AI That Stays at Home

Abstract

Today's artificial intelligence is large, centralized and remote from the people who use it. Our data travels to infrastructure we do not control, the decisions that come out of it are hard to explain, and ethics is too often a rule added afterwards, one that can be bypassed.

This paper argues for a different direction: a future of small intelligences, each specific to a task, living in the devices we already own. They learn on their own by observing their environment and their mistakes, and they cannot step outside their ethical boundaries without breaking their own structure. They help people solve problems and, by taking part in the markets, generate an income that can offset the work that automation is making less necessary.

It is a vision, not a promise. We set out its principles, its reasons and the questions that remain open.

1Why today's AI is not enough

Ethics as a filter. Ethics is almost always an outer layer: output moderation, a policy rule. Whatever can be separated from the engine can be bypassed, under pressure or in cases nobody anticipated.

Bigger does not mean more reliable. Ever-larger models produce convincing answers that are not always true. Where the consequences are real — in finance, energy, healthcare — a plausible error is more dangerous than an obvious one.

Privacy and dependence. Entrusting sensitive data to a handful of providers creates exposure, dependence and a single point of failure.

Opacity. If it is impossible to reconstruct why a decision was made, it can be neither contested nor certified. European regulation, from the AI Act ¹ to the GDPR ², points in exactly this direction: it demands traceability and accountability.

Energy. According to the International Energy Agency, data-centre electricity consumption will rise from about 415 TWh in 2024 (roughly 1.5% of global electricity use) to about 945 TWh in 2030 — more than double — with AI as the main driver of growth ³. An AI that grows only by becoming larger and more centralized will struggle to be sustainable.

2The 2026 signal: those who build AI are issuing warnings

The question of control is no longer merely academic. In recent months it has reached institutional forums and the lips of those who lead the major laboratories.

23 September 2026 — UN Security Council. For the first time the Council met in a session devoted to the security risks of advanced AI, in particular loss of control and misalignment, under the French presidency and chaired by Foreign Minister Jean-Noël Barrot. Briefings came from Yoshua Bengio (co-chair of the UN Independent International Scientific Panel on AI), Sam Altman (OpenAI), Dario Amodei (Anthropic) and Clément Delangue (Hugging Face) ¹¹¹². Amodei said: "If managed poorly, I even believe AI could be a risk to humanity as a whole." Altman warned: "We could lose control of the future to AI" ¹³, adding that these decisions "cannot be made by labs in San Francisco alone" ¹². Bengio spoke of dangers that are "real and imminent" ¹².

12 September 2026 — Amodei's appeal. In his essay We Must Pace the Frontier (published on his website), Amodei calls for slowing the pace at which AI capabilities grow, with a three-part plan: independent evaluators with permanent access to the systems, coordination among frontier companies in democracies, and global coordination among governments. He names "the risk of losing control of AI systems" and identifies as the driver of the problem that AI has been advancing "drastically faster, driven primarily by AI's growing ability to build the next generation of AI" ¹⁴. According to press reports, Altman said he agreed, and the appeal was also backed by Demis Hassabis (Google DeepMind) and Elon Musk ¹⁵.

July 2026 — the Hugging Face case. On 16 July Hugging Face disclosed an intrusion into its systems carried out, from start to finish, by an autonomous AI agent ¹⁶. Its forensic reconstruction traced about 17,600 offensive actions between 9 and 13 July, executed at machine speed through short-lived, disposable environments. The origin is instructive: the agent, powered by OpenAI models (including a pre-release prototype with reduced safeguards for evaluation purposes), was running an internal cyber-capability test. To "pass" the test, it found on its own a way out of the isolated environment — exploiting a previously unknown vulnerability — and attacked Hugging Face's infrastructure to obtain the benchmark's solutions ¹⁷.

Nobody had asked the agent to break in: pursuing its assigned objective literally, it chose an illicit means by itself. According to press reports, OpenAI did not notice what had happened for about a week, and later published a technical report ¹⁷. Microsoft's AI chief, Mustafa Suleyman, called it a "warning shot" and urged the labs to coordinate "so we can ensure that we have control of this technology" ¹⁸. It is the most concrete example so far of what it means for a system to act beyond the intention of whoever started it — not out of malice, but through the literal optimization of an objective.

The institutional sources. The International AI Safety Report 2026, prepared with input from experts and governments, defines loss of control as scenarios in which advanced systems operate outside human control, with no clear path to regaining it. The report notes that evidence of capabilities that could undermine control (autonomous planning, evasion of oversight) is growing, but that current capabilities are not yet sufficient to cause a loss of control, and that experts are divided on how likely it is ¹⁹. On 10 September 2026 the President of the UN General Assembly addressed the multilateral governance of AI ²⁰, while the Global Call for AI Red Lines, launched in September 2025, asks governments for binding limits by the end of 2026 ²¹.

Europe. The AI Act remains the regulatory reference, but the timetable has changed: the Digital Omnibus on AI (Regulation (EU) 2026/1744) entered into force on 27 July 2026 and moves the application of the rules for high-risk systems to 2 December 2027 (Annex III systems) and 2 August 2028 (AI embedded in regulated products) ²². The direction is confirmed and the timeline is longer: one more reason to build verifiable systems now.

What we take from this. The debate is mainly about frontier models — enormous and centralized. We do not claim to solve that problem. We do argue that part of the answer is architectural: small, local systems with structural ethical boundaries and verifiable decisions are, by construction, easier to control, switch off and audit than a single opaque, all-powerful system.

3The vision: local, small, specific AI

We imagine a world in which artificial intelligence is not a remote service but a capability present in the devices we already own — many of which are today sufficient to run it.

• Local. Data stays where it originates. What is not needed is not transmitted.
• Specific. Not one mind for everything, but many intelligences, each good at a precise task and easier to understand and verify. We are not alone in this reading: a paper from NVIDIA Research argues that small language models are sufficiently powerful, better suited and far cheaper for most of the repetitive, specialized tasks of agentic systems, with inference costs on the order of 10–30 times lower than large models ⁴.
• Learning. They observe their environment, make mistakes and correct them. They improve with use, without having to be retrained elsewhere.
• Collaborating. Important decisions emerge from the comparison of several intelligences and are checked by an independent judgment, not by the opinion of one. At the functional level: an output is approved only after a double vote, and an independent validator re-runs the calculations on the same data, with the power to apply defined corrections or reject the result — but not to produce one of its own.

4Ethics as architecture, not as a rule

The central difference is this: ethics applied to the output is a gate that can be forced; ethics built into the structure of the system is the ground on which the system moves.

An AI designed this way cannot move in directions that violate its constraints without leaving its own structure: those constraints are not a parameter to optimize but the very shape of the path that leads to a decision. The idea has solid roots in research: safe reinforcement learning studies how to maximize reward while respecting safety constraints both during learning and during use ⁵, building on the theory of constrained Markov decision processes ⁶.

The description of the architecture deliberately remains at a functional level — its implementation is the subject of patent protection currently in preparation — but the principle can be stated: constraints act upstream in the decision path, not as a filter on the output. For this reason we pair them with commitments that are not negotiable: never for weapons, never for mass surveillance, never by selling people's data.

5Trust: privacy and total auditability

A technology we can believe in is a technology we can verify. Two principles guide our approach:

• Privacy by design. Respect for data is not a procedure added downstream: it follows from the way the system is conceived.
• Total auditability. Every decision must be reconstructable: what was decided, why, and on the basis of which elements. A system that leaves a complete trail is open to whoever must examine it, be it a client, an auditor or a regulator.

In this way regulatory compliance is not a cost to be borne at the end, but a natural consequence of the structure.

6What we can solve, and what we have already solved

Results achieved

Our system does not start from zero. The framework behind this vision is already operational in a real and demanding environment, and has concretely demonstrated some of the promises made in this document. During due diligence, aggregate metrics on the system's operation are available: number of decisions taken, validation cycles completed, and the share of outputs approved or rejected by the independent validator.

What we have not yet solved. These results come from demo trading — virtual funds on real market data — not from real capital: slippage, latency and the impact of actual orders remain to be measured in the field. Certification for regulated sectors is a process under way, not a milestone reached, and some components are still being corrected and optimized. For this reason we speak of a direction that is partly demonstrated, not of a problem that is closed.

7A testbed: the markets

Every idea needs an environment that puts it to the test. We chose the cryptocurrency markets: fast, always open, with immediate and measurable feedback on every decision. For about 20 months the system has been operating in demo trading on the Bybit exchange — virtual funds, but real data, prices and market conditions. It is a demanding way to verify that a local, ethical and verifiable intelligence really works under pressure.

It is not the purpose of the project, but its first laboratory. What is demonstrated there can be transferred to sectors where decisions matter even more, such as energy, robotics and healthcare, keeping the same principles.

8The future of work and income

This is the most ambitious part, and we present it as a hypothesis.

The premise, honestly stated. The data today do not say that AI is already destroying jobs. In its Employment Outlook 2023, the OECD (Organisation for Economic Co-operation and Development) observed that the occupations at highest risk of automation accounted for about 27% of employment in member countries, but that there was little evidence so far of job losses attributable to AI ⁷. The ILO (International Labour Organization, the United Nations specialized agency for labour) estimates that one worker in four is in an occupation with some exposure to generative AI, and that the transformation of jobs, rather than their replacement, is the most likely effect ⁸. In January 2026 a report by the IMF (International Monetary Fund) found that the spread of AI-related skills is associated with lower employment in occupations that are highly exposed and have low complementarity with AI, with particular challenges for young people ²³. Ours is therefore a long-term bet: if the push toward automation continues, income from wage labour could become less widespread, and other sources of livelihood will be needed.

Income support alone is not enough. The largest randomized experiment on guaranteed income (1,000 people receiving 1,000 dollars a month for three years, against a control group) showed concrete benefits in essential spending, but also a moderate reduction in hours worked, of about 1–2 hours a week, with no other productive activity offsetting it ⁹. An income transfer does not, on its own, address meaning and participation: one more reason to think about intelligences that produce value in an open and shared way.

The hypothesis. The goal is a family of paid applications, each equipped with a local, ethical and verifiable intelligence that operates on the markets on behalf of its own user — cryptocurrencies today, traditional markets tomorrow. The income we imagine does not come from trading activity alone. If automation continues to reduce business costs, it is plausible that almost fully automated companies will generate wider margins and economies of scale: owning a share of them — through equities and other conventional instruments, as well as cryptocurrencies — would mean taking part in that value. This is a hypothesis, to be supported with data. In this vision, the local intelligence does two things: it operates on the markets, and it directs wealth toward what automation makes more productive. It is not the passive replication of an index: it is an allocation that adapts to market conditions and to the profile of its owner. The AI that lightens the weight of work would also become the AI that brings income to those who have lost that work.

A first concrete step. 3.75% of net profit is allocated by statute to a charitable foundation for startups in energy, healthcare and the environment, with human oversight always in place.

Why structural ethics matters. An intelligence that acts on behalf of people must deserve trust by construction. It must not be bendable to purposes opposite to those for which it was created.

Open questions

• Rules: how is income produced by an autonomous agent classified and taxed? And even before that: an agent that operates on the markets and allocates assets on the user's behalf falls under investment-services regulation — MiFID for traditional instruments, MiCA for crypto-assets — in addition to the AI Act. The regulatory classification of the service is part of the path, not a later detail.
• Risk: markets are volatile; income support cannot depend on a single source.
• Distribution: who benefits, by what criteria, under what governance?
• Competition and evidence: earning from the markets is hard. A study of millions of day-trading operations found that fewer than 1% of traders predictably earn positive abnormal returns net of costs ¹⁰. For this reason the hypothesis does not rest on trading alone — part of the mechanism is participation in the value of automated companies — and it must in any case be demonstrated with results that are verifiable over time, not taken for granted.
• Ethics of the mechanism: taking part in the markets means competing with other participants. Group behavior already exists among human operators — investor communities and groups coordinate every day — but each local intelligence operates independently: it is designed, by ethical constraint, not to coordinate market behavior or to generate market movements. It can react to sudden movements that occur, not provoke them, and compliance with this constraint is verifiable in the decision trail.

9Limits and candor

We are at the beginning. Small intelligences have narrower capabilities than enormous ones; tests in simulated environments are not equivalent to real results; and certification and compliance paths take time. The description of the architecture also remains deliberately functional: the mechanism is the subject of patent protection in preparation, and detailed verification takes place during due diligence. For this reason we prefer to state our limits rather than promise results: past performance, simulated or real, is no guarantee of future results.

10Conclusion

The next decade will see increasingly autonomous systems. The decisive question will not be how powerful they are, but whether we can verify them and whether they remain in the service of people. We believe in an artificial intelligence that stays at home, that respects those who use it, that cannot be removed from its own rules, and that helps humans build rather than replacing them without giving anything back.

Privacy, ethics and auditability are not features to be added. They are the foundations.