Photo by the author
# Entry
Are we all in a race to the bottom of our own making? Data scientists have been employed for years in developing enormous language models (LLM).
Currently, the number of open data items seems to be decreasing day by day. Of those advertised, most seem pretty terrible.
By terrible, I do not mean too low salaries or excessive technical expectations from candidates. No, I mean these vague phrases: “Comfortable with AI productivity tools,” “Ability to ship large amounts of code,” or “Additional rapid design skills.” Translation: The chatbot is your main coding partner, there will be no mentoring, no standards, just code churning.
The chatbot, our own creation, now limits us to simply copying and pasting its results. It doesn’t seem like a very meaningful or satisfying job.
Is it still possible to find a meaningful job in such an environment?
# What is Vibe encoding?
Andrzej Karpatysome OpenAI co-founder, coined the term “vibration coding”. This means you don’t code at all.
What you do: You drink a matcha latte, vibrate, give commands to the coding chatbot, and copy and paste its code into the code editor.
What the chatbot does: Codes, checks for errors and debugs the code.
What you don’t do: You don’t code, you don’t error check, and you don’t debug the code.
How does such work feel? Like full-time brain rot.
What did you expect? You’ve handed over all the engaging, imaginative, and problem-solving aspects of your job to the chatbot.
# Vibe coding has devalued coding
“It’s not as bad as weekend projects go, but it’s still pretty fun,” Andrej Karpathy said about vibration coding.
Still, companies you can trust—those that don’t think of their products as “throwaway projects”—have decided it’s still a good idea to start practicing vibration coding.
AI coding tools emerged and data scientists were thrown out. For those who remain, their main activity is talking to the chatbot.
Work gets done faster than ever. You meet deadlines that were previously impossible. The ability to pretend to be productive has reached a whole up-to-date level.
Result? Half-finished prototypes. Code that breaks in production. Data scientists who don’t know why the code isn’t working. Hell, they don’t even know why this code Is working.
Prediction: Professionals who can actually code will soon come back into fashion. After all, someone has to rewrite this code written “so quickly” by the chatbot. Talk about efficiency. Well, there is nothing more competent.
But how to survive until then?
# How to find a meaningful job now?
The principle is very basic: do the work that the chatbot cannot do. Here’s a comparison of what AI can’t do and what you can do.
Of course, doing all this requires some skill.
# Required skills
Finding meaningful work in the age of vibration coding requires these skills.
// 1. Writing technical specifications
Most of the applications you will deal with contain incomplete and ambiguous information. If you can turn this information into a precise technical specification, you will be praised for preventing conflicting assumptions and expectations during development. Technical specifications aid align all teams involved in the project.
Here’s what this skill covers.
Resources:
// 2. Understanding data flow
Systems don’t just fail because of bad code. They are probably more likely to fail due to incorrect assumptions about the data.
Regardless of vibrational coding, someone still needs to understand how the data is generated, modified and used.
Resources:
// 3. Production debugging
LLMs cannot debug in a production environment. This is where you come in, with the knowledge to interpret logs and metrics to diagnose the root causes of production incidents.
Resources:
// 4. Architectural reasoning
Without understanding their architecture, systems will be designed to work in a production environment (fingers crossed!) but often fail in real-world traffic.
Architectural reasoning determines system reliability, latency, throughput, and operational complexity.
Resources:
// 5. Draft scheme and contract
Poorly designed diagrams and definitions of how systems communicate can have a domino effect: cascading failures leading to excessive migrations, which in turn leads to coordination friction between teams.
Create a good design and you will ensure stability and prevent failures.
Resources:
// 6. Operational awareness
Systems always behave differently in production environments than in development.
Since the whole idea is to make the system work, you need to understand how components degrade, how failures occur, and what and where bottlenecks occur. This knowledge will make the transition from development to production less painful.
Resources:
// 7. Negotiating requirements
The principle “prevention is better than cure” also applies in this case. You can expect almost endless downtime and re-writing if the requirements were initially poorly defined. Trying to patch the system once it goes into production is hell.
To prevent this, you need to intervene skillfully in the early stages of development to adjust the scope, communicate technical constraints, and translate unclear requirements into technically feasible ones.
Resources:
// 8. Review of the Code of Conduct
You should be able to read code not only for its functionality, but more broadly for its impact on the system.
This way, you’ll be able to identify risks that don’t show up in linting or testing, especially with AI-generated patches, and prevent subtle errors that could otherwise ruin your production.
Resources:
// 9. Cost and performance evaluation
Your work has financial and operational consequences. You’ll be more valued if you show you understand them by taking into account computer usage, latency, bandwidth, and infrastructure bills in your work.
This is much more valued by companies than building costly systems that also don’t work.
Resources:
# Actual jobs that still seem to make sense
Finally, let’s talk about real-world jobs that still require the apply of at least some or all of the skills we discussed earlier. Attention may be diverted from the coding itself, but some aspects of the work may still seem significant.
// 1. Data analyst (real, not just on a notebook)
Artificial intelligence can generate code, but data scientists provide structure, reasoning, and domain understanding for unclear and often poorly formulated problems.
// 2. Machine learning engineer
AI can train a model, but what about data preparation, training pipelines, supporting infrastructure, monitoring, failure handling, etc.? This is the job of a machine learning engineer.
// 3. Analytics Engineer
AI can write SQL queries, but analytics engineers are the ones who guarantee correctness and long-term stability.
// 4. Data engineer
Data engineers are responsible for the reliability and availability of data. AI can transform data, but it cannot manage system behavior, upstream changes, or long-term data reliability.
// 5. Machine Learning Operations Engineer/Data Engineer
These roles ensure reliable pipeline operation and model accuracy.
You can apply AI to suggest fixes, but performance, system interactions, and production failures still require human oversight.
// 6. Scientist (Applied Machine Learning/Artificial Intelligence)
AI can’t really offer anything up-to-date, especially up-to-date modeling approaches and algorithms; it can simply remake what already exists.
For anything else you need expert knowledge.
// 7. Data Product Manager
This job description involves defining what data or machine learning-based products should do, which includes translating business needs into clear technical requirements and aligning the priorities of various stakeholders.
You can’t apply AI to negotiate scope or assess risk.
// 8. Governance, compliance and data quality roles
AI cannot ensure that data practices meet legal, ethical and reliability standards. Someone needs to define the policies and enforce them – that’s what governance, compliance and data quality roles are for.
// 9. Data visualization/decision science roles
Data must be associated with decisions to have any purpose. Artificial intelligence can generate charts as much as it wants, but it doesn’t know what is relevant to the decision it makes.
// 10. Senior Data Officer Roles (Director, Staff, Leader)
Artificial intelligence is a great assistant, but a terrible leader. More precisely, he cannot drive.
Decision making? Cross-domain leadership? Leading technical direction? Only humans can do this.
# Application
Finding meaningful work in the age of vibration coding is not uncomplicated. However, coding isn’t the only thing data scientists do. Try looking for job postings that, even though they require vibration coding, also require some skills that AI still can’t replace.
Nate Rosidi is a data scientist and product strategist. He is also an adjunct professor of analytics and the founder of StrataScratch, a platform that helps data scientists prepare for job interviews using real interview questions from top companies. Nate writes about the latest career trends, gives interview advice, shares data science projects, and discusses all things SQL.