Future in Drafts

Introduction

AI race continues at full speed between the U.S. and China with the ongoing advancements but with different strategies. A recent order from China was about to eliminate disorder on AI development. Efficiency perspective this sounds clear as developing and training models are the output of spending millions of dollars for a foundation model. On the other hand, innovation does not like hard rules and regulations. The U.S. has released an action plan and set the direction on acceleration of innovation and building American AI Infrastructure for leadership. The direction of countries will converge or differentiate, and we will learn this not later than 2028.

Order or Chaos in AI Innovation – Why Planning and Open Weight Models Matter

I will start with the core idea; innovation does not like hard rules and regulations. China has the materials and goods production knowledge at scale, and this has been achieved by mass production of specific items at specialized factories. A success story about this is happening with electric car parts being produced in scale and same parts used by different Chinese EV brands. Could the same if applied to AI developments produce similar efficiency? I don’t think it will work in the field of Data Science as the field is way different from physical production methodologies.

The path to innovation which comes with chaos will cause higher spending and less efficiency but if the target is to leap forward in Artificial Intelligence at earliest by taking global leadership, this method will have high chance of success.

China’s bold move for AI leadership came from the release of open weight models which pushed OpenAI to join the wave with its open weight models “gpt-oss-120b” and “gpt-oss-20b”. A crucial aspect of innovation is the usage of models which can thrive with global community interest, consumption, support and developer engagement. This has been a sub-race we’ve witnessed recently, and China played its hand wisely.

Innovation, Hardware and Infrastructure – Three Pillars of Success

Innovation requires freedom and global collaboration, but without a strong power grid and hardware technology and supply chain, a country cannot lead in AI. If current trends continue, the total power used by data centers worldwide could more than double compared to today’s levels. Nuclear power is in the talks for power hungry AI systems which can also help companies to fulfil their climate goals but realization of them take years up to a decade (https://www.technologyreview.com/2025/05/20/1116339/ai-nuclear-power-energy-reactors/). A challenge emerges at this point, the U.S. has a weak electrical grid, which could pose a major challenge to its AI leadership.

The U.S. is dominating AI chip technology and market, Nvidia at the forefront. China is making significant progress in producing competitive chips with the companies Cambricon Technologies, Biren Technology, and Alibaba. China does not yet mass-produce chips at the 5nm–3nm scale. This gives the U.S. time to strengthen its power grid, as the gap on chip technology could eventually be closed.

Different Vision – Different Race

We know from the Leadership of China; AI developments focus on practical use cases, set the target for improving efficiency and facilitating faster market adoption and investment returns. On the west side, the U.S. White House released “Winning the Race: America’s AI Action Plan” in July 2025 which lays out a broad federal strategy.

Neither the U.S. nor China has publicly set explicit goals or policies specifically targeting AGI (Artificial General Intelligence) or ASI (Artificial Superintelligence), nor have they formally defined a roadmap for achieving global leadership in these domains. At the leadership and policymaking level, this approach is reasonable, given that “AGI” still needs a precise definition. The lack of consensus on AGI creates an opportunity for divergence, with countries likely to pursue different objectives. And even divergence could be on practical use cases versus research investment on Artificial Superintelligence with intellectual level beyond human intelligence.

Conclusion

The innovation comes with a price, and it does not like hard rules but freedom within chaos. Supervision on technology companies for AI development efficiency may degrade the innovation and leave a country out of the race.

China has been more active in releasing open-weight LLMs publicly, while U.S. efforts dominate in closed-weight high-performance models and infrastructure. China’s move on the release of open weight models pushed OpenAI to join the wave. Open Weight Models’ race should not be underestimated as it is a sub-race on developer engagement with a strong influence on AI leadership.

Even if the countries are likely to pursue different objectives (Practical applications versus reaching AGI), the destination of AI by different countries will meet in battlegrounds with full autonomy, restricted within zones. If AGI and followingly ASI could not be achieved in next 5 years, practical use cases would win in the short term and countries invested in this strategy would better be positioned and control the upcoming market for AGI and ASI. We will get hints on this not later than 2028 by understanding if AGI could be a reality and if LLMs will not stay as powerful pattern recognizers.

Introduction

The MIT NANDA research report titled ‘The GenAI Divide – State of AI in Business 2025’ revealed that most Generative Artificial Intelligence (GenAI) pilots are failing to progress to full deployment, with only 5% of projects achieving success. 95% of pilots are failing sounds like a major deviation from what we read about AI takeover in various mediums therefore the MIT NANDA (https://nanda.media.mit.edu) research has made a lot of noise lately. Research provides valuable insights for crossing the “GenAI Divide” to be one of the successful organizations in GenAI deployments. (Thanks to Aditya Challapally, Chris Pease, Ramesh Raskar, Pradyumna Chari for this insightful report.)

Agentic AI Features for Business to Succeed – Persistent Memory, Feedback and Refinement Loops

An interesting inference from the research is that personal use of GPT-like LLMs has increased in recent years meanwhile Business fails 95% in GenAI pilots. I see this gap exists because of repetitive tasks like document processing or consulting to an AI expert without any human supervision and judgement fuel the use of GenAI by workforce.

In order to Business to increase deployment of pilots, Agentic AI systems built with adaptive agents are needed which should promptly be integrated into business workflows and handle new situations and tackle errors (I touched on this with a previous post here, under the section “Ending the Era of Spreadsheets“). Adaptive autonomous agents can process workforce feedback through LLM-driven iterative refinement loops. Persistent memory is also a key ingredient to track long-term goals and contextual awareness for user productivity. Further, fine-tuning the model will help to shape it with the domain and organization specific feedback but most of the time Feedback and Refinement Loops would be sufficient for most cases.

Preliminary AI Skills for Everyone – Familiarity with AI tools and Prompt Engineering

Considering the research, most successful integrations came from front lines in contrast with centralized automation departments. This is a signal of job interview questions to ask about Preliminary AI Skills. Familiarity with AI tools is becoming workplace staples. Therefore, be ready for your interviews in advance to explain which toolset you can work with comfortably and better take a course on Prompt Engineering before that.

Referrals and Peer Trust to Bridge the “Divide”? – No. Value must be the essential decision factor for an AI Product

The interesting part of the research is that interview questions are answered by leaders those failed AI Transformation in their Business. In that case, I expect the survey results may not reflect all the right actions for the future. Based on that, the method of selection of partners via Referrals and Peer Trust is one of them in my opinion. If “Limited disruption” is one of the patterns that define the GenAI Divide, then referral opinion led product selection might not make your company a disruptive one.

Conclusion

Agentic AI systems will be the key player in the future of AI adoption by Business. Nowadays Businesses are making new plans to increase their AI project deployments where teams to be expected to carry responsibilities as the frontiers. By considering this, it is going to be inevitable to learn Preliminary AI Skills by everyone and even go beyond that.

It is best to evaluate the report’s outputs cautiously as onboarding decision of an AI Product must not heavily depend on peer trust and recommendation. The value should be the key point where the pilot project target objectives are fulfilled by observing robust workflow loops not broken by edge cases.

I highly recommend you read the research report with many learnings and insights you can find in a single reading (I added the pdf file as document to my LinkedIn post).

Introduction

We are living in a co-working model with AI systems which the weight is increasing day by day for the advantage of AI. Meanwhile expectations from AI systems are heightened by the groundbreaking generative capabilities of GPT-like LLMs and the projected arrival of Artificial Superintelligence by 2027 (https://ai-2027.com/summary). Accompanied by all this, humans are still at the wheel in the age of AI.

Fully Autonomous, Unsupervised Driving – Regulatory Hurdles and the Need for Faster Hardware

Except for limited pilot programs, today’s autonomous driving systems are supervised. This kind of journey needs human-on-the-loop, requiring driver oversight which human monitors and be ready to intervene. Next-generation hardware is on the horizon which is expected to be ten times more powerful to arrive by the end of 2026 (https://x.com/elonmusk/status/1803856461333725615). With the arrival of next-generation hardware, robust unsupervised autonomy could bring better safety and security which could help for regulatory sign off.

Ending the Era of Spreadsheets – Advancements in Verifying Generalization and Tool-augmented Chain of Thought

I value spreadsheet software’s role in modern business more than any other tool. Ending it could serve as the canary in the coal mine for the decline of human-driven data analytics. If an advancement ends spreadsheet software usage by human operators, unfortunately this is going to be a time of “dark offices”. Such development affects many roles and accelerates the adoption of AI-driven analytics.

If someone claims that business grade spreadsheet software tasks can be fully automated, the workflow must demonstrate how errors are detected and corrected. To detect errors, you need to go beyond “guessing”. Therefore, in a business workflow, “guessing” is not enough but with cycles of formulate, try, check, and fix any error in formulas/macros/charts inside spreadsheet software are needed. This capability requires advancements in Verifying Generalization and Tool-augmented Chain of Thought. If an AI model faces an unseen task and writes a formula, the verifier can check if it produces the expected result by enforcing correctness checks. In addition, instead of thinking in the context only with complex reasoning, the Agentic workflow can leverage external tools in the intermediate reasoning steps to make the output accurate and trustworthy.

When AI Thinks Alike, Humans Stand Out – The Power of Human Creativity

Prompt engineering improves customization, yet all outputs still rely on LLMs trained on similar data sources, from web crawling, Wikipedia and various informative materials either printed 100 years ago and digitalized or natively in digital form. Human creativity is still at the top of the mountain and will continue to make us special at least to the time of Artificial Superintelligence arrival between 2027 – 2030 when I need to revisit my thoughts. Until then, we keep the wheel of creativity.

Conclusion

While impressive technologically, AI and hardware developments are unsettling for humans, as many jobs could vanish. Considering that, we are still steering the wheel, at least for the next few years. Complex and fully autonomous, robust, general-purpose self-verification systems are actively unfolding but not yet complete. With respect to speed in advancements of AI systems and next-generation hardware, fully autonomous systems are not far away. Artificial Superintelligence, my expectation is between 2027 and 2030, may dramatically change the human-AI dynamic. Roles are subject to change: as we expect to automate everything with AI and make it our workforce, we may also become its workforce by fueling the hunger of its vast energy demands and following its further demands. But first, let’s wait for the end of spreadsheet software! 

Introduction

Data Sources and Analytical Tools

This data story relies on the dataset which is based on real-world data and covers Data Science, Artificial Intelligence, and Machine Learning job listings from 2020 to 2025 hosted on Kaggle. It has been curated and uploaded to Kaggle by Adil Shamim: https://www.kaggle.com/datasets/adilshamim8/salaries-for-data-science-jobs. Data sources are AIJobs salary survey (CC0 license), 365DataScience, Payscale, KDnuggets, ZipRecruiter, and others.

Tableau is used for all data visualizations with “filtering” and “calculation field” features. Each visualization is located into its own tab for the public view.

Chapter 1 (Ch.1): Countries with Job Opening(s) – Uneven distribution in the world

On the world map, there're 97 countries with at least 1 job opening. If we consider there are 195 countries in the world, half of the world countries are not even looking for a Data Science professional. We could observe the economically developed and emerging countries are in the loop. Meanwhile low-income countries or countries having conflict/war are mostly not having a single job opening.

United States is clear leader in the posting of AI-related jobs with over 135k job openings in between 2020 to 2025.

Chapter 2 (Ch.2): What are the most wanted roles? – Practitioners lead the way

Amongst 422 job titles, the ones with the practitioner skills are preferred more like Data Scientist, Data Analyst and Data Engineer. This highlights the need for Data Science, Artificial Intelligence, and Machine Learning talents who will design, build up, run and maintain data science projects with hands-on experience.

Chapter 3 (Ch.3): What is the trend of years in between 2020-2025? – Foundational roles are the most wanted, new roles emerged

In the Sankey diagram, some titles — such as Data Analyst and Data Scientist — are spread across multiple years; both are considered foundational roles in the field of data science. The rationale behind that could be explained as whenever the roles were available in prior years are continued by leading the number of job listings as we see in Chapter 2. New specific titles like “AI Engineer” started its appearance by 2022 and showed a significant spike from only 1 listing and reached to 1394 listings by 2025.

Conclusion

The United States is leading the world in AI-related job listings, while many developing nations have a scarcity of such opportunities, with some having none at all. Great Britain and Europe are also engaged but significantly behind in the "job listings race". This highlights a significant disparity between the leading nation and the rest of the world.

Foundational titles like “Data Scientist” and practitioner related titles are leading the jobs as data collection, preparation and designing models take most of the time in the field. This also indicates the interest of companies in which they would like to analyze customer feedback, run predictions, make forecasts or detect anomalies within broader data science practices encompassing methodologies from Machine Learning and Artificial Intelligence. As highlighted in Chapter 3, the exponential growth of emerging AI-specific job titles like "AI Engineer" proves that companies are highly interested in AI talent by 2025. In the near future, I do not expect this to change, instead it is likely to create even more new job opportunities together with new AI-related job titles.

As a result, the US is the leading market for AI professionals, offering a high volume of job opportunities. This is followed by Europe, where the AI job market is also expanding. While demand exists for both foundational and emerging roles, the latter is seeing exponential growth in job listings, indicating a shift in the market toward newer, specialized practitioner positions.

Sertaç Levent
Founder, Vubion.ai