A mapping that questions the future of work
A visualization recently published by Andrej Karpathy has sparked heated debate in technology, business, and academic circles.
The former founding researcher at OpenAI and ex-Director of Artificial Intelligence at Tesla has developed an interactive tool to explore 342 American occupations using data from the Bureau of Labor Statistics. The mapping covers nearly 143 million jobs and assigns each profession a level of exposure to artificial intelligence.
The conclusion appears counterintuitive. The most exposed occupations are not necessarily the least skilled.
Developers, financial analysts, lawyers, consultants, accountants, and marketing specialists are among the most affected professions. Does this mapping herald a wave of disappearing intellectual jobs? Or does it reveal a much deeper transformation in value creation within companies?
To answer this question, we must cross-reference three levels of analysis:
- Andrej Karpathy’s exploratory work;
- recent academic research on occupational exposure to AI;
- field observations from the 100 entrepreneurs interviewed as part of the EntrepreneurIA project.
Who is Andrej Karpathy?
Andrej Karpathy belongs to the generation of researchers who contributed to the rise of modern deep learning. Trained at Stanford under the supervision of Fei-Fei Li, he specialized in computer vision and deep neural networks before joining OpenAI’s early work. He later led artificial intelligence teams at Tesla, particularly on perception systems used in the Autopilot program.
Today, Karpathy is also recognized for his educational work on large language models and generative artificial intelligence. However, his mapping of the labor market should be interpreted with caution.
The GitHub repository associated with the project explicitly states that this is not an economic report or academic publication, but an exploratory tool designed to visualize U.S. labor market data.
What the mapping actually measures
The exposure score proposed by Karpathy is based on simple logic.
The more an activity can be performed entirely through a computer using digital information, the higher its exposure to AI.
Conversely, jobs requiring physical presence, complex human interaction, or intervention in unpredictable material environments generally have lower exposure.
This approach has the merit of being readable. But it doesn’t measure risk of disappearance. It measures potential for transformation.
This distinction is essential.
What academic research actually says
Since the arrival of large language models, several research teams have been trying to measure their potential impact on work.
The study that has become a reference, GPTs are GPTs, was published by researchers from OpenAI, the University of Pennsylvania, and the Princeton Laboratory for Artificial Intelligence.
The authors estimate that approximately 80% of the American workforce could see at least 10% of their tasks affected by large language models, without necessarily concluding automatic job disappearance. However, the authors urge caution. Their work constitutes neither a forecast of job losses nor a scenario of worker replacement. It aims to measure the extent to which certain tasks could be transformed by large language models in the coming years.
This nuance is fundamental. The most recent work converges toward an analysis by tasks rather than by professions. Artificial intelligence generally doesn’t replace an entire job.
It modifies certain activities that make up that job. A lawyer continues to exist. But part of legal documentary research becomes automatable.
A consultant remains necessary. But information gathering and certain syntheses are now accelerated.
A developer retains a central role. But part of the coding becomes assisted.
The OECD’s work goes in the same direction. Its new framework for assessing occupational exposure to AI is based on comparing AI’s actual capabilities with the skills required by professions. The goal is precisely to avoid simplistic interpretations based solely on theoretical substitutability.
Skilled workers are paradoxically among the most exposed
One of the most surprising results of recent studies concerns the level of qualification. Unlike previous waves of automation, the most educated professions often appear among the most exposed.
OpenAI researchers notably observe that high-income professions frequently show higher levels of exposure to large language models. OECD analyses also show that highly digitized occupations can simultaneously experience high exposure and employment growth.
In other words, exposure doesn’t mean decline. It can also mean increased productivity and accelerated transformation.
What the 100 entrepreneurs interviewed in EntrepreneurIA reveal
The interviews conducted as part of the EntrepreneurIA project allow us to observe these transformations directly in the field. The first observation is striking.
No entrepreneur interviewed describes a massive disappearance of jobs today. All, however, mention a rapid transformation of tasks, skills, and decision-making methods.
Developers are changing roles
In technology companies, code generation assistants are now widely integrated into development processes. They accelerate the production of standardized code, automate certain repetitive tasks, and improve team productivity. Yet, demand for engineers remains strong. The most sought-after skills are gradually evolving toward software architecture, cybersecurity, data governance, AI model supervision, and complex system integration. Rather than a disappearance of the developer profession, entrepreneurs observe a shift in value. It is less and less located in writing the code itself and more and more in the ability to design, orchestrate, and secure increasingly sophisticated technological environments.
Consulting is transforming
Consulting professionals now use artificial intelligence daily to accelerate documentary research, produce syntheses, conduct preliminary analyses, or prepare certain deliverables. This automation saves considerable time on low-value tasks. Nevertheless, leaders continue to seek what models cannot provide alone: field experience, discernment, fine understanding of the company’s context, human support, and the ability to manage uncertainty. As information becomes more accessible, its relative value decreases. On the other hand, the ability to interpret it, contextualize it, and inform decision-making becomes an increasingly strategic advantage.
Marketing enters a new phase
Entrepreneurs specializing in marketing and communication probably describe the most visible transformations related to artificial intelligence. Content creation, monitoring, competitive analysis, message personalization, and audience segmentation are experiencing spectacular acceleration. However, a more subtle phenomenon is emerging. As the same tools become accessible to all market players, production capacity ceases to be a real differentiating factor. The competitive advantage then shifts toward strategy, vision, creativity, and the ability to design truly distinctive experiences.
Healthcare highlights the limits of automation
Healthcare professionals interviewed as part of EntrepreneurIA observe significant advances in artificial intelligence in documentary analysis, clinical research, diagnostic assistance, and medical imaging. Nevertheless, the therapeutic relationship remains at the heart of their practice. Listening to patients, trust, empathy, and support remain profoundly human dimensions that technology cannot fully reproduce. In this context, AI appears above all as a tool for augmenting healthcare professionals’ skills, rather than as a substitute for the relationship between caregiver and patient.
The five mistakes leaders make when interpreting AI’s impact
Mistake 1: believing AI replaces entire jobs
The observed transformations mainly concern specific tasks.
Mistake 2: launching tools before mapping processes
The issue isn’t the tool.
The issue is identifying activities that truly create value.
Mistake 3: considering AI as an IT project
The most profound transformations are organizational and managerial.
Mistake 4: underestimating human skills
Critical thinking, creativity, leadership, and emotional intelligence gain more value as AI progresses.
Mistake 5: thinking all jobs are exposed in the same way
Exposure depends on actual tasks performed and not solely on job title.
The Yunova AI maturity matrix
Analysis of testimonies collected in EntrepreneurIA identifies four maturity levels.
| Level | Dominant characteristic | Main risk |
| Observation | Monitoring and experimentation | Underestimating the speed of change |
| Assistance | Individual AI usage | Shadow AI |
| Augmentation | Integration into business processes | Insufficient governance |
| Transformation | Strategic organizational overhaul | Excessive dependence on models |
The most advanced companies are no longer just trying to automate certain tasks. They are gradually rethinking their organization around new complementarities between human intelligence and artificial intelligence.
An important limitation: exposure measures remain debated
A new generation of researchers is now calling for greater caution.
Several recent works criticize exposure measures based solely on assessments produced by language models. According to these researchers, exposure scores should be regularly recalibrated based on actual observations of AI adoption in companies and not solely on theoretical predictions.
This criticism is important. It reminds us that AI’s real impact depends as much on economic uses, regulatory constraints, organizational choices, and human factors as on technical capabilities themselves.
The new economic scarcity
For decades, scarcity was based on access to information. Today, information is becoming abundant. Artificial intelligence significantly reduces the production cost of much knowledge. Scarcity is shifting.
It now concentrates on profoundly human qualities: judgment, creativity, emotional intelligence, sense of responsibility, ability to make decisions under uncertainty, and inspire trust. This evolution concerns both SMEs and large organizations.
Conclusion: Karpathy’s real lesson
Andrej Karpathy’s mapping doesn’t constitute a prophecy about disappearing jobs. It acts more as a revealer. Recent academic research converges on the same idea: artificial intelligence transforms tasks first before transforming professions.
The testimonies collected in EntrepreneurIA reinforce this observation. The question is no longer which jobs will disappear. The real question is understanding where human value remains when cognitive execution gradually becomes abundant.
The companies that succeed probably won’t be those that use the most artificial intelligence. They will be those that best understand how to articulate human intelligence, governance, business expertise, and artificial intelligence.
This is probably the main strategic lesson that leaders can draw today from Andrej Karpathy’s work.
References
US Job Market Visualizer by Andrej Karpathy
Jobs’ AI Exposure Should Be Measured from Evidence, Not Model Priors
