Teaching machines with Andrei Manea – he began his career testing systems whose behaviour could be specified, reproduced, and checked. Since then, he has moved through software testing, development, and DevOps into AI research, where the same expectation of certainty no longer applies.
That progression has made him particularly attentive to the distance between what an AI system can demonstrate and what it can be trusted to do. We spoke with Andrei about probabilistic engineering, the value of efficiency, the limits of automated judgment, and the responsibility of the people building systems whose consequences may become visible only after deployment.
This interview is part of The Art of Solving, a series about the people behind Bytex and the experiences, instincts, and principles that shape how they approach difficult problems.
You began in systems testing, moved through software testing, development and DevOps, and now work with AI. What has each stage taught you about solving problems?
I studied Automation and Systems Engineering at the Technical University of Iași. The program gave me exposure to many layers of technology, from hardware and low-level programming to higher-level software, even if we did not go deeply into every one of them. That breadth helped me build a framework I could continue developing on my own.
My first professional role was in automotive systems testing, working largely with mathematical simulation environments such as MATLAB. It taught me how much coordination is required in a large organization. Hundreds of people may depend on the same resources, databases, procedures, and development schedules, and even a technically simple task can become difficult when those dependencies are not aligned.
I later moved into software testing, including unit testing in C and C++, and then began developing internal tools in higher-level languages. That was the point at which I understood that being familiar with many areas was useful, but I also needed to choose where to develop real depth.
At Bytex, particularly through my work with Cisco, the range opened up again. I worked across web development, security, deployment, Kubernetes, and other technologies, often with teams in different time zones. For the past two and a half years, my work has focused on research. Each stage added something different: systems thinking, specialization, independence, and the ability to learn quickly when the problem changes.
Traditional software is expected to behave according to defined rules. What changed in your engineering mindset when you began working with models whose outputs are probabilistic?
It felt, in a strange way, like working with people again. When someone joins a team and does not yet know the job, their response can be difficult to predict. With conventional software, I am working with a program whose rules can be inspected and tested. With an AI model, I am dealing with a system that can behave plausibly without behaving consistently.
Models have become better and more specialized, and people who work with them develop an intuition for how to guide them. Even so, you are still working with an imperfect approximation of human reasoning. The engineering task changes because producing the desired result once is not enough. You have to understand how often it will produce that result, what happens outside the expected context, and how the surrounding system responds when the model is wrong.
AI demonstrations can appear convincing under controlled conditions. What separates an impressive prototype from a system that can be trusted in real use?
The uncomfortable answer is that, from a customer’s perspective, trust can be influenced by how well the system is presented and sold. Benchmarks and demonstrations can be framed selectively. That gives people good reason to be cautious about some of the claims in the market.
A deterministic system can be tested against clearly defined behaviour. Depending on how it is licensed, its internal logic may also be inspected. With a large model, you cannot meaningfully examine billions of parameters and all the relationships between them. The burden therefore moves to evidence: how the model performs across different contexts, how the system handles failure, and whether the claims made about it are supported by testing that resembles actual use.
There is another complication. Models are becoming more fluent and more convincing, which makes them easier to trust emotionally. People can experience a sense of humanity in the interaction and begin using these systems for personal advice, including decisions about relationships. Fluency can create confidence before reliability has been established. A trustworthy system needs safeguards against that gap.
The competition between AI models is focused on efficiency as well as capability. Which matters more: having the most advanced model or designing the strongest system around it?
Models from different companies can be exchanged or integrated into the same system. The relevant cost includes the price of running them, but also compute, energy, water, and the human infrastructure required to keep them operating.
Capability will also reach a point of diminishing returns. If one model performs at 89% of the best model’s reasoning capacity but costs a fraction as much to run, that difference can be decisive. In most ordinary applications, a small gain in benchmark performance will not justify a dramatically higher operating cost.
There are also decisions for which a model should never be given authority, regardless of how precise it appears. In those cases, additional capability does not solve the underlying governance problem. The stronger system is the one that uses an appropriately capable model efficiently and sets clear limits on what the model is allowed to decide.
As AI becomes easier to integrate, what judgment should precede the decision to use it?
The first question should be whether the task requires AI at all. Much of the public use of generative AI is trivial, while the resources required to train and run the models are substantial. We should be more deliberate about whether the value of a use case justifies those costs.
There are useful applications. A model can summarize information, help someone examine a problem from several perspectives, generate initial ideas, or identify considerations the user may have missed. Those are supporting functions. The distinction I would make is between helping a person think and replacing the person who is responsible for the judgment.
Other forms of AI, including specialized predictive models, can have clearer value. They may be trained on carefully prepared data produced by experts for a defined purpose. That is a very different proposition from a general-purpose model trained to accommodate almost any public request.
What concerns you about allowing AI to replace human decision-making?
Judgment weakens when it is no longer exercised. We make judgments constantly, consciously and unconsciously, and that capacity can deteriorate surprisingly quickly when it is delegated.
An AI system can make a process faster, but speed does not answer the moral question. If the system makes the decision, responsibility becomes harder to locate, while the person using it gradually loses the practice of evaluating the situation independently. AI can assist with evidence, options, and patterns. It should not become a convenient way to remove the accountable decision-maker.
Medicine, education, and public administration are often described as major opportunities for AI. Which forms of work are ready for automation, and which require a higher threshold of evidence?
Medicine already offers credible examples. Specialized models can help identify diseases in medical imagery, for instance. Even there, the use case needs to be narrow, the evidence strong, and the final responsibility clear.
Education depends heavily on the type of model and the source of its knowledge. I would be open to a specialized system trained on a transparent body of educational material, with public methodologies and inspectable sources. It could support explanation, summarization, and guided study. I would not want a general-purpose model, shaped by opaque data and changing commercial decisions, to become an unquestioned teacher.
The higher the consequence of an error, the more demanding the standard should be. Helping a student organize information and making a medical or administrative decision about a person cannot be treated as equivalent deployments.
Every model reflects its training data and the choices made during development. Can bias ever be removed?
I do not think completely unbiased data exists. Scientific data comes closest, although even scientific systems begin with axioms and methodological choices. Once we enter social, political, or cultural questions, every dataset reflects a perspective.
The problem is not simply that bias exists. It is that some sources and decisions are given more weight than others, sometimes intentionally, through the choices made by developers, trainers, model providers, and the institutions supplying the data. A model can then present an average as though it were an objective answer.
That becomes dangerous because an average is not the same as a moral truth. If we ask a system to become “better” than people, we still have to ask: better for whom? Is it optimizing for an individual, for the majority, or for a supposedly universal principle? Every answer introduces another set of assumptions, and someone still has to choose the standard.
Could AI help people examine their own assumptions rather than deciding on their behalf?
In theory, yes. You could imagine systems that make different ethical frameworks explicit and help a person understand how the same problem appears through each one. That could encourage users to recognize their own assumptions instead of receiving one answer that appears neutral.
The practical difficulty remains governance. Who defines those frameworks? Who decides that they have been represented fairly? The idea is intellectually attractive, but implementing it without simply embedding a new authority inside the system is much harder.
Many careers depend on learning through repetitive or junior-level work. If AI takes over that layer, how will future specialists develop the experience required to make difficult decisions?
We should distinguish repetitive work from formative work. Endless coding, memorizing syntax, or fixing mechanical details does not automatically develop good engineers. Some of those tasks provided jobs and an entry point into the industry, but performing them indefinitely was never the same as learning how to reason about a system.
The danger is that people may end up serving the model rather than learning from the work. Career development is far more than a simple progression in which everyone moves from entry level to junior, middle, senior, and then leadership. People develop differently, and removing the early layer without redesigning how they learn will leave many of them without the experience required for the next one.
If routine implementation becomes more automated, systems thinking has to be taught earlier. Juniors will need to understand architecture, dependencies, failure, and trade-offs from the beginning. That requires a different educational model.
Engineers often see the consequences of a technology before the rest of society understands them. What responsibility does that give the people building AI?
This is difficult because responsibility and power are distributed unevenly. Individual engineers may understand parts of the system very well while having little influence over the commercial or political decisions behind its deployment.
The phrase “if I do not build it, someone else will” is tempting, and people use it to live with work they find ethically uncomfortable. I understand why. In a large organization, a lower-level contributor may be responsible for one component that is harmless in isolation. The potential for harm can emerge later, when the model owner changes a safeguard, a buyer integrates the component into another system, or the finished product is used in a context the engineer never controlled.
That does not erase personal responsibility, but it shows why individual conscience cannot be the only safeguard. Engineers need credible ways to raise concerns, refuse certain work, and know that those decisions will matter. Otherwise, the people controlling budgets and deployments hold the practical authority while responsibility is dispersed across everyone below them.
If an AI system causes harm, where should responsibility sit?
Responsibility should follow control. The people who finance the system, determine its purpose, remove safeguards, authorize deployment, and benefit from it should carry the greatest share. A model provider cannot treat the system as neutral if it retains control over its capabilities and restrictions. The deploying organization can’t claim it merely bought a tool if it chose the context and gave the system authority.
There will be cases in which an engineer knowingly designs a directly harmful function, and personal responsibility becomes clearer. In many complex systems, however, no single technical contribution produces the harm on its own. Accountability therefore has to reflect the complete chain of decisions rather than concentrating on the last person who touched the code.
Do you think society will develop a healthier relationship with AI as the technology matures?
I am somewhat pessimistic about the immediate future. The investment is large, governments are deeply involved, and deployment is moving faster than ethical or social governance. I do not expect restraint to emerge automatically from technological maturity.
We have governed other powerful technologies by creating limits around their use, but AI is diffuse and easy to integrate into many parts of society. That makes agreement harder. If the current economic momentum continues, I expect the technology to advance before institutions have settled the questions surrounding it.
Give us a best-case scenario five years from now. What would responsible AI deployment look like?
I do not expect artificial general intelligence within that period. A more credible and desirable outcome would be continued progress toward smaller, less expensive, and more efficient models that require less compute while performing defined tasks well.
The best case is a slower, steadier form of adoption. Specialized systems would be used where the need is clear, their evidence can be examined, and their role is limited. People would remain responsible for consequential decisions. Efficiency would improve enough to reduce some of the resource pressure, while governance would begin catching up with deployment.
Technology will still be political because it is created, funded, and used by people. Its design reflects choices about whose problem matters, which outcomes are desirable, and who receives power. Five years from now, I would like us to be more honest about those choices, rather than treating AI as an independent force that simply arrived.
