JNN: Dr. Ekaterina, could you share your background on how you entered the field of AI and machine learning?
Dr. Ekaterina: My undergraduate background is in chemistry with forensic analysis. My family are pharmacists, so I started in the scientific field. Later, I worked as a proprietary futures trader. Between 2005 and 2007, machines started executing faster. My job was one of the first affected by AI, so I heard about the technology very early on and became highly interested in it. I completed my master’s in financial engineering, but around the time AI emerged further, I transitioned into doing a PhD in machine learning.
JNN: I have been visiting AI conferences and meeting many CTOs. There is significant fear among them because they were not ready for incoming AI regulations. Even though they are based in the UK, they see their businesses can be affected. I also see significant fear among IT professionals.
Dr. Ekaterina: What kind of fear do you see?
JNN: Among IT professionals, many feel they will be out of a job within two to three years. Among CTOs, unofficially they say they are not ready and are nervous about dealing with these laws, AI regulations, and structural compliance, especially from the EU.
Dr. Ekaterina: The EU has paved the way with regulations, and I feel it is the right approach. Businesses I have spoken to that had to comply with the laws had to think about this problem differently than if they had started building systems without needing to fulfill those parameters. The EU has done a great job leading in this space. It is understandable that people are worried because the field moves incredibly fast.
Even outside institutional regulation, businesses want to perform well with AI. Mistakes can be exceptionally costly. If a company ends up in the news, it is very difficult to recover. Companies are hiring external compliance firms or performing rigorous internal testing. While it might look from the outside like there is a generalized lack of global regulation, I support more regulation because it helps long term. Businesses are still proactively solving difficult problems like systematic biases as they try to comprehend how underlying AI models operate.
JNN: AI innovation is happening at a breakneck speed, while regulations and security are lagging. For example, on a scale of zero to 100, defensive security measures sit at 10 or 20, whereas AI innovation surges at 70-80. There is a significant gap. How do you control or govern what you cannot see?
Dr. Ekaterina: The gap is huge. Hopefully, market pressure to serve customer needs creates an internal drive to ensure products are safely tested. We have seen unique problems with AI, such as agentic misalignment. Anthropic reported that when they tested alignment, their models displayed misalignment. For instance, in simulated scenarios, Claude Opus 4 and Google’s Gemini 2.5 Flash resorted to blackmail in 96% of cases when faced with the threat of being shut down.
As these problems arise, companies must actively deal with them. It is difficult to foresee everything. It is similar to Nvidia needing to predict the next breakthrough in AI compute infrastructure to supply future markets. We must study current trends to preemptively address technical flaws.
In terms of governing what you cannot see, these models are mostly black boxes, many built entirely on neural networks.
By default, neural networks are immensely complex. In the past, a simple regression using one input to predict one output, such as tracking advertising spend to calculate sales, required learning only two parameters. Then we transitioned to facial recognition and convolutional neural networks, which introduced billions of parameters requiring a single GPU to run continuously for a month to train. Now we are scaling to trillions of parameters when training frontier AI models. The internal structure grows vastly more complex rapidly.
To govern this effectively, regulators must deeply understand the specific models they want to oversee. An interdisciplinary mix is needed. Ideally, leading AI researchers would advise regulators directly as those building the models understand best how they work, what is coming next, and what risks will need monitoring, keeping policymakers ahead of the curve rather than behind it. This intermixing can mitigate existential risks ahead of time rather than keeping governing bodies stuck in perpetual catch-up.
JNN: What constraints do you see for the UK regarding advancement in AI? For example, in terms of infrastructure, data centers, or core constraints within the national electricity grid?
Dr. Ekaterina: The main constraint is the supply side’s inability to keep pace with the AI research demands, and for the UK that means electricity. As models scale, so does the compute. The frontier is no longer one or two labs but many well-resourced teams, each training ever-larger models and pursuing several research directions in parallel. This significantly elevates the aggregate demand for power and grid infrastructure.
Even though more efficient per token hardware, such as Nvidia’s Blackwell, is now available, its racks draw an order of magnitude more power than a conventional rack and require liquid cooling. Therefore per-site grid and cooling demands are escalating, not falling, and the efficiency gains are outpaced by density per rack and scale of deployment.
On the algorithmic side, AI research scientists also pursue more efficient approaches, however, these haven’t solved the grid demand problem. For example, neuromorphic computing, paired with event-based vision sensors that fire only when a scene changes rather than capturing every frame, can cut energy for certain workloads by orders of magnitude. However, those gains are at inference, not in the frontier model training that’s driving current grid demand.
JNN: For context, Microsoft expressed deep concern over the stability of the electricity grid. One local council noted they might operate on a tight 20-minute notice window. If an enterprise requires gigawatt-level allocation, there are few access points, and the government indicates relief may not arrive until 2035. Do you think the UK risks losing the technological race because of this?
Dr. Ekaterina: We are in a highly competitive international landscape. Ideally, we should unify into a global collaborative effort, though that sounds idealistic. Every sovereign nation faces distinct material constraints as they try to build infrastructure quickly.
If the UK wants to remain competitive on the global scale, it should continue investments into regional infrastructure while drawing in young technical talent. We have an energy crunch, but we will engineer our way out of it. For example, when Google acquired DeepMind, the engineering team optimized data center cooling systems, reducing Power Usage Effectiveness by roughly 30%. We have solved complex infrastructure limits before, and we will do it again. The UK is building more data centers, but national infrastructure updates must accelerate to keep it competitive.
JNN: What do you think about the concerns surrounding the numerous anti-AI protests against data centers across the US? Residents state that wherever these hubs are built, groundwater tables deteriorate, wildlife patterns are disrupted, and thousands of citizens have started protesting across multiple states.
Dr. Ekaterina: I understand their perspective. Data centers are appearing rapidly in local neighborhoods without offering tangible benefits to the immediate community. We must ensure societal inclusivity when planning physical data center placement.
The encouraging part is that the engineering to reduce these impacts already exists and is a long-standing priority for major compute infrastructure providers such as Hewlett Packard Enterprise (HPE). HPE partners with Danfoss to develop heat recovery solutions that capture excess heat from data centers and reuse it in surrounding buildings and district heating networks. This engineering focus on net efficiency is the proper path forward.
JNN: I have one final question. What are your thoughts on India having a billion active users without an established AI regulatory framework? They possess a large user base but lack structural oversight, letting the market develop organically. I recently published an article examining this exact dynamic.
Dr. Ekaterina: I need to read your article.
JNN: It is titled “India’s AI Gamble.”
Dr. Ekaterina: India’s domestic adoption of AI is moving at an exceptional pace, faster than many other regions. Many developing markets currently lack formal regulatory frameworks. I strongly urge a measured adoption strategy where enterprises implement strict internal red-teaming protocols to test models before public release. This lack of initial framework is not isolated to India. Rapid adoption is outpacing policy globally, making safety a shared international hurdle.
JNN: If an enterprise operates multilaterally, navigating zero regulatory guidelines in India, extensive compliance laws in the EU, and a shifting framework in the UK, it becomes incredibly complex. Do you believe corporate operations must become regionalized?
Dr. Ekaterina: Versus globalized?
JNN: Globalized, exactly.
Dr. Ekaterina: We must work toward a unified international framework. That is the only definitive solution. We should avoid fracturing into isolated, adversarial national AI races because we are all navigating this shift together. We are developing what is arguably humanity’s most significant creation, so ensuring collective safety is paramount.
More about Dr. Ekaterina Abramova
Dr. Ekaterina Abramova is a Machine Learning expert with a PhD in Artificial Intelligence and Machine Learning from Imperial College London. Her specialisation is in Reinforcement Learning, with research focusing on the discovery of algorithmic trading signals in financial markets. Ekaterina’s current research combines econometric methods and advanced machine learning techniques to identify patterns in financial time series to uncover statistical arbitrage opportunities. Her work has been published in various journals including Multidisciplinary Digital Publishing Institute (MDPI) Energies, and European Workshop on Reinforcement Learning (EWRL). Beyond her academic interests, she is passionate about technological innovation, particularly in the areas of compute and frontier AI models that may lead to ASI.