Ep10: May 25 – Automation Blackout?

• Intel is in restructuring, and cost cuts (20% staff to go), delayering, and accelerating on focusing on business units relevant for the future.
• TSMC has the journey to accelerate production in the US, addressing the AI chipset restrictions from the US to China and the tariffs on Taiwan. Anyway, the challenge to produce in the US, especially the 2nm, with restrictions on China’s export and the Taiwan government trying to keep the most advanced technology limited to its own country, could be a challenge to TSMC that is already raising, for example, on the limits of control of the AI chipset export to China.
• Huawei, meanwhile, aims to produce an AI chipset in China, competing with Nvidia. In general, China is working to produce AI chipsets locally and gain independence, and there is high tracking from the US to look for their progress, even monitoring via satellite for the development of production lines in this area.
• Nvidia getting squeezed between US regulations on export of H20 chipsets for AI to China and China’s development needs. On top of the base of the chain, there is still TSMC in many cases.
• So What?: Collaboration between Intel and TSMC could become a reality, and I predicted that here, where I supposed a piece of Intel would be sold to TSMC.
• My Thoughts: China has already proven in the past to be able to step up its own supremacy. The fact that there is a gap in AI but a company like Huawei is strongly on it, and there is also the capability to build the underlying semiconductor part, could make the delay from the US to China on the run for AI supremacy, maybe not that long. At the end is not a run for the next 12 months, but is going to continue for the next many years to come, and distances could shrink during such time.

I can only mention this month to be careful in this market and where to invest. I really prefer to avoid continuing speaking about it, especially as I see speculations ongoing on meme coins and people still brutally landing on them as a way to invest. Stay curious but also vigilant on this.

• The acceleration in the run for supremacy in energy production is quite big. China recently ranked highest on nuclear power production rate. Many local mini-nuclear productions are ramping up, reducing dependency on big grids. Also, nuclear fusion (being the future target for clean energy) is also a big toss where Europe looks a little bit behind, apart from France, against China and the US. Just to take into consideration that also in the area of nuclear fusion, there are a lot of semiconductors needed that could make the chipset war even more aggressive, as the AI, chipsets, and energy are tightly connected and depend on each other while increasing demand.
• On the path to balance production and consumption of energy, an interesting decision came from Kuwait that has forbidden mining of crypto recently as cause of blackout from the government perspective.
• Recently, we also saw Spain/Portugal struggle with blackouts caused, apparently, by remodulation problems from the renewable energy not balancing as expected in the grid. From what we can read so far, the increase of renewable energy (that happened anyway over the years) has caused a different way to react in case of peaks (positive or negative) than in the past, which has not been properly captured by the expert system, basically an AI system working in automation. Here we are in the phase of assumptions, and there is yet not a formal root cause summary, and the EU just started the investigation formally. Hopefully, this will be proof to be NOT cyberattack-driven, as that would be even more fragile from a perspective of impact.
• My Thoughts: Here, I would like to give some level of predictions and speculations on the general AI. A grid is a complex engine, with a lot of components, sensors, automations, and acts as a complex expert system with several levels of robustness and fault tolerance, and working within specific tolerances to automatically decide. In this specific case, the down happened after 5 seconds from the start of the problem. If the assumptions we read are right, it could look like that environment over time changed, and the tolerances didn’t adjust. It’s a super over-simplification to make my reflection more understandable. Now, if an unexpected change in the environment caused such a problem in such a complex but monitored, maintained, and powerful automated engine, we need to take into consideration that automation in day-to-day jobs in enterprises also has some level of risk if not properly designed and maintained. Autonomous engines, like Agentic, have possibilities to adjust automatically in unpredictable situations but will also have the possibility to act in unpredictable ways when outside safeguards, so basically boundaries. So will be relevant together with the data training, the reasoning, and the maintenance to keep such environments under proper monitoring. I give these examples to mention that in some rare cases, behavior can go against the plans, and cyber can also force behavior-changing perception of the environment with false information. I analyze more in the section Job Evolution on the transformation around Agentic for enterprise job activities that are clearly less complex and business critical than such an infrastructure of a grid. Here are my predictions that the blackout will become more real once we are clear on the root cause. As I said, I hope it was not a cyber event that manipulated the environment, making the AI engine fail.

In this context, we saw the US starting to dig deep into the oceans, and such action is considered high environment impacting. Even China was against it. From a technology point of view, that remains my focus, the competition is again for semiconductor minerals as a reason to act to deliver for the different areas of need.

One of the major changes coming with the Agentic is the autonomous capability to understand, remember, make decisions, and evolve. There is a big dependency on the interpretation of the request, the data source for training, restrictions, safeguards, and the generation of the output. A special evolution in the area of LLM, which is also influencing Agentic, is the reasoning, that is, the capability of making complex correlations and also describing the steps executed in the chain of decisions. Recently, Anthropic made an analysis of the fact that the reasoning models sometimes hide some of the steps they used to make a decision. Who follows me for some months know that I often report the challenges around governance of Agentic, especially if we start to have them combined in a swarm approach. If we cannot derive the sources (due to the distilling approach on LLM), we are definitely limited in our understanding of how an agentic took a decision.

My Thoughts: Agentic is used in core processes, making decisions that cannot be fully audited because part of the reasoning process may not be visible. On top of deriving the source of data, training could be impossible. As a consequence, agentic could be considered not reliable and not auditable. I recommend applying for non-critical agentic at this stage and working heavily on the quality of the training data and the boundary to limit where to look for decision paths. On top of feedback to AI producers will be relevant to drive toward a direction of completeness.

Microsoft released an interesting article on the future firms marked as Frontier. In such article, based on big amount of linkedin data (they own) and office 365 data of millions of users, they can see the influence that AI will have in agentic mode in many activities based on the actual level of repetitive tasks, the change of demands of skills required over internet (Linkedin) and the general way to operate for people in the offices. They also analyzed some aspects of augmented people with AI. A key aspect of the logic is the shift of the job (and I will speak more in the Job Evolution section in this newsletter) from “I send emails” or “I write documents” to “I create and manage agents“. I will speak more in the section as I mentioned. What I reflect here is the importance that such a transformation will be properly regulated because there is an intrinsic difference between sending emails (where we own the action) and building an agent that is autonomously acting and could execute things not in full alignment with the creator.

The direction toward “off the shelf” AI will be progressively more relevant, and that will drive the build of the core components.

Last month, we also saw a progressive interest in creating action figures by many people, and social media was filled with them. At this stage, there are hidden problems of privacy risk, and people often underestimate the way such information could be used later. I remind you here because in the last few months, I mentioned in my newsletter how many cases we had of data not formally authorized but used to train AI. This is specific, for example, for many newspapers, authors, and artists who started a fight against AI LLM engines using copyrighted data for their training. The regulations around AI are still quite weak, and although the EU made an effort, maybe quite formal, many other countries are completely out of such agreement and are not respecting some basic aspects yet.

On this topic, there is a proximity between tariff wars and robots for automating more and more production and reaching even higher cost-effectiveness. In this sense, China is progressing. I assume that if such a process were fast enough, it would make it easier for the US to bring back production in the US without the need for expert workforces, as that could be fully automated and potentially autonomous with the AI capabilities out there, meanwhile. Tim Cook, some time ago, mentioned the level of specialization needed for bringing together the iPhones, and that was why he would still use local production in China (cost-effectiveness and level of precision). In the years he started to build also in India, and recently shifted China production for the US to India, to come toward the US administration that wants to limit China’s dependencies. Definitely, I agree that robots can mitigate the difference in cost of workforce to operate because the cost would be similar in any country where today people have a different cost in different places of the world. Such robots would make offshore productions less convenient. It’s not clear how fast this would really be possible, and how the gaps in jobs (where robots would replace people) would be absorbed with new jobs, and how fast. There is, in general, a multifaceted capability in humans that is far from one task robots but also by one task AI, even if those tasks can be chained.

I recap what I started to mention in the AI section. Microsoft Work Index Report mentions the evolution toward Frontier enterprises, where the need for distinct roles in the organization is reducing, thanks to AI. In such a model, humans with AI become more capable of executing actions in different areas, and there is a transformation from “send an email” to “create and manage an agent” in the way of operating several tasks. This paradigm shift is justified by the fact that each human works with augmented AI in a sort of modern RPA, low-code or better no-code way.

Now, another article linked to that is the Harvard study on Cybernetic Teammate. This study, quite interesting even if limited in the number of examples and length of the test, is showing that a human with AI is beating humans without AI and even teams (of two people) without AI.

In the study, the team is based on one commercial and one R&D, and if one human of the two types without AI would make decisions quite influenced by his/her own skills, one human + AI can make decisions similar to a team composed of the two different competencies and even beating them (if the team is without AI). So AI is leveling competencies and reducing gaps, allowing for making decisions not influenced by the specific competencies of the human, but working more across functions.

My 2 cents on the study is about the fact that it has been quite short in time and only on 2-member teams, so the correlation and mutual feeling that a team could build would be too limited in time to prove the augmented value from the team with or without AI versus the individual.

Definitely clear instead is that an individual without AI is definitely less performing than an individual with AI in such a test, and that is a key aspect why the upskilling is going to be fundamental in the coming future, as I mentioned many months ago, also referring to official reports like the one from the WEF 2025.

There are many challenges to understand. First of all, the reliability of AI answers, the training data, the governance and regulations that we already mentioned in the AI section, and finally, most relevant for me, the accountability. We can build an agent and make it autonomous to operate, replacing some actions we take, even the example of the send email or writing a document. However if we send a wrong email or write a document with wrong contents, we are accountable for the contents and if we create an agent doing on our behalf a reply that one day reacts unexpected, we are still responsible for the contents we send out and that is based on the engine selected by the enterprise we work in, the proper safeguards and training, the reasoning (as well) and all the possible unpredictability coming from an autonomous agent. On top, we need to take care that an agent could answer emails on our behalf on the basis of knowledge, but could be manipulated by who is requesting information to get around the safeguards. That is making it even more relevant to make such an agency progressively properly maintained.

The other challenge I see is about the fact that an AI leveling the contents and knowledge is reducing the difference and specialization between the departments. In this sense, the Microsoft article speaks about a new blueprint for organizations. That has to be clearly foreseen in advance by companies when they design the future way to operate, as specialization is anyway something bringing a variety of perspectives, and the AI needs to be seen to complete the human perspective, not to make it irrelevant.

Another aspect related to the same topic is about the way the new automations will transform in terms of resources. There is blur area in the way humans act that is combining a chain of sequential skills, and that is not typically immediate to be AI-replaced. No one is doing just a task, but it is often a mix of some repetitive tasks (automation capable), some reasoning, and some other action as a result of their own correlations. If we take only some specific skills and we automate or make autonomous, we can believe in the 1 hour saving per day that could be achieved, for example, by repetitive tasks we do regularly, and that level of efficiency needs to be balanced by more productivity (more volume or higher quality or both). At the same time, the effort to maintain our agentic reliability requires some level of evaluation, especially in the first phase, of their reasoning, like we are following a junior employee in some basic tasks in the daily job, then we leave him/her more free, but from time to time we check if everything is right.

Key aspects going forward are to describe key processes where scaling capabilities are needed, and be able to design which parts will be more relevant to be automated to accelerate the scaling. Clearly, enterprises of smaller size will have more agility to transform, but the impact on big enterprises will be highly beneficial as more structured and formal processes are in place, and they can better estimate the improvement from automation.

One final aspect that I consider relevant in the workforce upskilling that will have to happen in the near and medium future is about the junior and, in general, Gen Z employees. As many of them will enter enterprises with already AI integrated in the way to operate, they will be requested to make more complex decisions from the early stage, but will miss part of the initial experience that would help them to mature on taking decisions. I believe the job of the higher management will have to include the effort to train such a workforce to balance their missed experience, and will be an opportunity to create long-lasting relationships as part of a generation development that each enterprise will have to execute.