Jean-Christophe Huc

CTO / VP Engineering | Real-Time Systems, Simulation, 3D, AI & Robotics | Ex-Airbus → Scaling Eng Orgs | Paris 🇫🇷

[Jean-Christophe Huc @cto_software]

C’est aussi un mécanisme économique assez classique : quand le coût d’une tâche s’effondre, la valeur se déplace vers les capacités complémentaires qui restent rares.
Si la démonstration formelle devient largement automatisable, la rareté migre vers la formulation des bons problèmes, les abstractions, les conjectures, le jugement et la capacité à décider quelles pistes méritent d’être explorées.
L’automatisation ne supprime pas nécessairement une discipline. Elle en déplace la frontière de valeur.

[Jean-Christophe Huc @cto_software]

@Ronald_vanLoon @antgrasso Exactly. Authentication proves who entered the system. It does not tell us what that identity should be allowed to do once inside.
With AI agents and automated workflows, security increasingly depends on runtime permissions, continuous authorization, observability and revocation.

[Jean-Christophe Huc @cto_software]

@Corix_JC @InfoWorld @mjasay Enterprise trust cannot be purchased as a technology layer.
It comes from how the system operates: clear authority, observable decisions, explicit boundaries, accountability and the ability to recover when things go wrong.
Trust is ultimately a property of the operating model

[Jean-Christophe Huc @cto_software]

Exactly. The gap between notebook and production is usually where hidden assumptions surface: interfaces, data contracts, observability, failure modes, ownership, deployment constraints and recovery.
A better model rarely fixes those boundaries.
Production reliability is often an interface problem before it is a model problem.

[Jean-Christophe Huc @cto_software]

Trust in AI agents will depend less on persuasion and more on control.
Users need to know what the agent can do, what it cannot do, what evidence it used, and whether its actions can be challenged or revoked.
Trust becomes durable when autonomy is observable, bounded and accountable.

[Jean-Christophe Huc @cto_software]

The interesting shift is that AI is forcing CIOs to redesign work, not just technology.
Once agents can coordinate tasks, make recommendations and act across workflows, the operating model becomes part of the architecture: decision rights, accountability, handoffs, controls and escalation paths all matter.
AI transformation becomes real when the organization itself is rearchitected around faster, clearer decisions.

[Jean-Christophe Huc @cto_software]

What resonates most with me is the distinction between comfort and agency.
In system-design terms, human agency cannot simply mean keeping a person “in the loop.” It means retaining the ability to define objectives, set boundaries, challenge decisions, choose alternatives — and revoke delegated authority.
Otherwise we may automate more and more of the decision process while leaving humans with little more than consent at the edges.
Abundance can improve outcomes. Agency determines whether people remain authors of those outcomes.
That may be one of the most important governance questions of the AI age.

[Jean-Christophe Huc @cto_software]

Long-horizon agents make reliability a state-management problem as much as a model problem.
The longer the horizon, the more opportunities there are for context drift, stale assumptions, permission changes and silent divergence from the original objective.
That is why verifiability needs checkpoints across the execution path — not only evaluation at the end.
Reliable autonomy requires continuous evidence that the system is still doing the right thing for the right reasons.

[Jean-Christophe Huc @cto_software]

@Corix_JC @CyberScoopNews Agentic AI turns legal accountability into an architecture problem.
If we cannot reconstruct who authorized an action, what permissions applied, what evidence the agent used and why it crossed a boundary, liability will remain ambiguous.
Autonomy needs auditability by design.

[Jean-Christophe Huc @cto_software]

@antopatrex1 The data bottleneck is real, but volume alone won’t solve it.
What matters is whether the data covers the failure modes, edge cases and regime changes the robot will actually encounter in deployment.
More data improves learning. Better coverage improves reliability.

[Jean-Christophe Huc @cto_software]

https://t.co/Vk4eiOZgoq

[Jean-Christophe Huc @cto_software]

https://t.co/IsfoEGPZae

[Jean-Christophe Huc @cto_software]

This is a great example of why “the system is up” is not the same as “the system is working.”
Silent fallbacks are especially dangerous in AI pipelines because they preserve availability while degrading decision quality.
The runtime needs to surface not only exceptions, but also mode changes, confidence loss, fallback activation and output quality drift.
Technical availability without semantic correctness is a hidden failure mode.

[Jean-Christophe Huc @cto_software]

This is a key gap.
If AI is going to participate in scientific discovery, evaluating the final answer is not enough. We also need to evaluate the process: which evidence it selected, which hypotheses it formed, which experiments it proposed, and how results changed the next decision.
Scientific capability is not just producing an answer. It is running a disciplined feedback loop.

[Jean-Christophe Huc @cto_software]

@0ivankaTrump1 6 holes in the shirt and 494 out.

[Jean-Christophe Huc @cto_software]

Trust in AI will not come from reassurance or policy alone.
It has to be engineered into the operating model: clear authority, traceable decisions, runtime observability, verification and recovery.
The more autonomous systems become, the more trust becomes a systems property — not a communication problem.

[Jean-Christophe Huc @cto_software]

Exactly. The hard part is not executing the nominal task — it is handling the gap between the expected world and the observed one.
A robust robotic system needs to detect that divergence, reassess the state, adapt the plan, and know when recovery is still safe.
The demo proves capability. The edge cases reveal whether you actually built a system.

[Jean-Christophe Huc @cto_software]

This is not only a workforce issue. It is an operating-model issue.
Many entry-level tasks were inefficient, but they also created the apprenticeship path through which people learned systems, judgment and context.
If AI removes that layer, organizations need to redesign how capability is built — through mentoring, simulation, progressive responsibility and explicit feedback loops.
Automating the first rung without rebuilding

[Jean-Christophe Huc @cto_software]

The training environment may become as important as the model itself.
If the environment fails to preserve the constraints, failure modes and uncertainty that matter in deployment, the system can learn the wrong lesson very efficiently.
Simulation creates leverage only when the learning loop closes back to reality through verification and transfer.

[Jean-Christophe Huc @cto_software]

NO83 / Decision Note #01

I’m building NO83, an experimental crypto decision system.

The goal is not to “predict the market”.

It is to answer a harder question:

How do you make better decisions when the future remains uncertain?

One principle became fundamental very quickly:

A winning trade can still be a bad decision.
A losing trade can still be a good one.

At T0, NO83 only knows:

• signals
• market context
• competing hypotheses
• risk

The outcome comes later.

So the system keeps two things strictly separate:

WHAT WE KNEW WHEN WE DECIDED

from

WHAT HAPPENED AFTER

If you judge a decision using information that only existed afterwards, you are not learning.

You are rewriting history.

Decision quality ≠ outcome quality.