AI adoption has become an existential question for enterprises. That was the central thread running through a recent InfoQ podcast episode featuring Meryem Arik of DoubleWord (formerly TitanML), Clara Higuera Cabañes of BBVA, and Jeff Smith of C Proof. Host Olimpiu Pop gathered the three experts to break down what is actually happening as companies rush to integrate AI into their operations, and what comes next as the technology matures.
Adopting AI Because Everyone Else Is
Arik, who has spent the last three to four years working on inference optimization at DoubleWord, framed the urgency plainly. Companies are not adopting AI simply to gain efficiency. They are adopting because their competitors are doing the same, and the downstream effects on competitive advantage are already materializing. Every business leader recognizes that if they cannot adopt AI in a way that meaningfully changes their operations and output, they risk being outcompeted within the next decade.
Arik described the moment as an industrial revolution, though she preferred the word augmenting over replacing. The workforce is not being eliminated, but the fundamentals of how work gets done are changing. Companies that continue operating as they did before AI became available will find themselves at a disadvantage that compounds over time.
From Model Selection to Trust and Governance
At BBVA, Higuera Cabañes leads the responsible AI initiative. Her work focuses on translating abstract principles like fairness and privacy into concrete product outcomes. She noted that BBVA has had machine learning models deployed for roughly a decade, and the arrival of generative AI has shifted the conversation. The problem is no longer about choosing the right model. It is about reliability, trust, governance, and how organizations actually adopt these tools.
Her central concern is whether AI is the right solution for a given problem. She pointed out that responsible AI principles, when operationalized, can prevent real product failures. Fairness relates to discrimination, privacy relates to data leakage. In a financial context, these are not abstract ideals. They affect whether a credit risk model treats people equitably or whether a customer-facing chatbot handles different demographics differently.
Higuera Cabañes also raised the question of whether product managers should be able to detect and control bias in large language models. Some models, she noted, treat younger people differently from older people. Whether that bias is acceptable depends on the product and its users, and that determination needs to happen before deployment, not after.
Generative Code and the Shift to Non-Deterministic Programming
Jeff Smith, whose company C Proof focuses on verified numerical computing and understanding what generated code actually produces, offered the most pointed critique of the current wave. He described the impact of generative AI on software engineering as an orders-of-magnitude improvement in velocity. Some teams are seeing tenfold gains. But that speed comes with a meaningful increase in defect rates.
Smith argued that the replace-versus-augment framing misses the point. Software engineering has always been about raising levels of abstraction. Developers used to write instructions for specific machines, then moved to code that runs across multiple machines, then relied on compilers and type systems to handle lower-level concerns. Each previous abstraction layer was largely deterministic. The current shift introduces non-determinism into the process, and that changes the nature of the quality problem entirely.
The result, Smith said, is an industry-wide reorganization. The profession is restructuring around the fact that large portions of what used to require human code authorship can now be reliably produced by machines. Developers need to focus on the parts of software engineering that are not machine solvable, much as they no longer write assembly code by hand or build their own compilers.
Open Source, Sovereignty, and the Problem With Proprietary Models
The conversation turned to the growing tension between proprietary models and open-source alternatives. Arik expressed frustration with opaque changes in proprietary systems, referencing recent developments around Claude that she described as troubling. The criticism centers on a practice sometimes called nerfing, where model behavior changes without clear communication, making it difficult for enterprises to rely on consistent outputs.
This is driving momentum toward open-source models, which offer better control over the software supply chain, improved cost efficiency, and what the panel described as sovereignty over the systems enterprises depend on. The ability to run models locally, inspect their behavior, and maintain governance without depending on a single external provider has become a practical concern, not just a philosophical one.
What Comes After Transformers
Beyond the current wave of generative AI, Smith pointed to a possible next phase. The transcript cut off before he could fully explain his view, but the key takeaway from the episode's summary points toward more deterministic, principled approaches to AI. Approaches like Kolmogorov-Arnold Networks are being explored as alternatives to current transformer architectures, with the goal of improving inference profiles and bringing more mathematical rigor to the field.
For enterprises and developers navigating this landscape, the message from the panel was consistent. The tools are real, the benefits are measurable, and the risks are manageable but require deliberate attention. Companies that treat AI adoption as a systems problem rather than a feature to bolt on will be the ones that realize its full potential.