that traffic added about 0.09 ms to each request, handled about 11,000 requests a second and peaked at 29 MB of memory [20].
• Hosts. On ADFA-LD, the host intrusion dataset published in 2013 by UNSW Canberra at the Australian Defence Force Academy, attack sessions ranked above normal operation at an AUC of 0.9246, learned from normal behaviour only [28].
• Agents too. An automated agent working a replica portal, with synthetic normal traffic, was blocked by its third request, with no normal session affected [28].

There is no token meter in any of this. Ten times more traffic is more work for a CPU core, not ten times the invoice.

The security case: nothing leaves

For a business decision maker the deepest difference may be the simplest. An engine that runs on a remote model has to send your data to it. An engine that runs inside your network does not.

• Data sent out does not come back. In 2023 Samsung restricted generative AI tools after engineers uploaded internal source code to a public chatbot, concerned that data on external servers is hard to retrieve and delete [29]. Australia's privacy regulator now recommends that organisations do not enter personal information, and particularly sensitive information, into publicly available generative AI tools [30].
• AI is becoming the breach. IBM's 2025 Cost of a Data Breach study found that 13% of organisations reported breaches of AI models or applications, and 97% of those lacked proper AI access controls. Organisations with high levels of unsanctioned "shadow" AI saw breaches cost on average US$670,000 more, against a global average breach of US$4.44 million [31]. Stanford reports that AI incidents logged by the AI Incidents Database reached 233 in 2024, a record, up 56.4% [10].
• Language models take instructions from what they read. The OWASP Top 10 for large language model applications ranks prompt injection first and sensitive information disclosure second [32]. A model that reads an attacker's log line can be told what to do by it.

Senua AI makes no API calls to anyone's model; it has no trained model to call. It does not call home. It does not take instructions from the data it reads; it measures that data against the system's own normal. What it learns stays on your hardware, sealed with AES-256-GCM under keys sent with ML-KEM-1024, the post quantum standard the Australian Information Security Manual names, in our own cryptography with no third party library [33].

Adding it up

Put the cost lines a business actually pays side by side, and the shape of the difference is clear.

Some of these multiples are well past ten. Some are not multiples at all, because one side of the line is zero. We call it ten times because that is the floor we are confident of, in the places we work, on the measures above.

What we do not claim

We do not claim that causation replaces language models everywhere. It does not write your marketing copy or chat with your customers. We do not replace a robot's control loops or a security team's rules; we work beside them, where the job is to learn a system and judge what it does. We do not claim that every number we publish is ten times; each one names its dataset, its method and its limits, and where a result was close, the datasheet says so.

And we do not claim more checking than has been done. Every Senua AI figure in this essay is our own measurement, published with its method and, for most, the data to rerun it. None has yet been reproduced independently. Anyone who wants to check them is welcome to, and we will help.

Why being first matters

The AI industry measures the cost of correlation in billions of dollars, terawatt hours and breach reports. An MIT report says the missing piece is learning. Gartner placed causal AI at the start of its 2024 hype cycle for AI. Senua AI is causal AI in its pure form: a causation engine with no trained model of any kind inside it, built in Australia from the programming language up, running today on hardware your business already owns.

If your product or your operation reads streams and has to decide, whether that is a robot, a machine, a network or a security operation, talk to us. We license the engine, it runs on your own hardware, and your data never leaves.

We do not use models. No LLMs. No SLMs. Of any kind.

No GPU. No data centre. No AI token costs, ever. Senua AI runs on low end CPUs.

References

1. Gartner, Hype Cycle for Artificial Intelligence, 2024 (17 June 2024), which places causal AI at the innovation trigger; definition as quoted in causaLens recognized in Gartner Hype Cycle for AI 2024, causaLens, 5 July 2024. The report itself is for Gartner subscribers.
2. J. Pearl, The seven tools of causal inference, with reflections on machine learning, Communications of the ACM 62(3) (2019) 54 to 60.
3. J. P. Crutchfield and K. Young, Inferring statistical complexity, Physical Review Letters 63 (1989) 105 to 108. Peer reviewed.
4. C. R. Shalizi and J. P. Crutchfield, Computational mechanics: pattern and prediction, structure and simplicity, Journal of Statistical Physics 104 (2001) 817 to 879. Peer reviewed.
5. Stanford HAI, AI Index: state of AI in 13 charts, AI Index Report 2024. Training costs estimated with Epoch AI; compute only.
6. B. Cottier, R. Rahman, L. Fattorini, N. Maslej, T. Besiroglu and D. Owen, The rising costs of training frontier AI models, Epoch AI, arXiv 2405.21015 (2024).
7. Meta, Llama 3.1 405B model card: training hardware and greenhouse gas emissions.
8. International Energy Agency, Energy and AI: executive summary, April 2025.
9. A. S. Luccioni, Y. Jernite and E. Strubell, Power hungry processing: watts driving the cost of AI deployment?, ACM Conference on Fairness, Accountability, and Transparency (FAccT 2024). Peer reviewed.
10. Stanford HAI, AI Index 2025: state of AI in 10 charts, April 2025.
11. A. Challapally, C. Pease, R. Raskar and P. Chari, The GenAI divide: state of AI in business 2025, MIT NANDA, July 2025. Preliminary findings from 300+ public initiatives, 52 interviews and 153 survey responses; not peer reviewed.
12. Gartner, Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027, press release, Sydney, 25 June 2025.
13. K. Vafa, J. Y. Chen, A. Rambachan, J. Kleinberg and S. Mullainathan, Evaluating the world model implicit in a generative model, NeurIPS 2024. Peer reviewed.
14. MIT News, Despite its impressive output, generative AI doesn't have a coherent understanding of the world, 5 November 2024.
15. Z. Wu, L. Qiu, A. Ross, E. Akyürek, B. Chen, B. Wang, N. Kim, J. Andreas and Y. Kim, Reasoning or reciting? Exploring the capabilities and limitations of language models through counterfactual tasks, NAACL 2024, 1819 to 1862. Peer reviewed.
16. R. Geirhos, J. H. Jacobsen, C. Michaelis, R. Zemel, W. Brendel, M. Bethge and F. A. Wichmann, Shortcut learning in deep neural networks, Nature Machine Intelligence 2 (2020) 665 to 673. Peer reviewed.
17. B. Schölkopf, F. Locatello, S. Bauer, N. R. Ke, N. Kalchbrenner, A. Goyal and Y. Bengio, Toward causal representation learning, Proceedings of the IEEE 109(5) (2021) 612 to 634. Peer reviewed.
18. M. Chahine, R. Hasani, P. Kao, A. Ray, R. Shubert, M. Lechner, A. Amini and D. Rus, Robust flight navigation out of distribution with liquid neural networks, Science Robotics 8 (2023) eadc8892. Peer reviewed.
19. R. Firoozi, J. Tucker, S. Tian, A. Majumdar, J. Sun et al., Foundation models in robotics: applications, challenges, and the future, International Journal of Robotics Research (2025). Peer reviewed.
20. Senua AI, What it takes to run: the Raspberry Pi Zero 2 W run of 21 September 2026 and the CSIC 2010 throughput on one core; and the datasheet The whole engine on a Raspberry Pi Zero 2 W, 24 September 2026, with its data.
21. Senua AI, Processor time against a local language model, datasheet, 24 September 2026 (llama3.2:1b; data published beside it).
22. Senua AI, Datasheets: edge, integrity and learning: companion power on a drone (DS-001).
23. Raspberry Pi, New product: Raspberry Pi Zero 2 W on sale now at $15, 28 October 2021.
24. CNX Software, NVIDIA increases the price of Jetson modules and devkits by up to 101%, 22 July 2026: Jetson Orin Nano Super developer kit US$399, Jetson AGX Thor developer kit US$5,499.
25. R. Sommer and V. Paxson, Outside the closed world: on using machine learning for network intrusion detection, IEEE Symposium on Security and Privacy 2010. Peer reviewed.
26. D. Arp, E. Quiring, F. Pendlebury, A. Warnecke, F. Pierazzi, C. Wressnegger, L. Cavallaro and K. Rieck, Dos and don'ts of machine learning in computer security, USENIX Security 2022 (distinguished paper); summarised as Pitfalls in machine learning for computer security, Communications of the ACM 67(11) (2024). Peer reviewed.
27. OpenAI Help Center, What are tokens and how to count them: about four characters of English text per token, a rule of thumb.
28. Senua AI, Measured results and security datasheets, each with its data and a pass mark set before the run.
29. CSO Online, Samsung bans staff AI use over data leak concerns, May 2023, reporting Bloomberg.
30. Office of the Australian Information Commissioner, Guidance on privacy and the use of commercially available AI products, 21 October 2024.
31. IBM, IBM report: 13% of organizations reported breaches of AI models or applications, 97% of which reported lacking proper AI access controls, 30 July 2025 (Cost of a Data Breach Report 2025, with the Ponemon Institute).
32. OWASP GenAI Security Project, OWASP Top 10 for LLM applications 2025.
33. Senua AI, Our first quantum-safe steps are live, ahead of their dates, 21 October 2026, and the cryptographic register.
34. J. Bjorck et al. (NVIDIA), GR00T N1: an open foundation model for generalist humanoid robots, arXiv 2503.14734, March 2025.
35. K. Black et al. (Physical Intelligence), π0: a vision-language-action flow model for general robot control, arXiv 2410.24164, October 2024.
36. Gemini Robotics Team (Google DeepMind), Gemini Robotics: bringing AI into the physical world, arXiv 2503.20020, March 2025.
37. Microsoft, Microsoft Security Copilot pricing: provisioned and overage security compute units billed by the hour, with a worked example at US$4 and US$6 an hour; an allowance for Microsoft 365 E5 customers. Checked 21 October 2026.
38. Senua AI, It starts level with the rule set. Then it learns., and the datasheet Against the OWASP Core Rule Set, and after learning, 24 September 2026, with its data: CSIC 2010, the OWASP Core Rule Set on ModSecurity v3 at paranoia level 1, nothing tuned on either side.