Jon Krohn in Conversation: How LLMs Are Reshaping the Value of Data Scientists

Claude Opus 4.6 strips twenty years of technical skills of their moat, shifting value to context and business communication.
In a Super Data Science live conversation, Jon Krohn argued that models like Claude Opus 4.6 have eliminated the moat around traditional skills such as MATLAB, Python, and machine learning, shifting practitioner value toward understanding requirements, providing context, improving experience, and communicating consensus. Drawing on his path from neuroscience to AI, he stressed that fundamentals still enable going deeper, and that vibe coding suits rapid prototypes but makes later changes costly. The talk also covered AI orchestration and agentic AI as today's best investment, plus a tipping-point narrative of accelerating AI progress.
When Twenty Years of Technical Expertise Gets "Disintermediated" by a Single Model
In a recent live conversation hosted by Josh Starmer and Luis Serrano, Jon Krohn—host of the Super Data Science podcast and author of Deep Learning Illustrated—dropped an unsettling verdict: since Anthropic released Claude Opus 4.6, the technical skills he spent nearly twenty years building—MATLAB, command-line tools, R, Python, statistics, and machine learning—no longer constitute any kind of moat.

"LLMs can do it better, faster, and cheaper," Krohn said. "That's scary." He admitted that these technical skills used to be surrounded by a wide moat, and that wall is now being systematically torn down. But rather than dwelling in anxiety, he offered a clear direction for where value is migrating: practitioners' value is shifting toward things closer to the essence of business—understanding requirements, improving user experience, communicating ideas, building consensus, and handling the messy, complex real-world context of actual projects.
The New Core Skill: Providing Context for LLMs
Krohn described how his own way of working has changed: a substantial portion of his week is now spent thinking about "what context does this LLM need to successfully complete the task?" Which meeting notes are needed? Which resources should be pointed to? What information exists only in his head, or only in a conversation from a phone call?
This sounds like a technologist turning into a project manager, but Krohn sees exactly that as the value. His AI software consultancy Y Carrot (a pun on the "Y hat" symbol, with the domain ycarrot.com conveniently available) is a practice of this philosophy: the team all comes from the pre-LLM era, understands underlying principles and common pitfalls, and helps clients turn ideas into working prototypes from scratch, then push them into production.
He specifically called out the limits of "vibe coding": a client spends two hours on a weekend getting Claude to generate an app that looks viable, but the moment you need to modify or extend it, it often turns out to be more expensive and slower than rebuilding from scratch. Josh Starmer drew an analogy from high school math—plot the function first, then solve the equation, and you know how many intersection points to expect. Understanding the fundamentals isn't about writing every line of code yourself; it's about knowing what the output should look like and where errors might hide.

Context Engineering is a concept that emerged as LLM applications deepened. It refers to carefully designing, organizing, and injecting the background information, constraints, and reference materials a model needs to complete a task, in order to maximize output quality. Unlike prompt engineering, which focuses on wording tricks, context engineering is more concerned with information architecture: which documents, conversation records, domain knowledge, examples, and tool descriptions should enter the model's context window, in what order and format they should be presented, and how to avoid irrelevant information interfering. In agent systems, context engineering also includes managing long-term memory, retrieval strategies for retrieval-augmented generation (RAG), and state passing for tool calls. Krohn's description of "thinking about what context the LLM needs" is precisely the core of this practice.
Vibe coding was a term coined by Andrej Karpathy in early 2025, referring to a style of programming that relies heavily on describing intent in natural language and letting AI generate code, with developers reviewing it line by line less often. Its typical use case is rapid prototyping or one-off scripts: the user describes what they want, the model produces runnable code, and the developer judges by "feel" whether the result is reasonable. This dramatically lowers the barrier to entry, but the hidden risk is that the code's internal structure, dependencies, and edge cases are often not fully understood. When requirements change or the system needs to scale, developers lacking foundational understanding struggle to locate problems, and refactoring costs can far exceed rewriting. Krohn's comments on the limits of vibe coding are grounded in exactly this context.
Is It Still Worth Learning Data Science Now?
Faced with an audience question about whether one should still learn ML from scratch, Krohn admitted he "really doesn't know the answer," but he hopes the answer is yes. The reasoning aligns with Starmer's analogy: if anyone can produce a data science solution with vibe coding, differentiation comes from being able to go one level deeper—being able to say, "Hold on, this reminds me of something I've read or experienced, and there's more we can do here."
Luis Serrano added an often-overlooked angle: communicating results matters more than generating them, and explanations need to be tailored to people at different levels. Your boss, your boss's boss—each needs a different framing. Without understanding the fundamentals, you can't adapt flexibly beyond the script. Serrano also shared his own analogy: if you've run marathons for ten years and then marathons are cancelled and replaced by soccer, the 42 kilometers of mileage itself is useless, but the physical capacity and willpower you built will transfer.

Krohn cited analogies from law and medicine: we want lawyers and doctors to have AI assistance, to be able to consult a broader, less biased range of opinions, but we still want a human to bear responsibility and to contextualize explanations for our specific situation.
From Neuroscience to AI: An Atypical Path
Krohn's résumé is rather unique: an undergraduate degree in neuroscience from Wilfrid Laurier University, a PhD in neuroscience from Oxford, where his research focused on machine learning and computational statistics for brain imaging and genetic data. He emphasized that he didn't "switch careers" into AI—what he was doing during his PhD was already machine learning. This background deeply influenced his writing: the reason Deep Learning Illustrated presents AI concepts with extensive visuals is precisely that artificial neurons, neural network connectivity, and machine vision system architectures all drew inspiration from biological neuroscience.
Worth noting: all 200-plus illustrations in that book were hand-drawn by Belgian artist Agley Bassens—this was before text-to-image models like DALL·E existed. Krohn joked that an Amazon reader gave it a one-star review on the grounds of "where are the illustrations you promised," presumably expecting a comic book.
Two Narratives About the Future
In his keynote talks, Krohn repeatedly cites Matt Schumer's blog post "Something Big is Happening": in February 2020, everything in Europe and America was normal, and no one could imagine that a month later the whole world would shift to Zoom work and lockdowns. Schumer's argument is that AI is at the same kind of tipping point—Opus 4.6 has arrived, the massive change hasn't yet shown itself, but it is inevitable.
He cited Artificial Analysis data to illustrate the pace of change: Opus 4.6, the model that "disintermediated" all his technical skills, scored 32.5 on the composite intelligence index, while the current leading model has reached 58, and dozens of open-source models have already surpassed the former.
For the further future, Krohn cited a forecast timeline from Epoch AI researcher Jaime Sevilla: within months, a moment when agents massively replace office tasks; AI annual revenue reaching the trillions of dollars; millennium prize math problems solved by AI; around 2030, tens of millions of high-precision robots and job displacement becoming the biggest political issue; by 2035, AI reaching parity with human cognition in all domains, with economic growth potentially rising from today's 3% to 10% or higher. Krohn made clear this is not his own vision, but if even a tenth of it comes true, the world will be unrecognizable.
One moving interlude in the conversation: Krohn shared a music video generated with Opus 5.5—the AI autonomously chose Voyager 1 as its theme, wrote the lyrics, and used code to render every frame based on real NASA data, telling the story of a machine growing old, drifting farther and farther from us, where every message takes a day to arrive and the battery is running out. "I cried watching this machine-made video," he said.
Artificial Analysis is an independent analytics organization that scores mainstream large language models through standardized benchmarks and human evaluations. Its Intelligence Index aggregates performance across reasoning, knowledge, coding, math, and other dimensions, and is often used for cross-model comparisons. Krohn cited this data to illustrate the speed of model iteration: Opus 4.6, which made him feel his technical moat collapse, scored only 32.5 on the index, while the leading model at the time had reached 58, and dozens of open-source models had surpassed the former. This contrast reveals the rapidly shifting landscape of AI capability—today's "disruptor" is quickly overtaken by newer models, and practitioners face not a one-time shock but a continuously accelerating migration.
Advice for Young People—and for Himself
To confused students, Krohn's advice is direct: if you just want to make sure you learn something useful for the coming months and years, that's AI orchestration and agentic AI—how to trust AI systems to make more decisions autonomously, establish the right evaluation mechanisms, and ensure systems are aligned with what we actually want. He is co-writing Engineering AI Agents with Ed Donner, and serves as series editor for Pearson's "Jon Krohn AI Signature Series," whose first two published books are Becoming an AI Orchestrator by Sadie St. Lawrence and Building Agentic AI by Sinan Ozdemir.

Serrano and Starmer added a more down-to-earth view: don't worry about skills becoming obsolete; focus on the problem itself. Starmer said, "Skills come and go; they don't matter. You learn them when you need them and forget them when you don't. But if you focus on the problems you believe need to be solved, the skills will come." Krohn finally admitted that he is "like everyone else, clinging to a plank and drifting at sea"—uncertain, but still paddling.
One-line Summary
Claude Opus 4.6 has stripped twenty years of technical skills of their moat, shifting value toward providing context and business communication.
Paragraph Summary
In a Super Data Science live conversation, Jon Krohn argued that models like Claude Opus 4.6 have eliminated the moat around traditional technical skills such as MATLAB, Python, and machine learning, and that practitioners' value is shifting toward understanding requirements, providing context, improving experience, and communicating consensus. Drawing on his own path from neuroscience to AI, he emphasized that learning fundamentals still enables going one level deeper, and that while vibe coding suits rapid prototyping, the lack of underlying understanding makes later modification costly. The conversation also touched on AI orchestration and agentic AI as the most worthwhile areas to invest in now, and on the tipping-point narrative of AI advancing far faster than intuition suggests.
Related articles

Vergecast Ebook Library Tips: Multiple Virtual Library Cards and Kobo Integration
Vergecast shares ebook library tips: buy directly to support authors and indie bookstores, use multiple virtual library cards via Libby/Overdrive to expand library borrowing, and Kobo e-readers deeply integrate library resources with multi-card support.

Support Vector Machines (SVM): Core Concepts and Python Hands-On Practice
From maximal margin classifiers to the kernel trick, this article explains SVM core concepts in an accessible way and demonstrates a complete polynomial-kernel SVM workflow on the diabetes dataset in Python, including standardization, model training, and decision boundary visualization.

GRPO Reinforcement Learning Explained: Group Relative Policy Optimization That Teaches LLMs to Reason
A deep dive into the GRPO reinforcement learning algorithm: how in-group relative advantages are computed, how clipping and the importance sampling ratio work in the loss, the role of KL divergence constraints, and GRPO's training-cost advantage over PPO—with round-by-round mathematical derivations.