AI, Jobs, and Your Career as an ML Engineer

Right now, ML engineering is one of the best jobs in the economy. Demand is up 89% in the first half of 2025. Median salary sits at $187,500. Senior roles at the major tech companies pay $320K to $550K in total comp. Generative AI specialists command 40 to 60% premiums on top of that. The talent deficit is 3.2 jobs for every qualified candidate, which means you have real leverage in negotiations.

And yet, the role you were hired for is not the role that will exist in three years. You are building the systems that will eventually automate parts of your own work, and the window for repositioning is measured in months. That tension — between a historically great job market and a rapidly shifting foundation underneath it — is what this episode is about.

Before we get into the evidence, here's the short version for anyone whose commute is about to be 30 minutes of escalating anxiety. You're fine right now, and you have time to act, but you need to act. The concrete version: get your hands dirty with production ML systems — deployment, monitoring, evaluation, not just training. Pick a domain outside pure tech — healthcare, climate, finance, robotics — and learn it well enough to have opinions about what models should do, not just how they work. Build something end-to-end that ships to real users, because the gap between "can train a model" and "can run a model in production" is where the job security lives. Get fluent with AI coding agents, not as a novelty but as your daily workflow. And start developing the product instinct to know what's worth building, because as code gets cheaper, knowing what to build becomes the scarce skill. We'll get much more specific later in the episode, but if you remember nothing else: the safe zone is the intersection of technical depth and something AI structurally can't do yet. Stay there, keep moving, and you'll be ahead of most people in this field.

Now, the evidence. We're going to look at the displacement data as it actually stands today, walk through what happened to the professions that got hit first, examine why the common escape routes have serious problems, and then get very specific about what ML engineers should do with the next few years.

What the displacement data actually shows

The headline numbers sound manageable. Through November 2025, Challenger, Gray & Christmas counted about 55,000 U.S. job cuts directly attributed to AI. That's 4.5% of total layoffs. Yale Budget Lab found, in their words, no discernible disruption at the macro level. Oxford Economics says firms don't appear to be replacing workers with AI on a significant scale. If you stopped there, you'd conclude this is overblown.

But the aggregate data is hiding what's happening at the edges, and the edges are where careers live. Stanford's Digital Economy Lab analyzed ADP payroll data covering millions of workers and found a 13 to 16% relative decline in employment for workers aged 22 to 25 in the most AI-exposed jobs since late 2022. Workers over 30 in the same occupations saw employment grow 6 to 12%. That asymmetry is the key to understanding what's really going on. The displacement isn't showing up as mass layoffs. It's showing up as jobs that simply never get posted. The Bureau of Labor Statistics reported the lowest rate of job openings in professional services since 2013, down 20% year-over-year in January 2025. Hiring for positions paying $96K and above hit decade-low levels. In the UK, graduate job postings dropped 67% since 2022.

The company-level evidence makes the pattern even clearer. Salesforce cut 4,000 customer support workers after AI took over 30 to 50% of the company's workload, reducing support headcount from around 9,000 to about 5,000. Microsoft eliminated roughly 15,000 roles through 2025, and Bloomberg analysis showed that more than 40% of their Washington state cuts targeted software engineers specifically. CEO Nadella confirmed that about 30% of Microsoft's code is now AI-generated. Amazon cut 30,000 corporate jobs while pouring over $100 billion into AI infrastructure. Shopify's CEO mandated that employees must demonstrate AI can't do a job before requesting new headcount. A survey of more than 2,000 C-suite leaders found that 46% have already reduced headcount due to AI, with 54% expecting fewer employees within five years.

The pattern across every company is the same sequence: contractors and freelancers go first, then entry-level positions, then attrition without backfill. Klarna eliminated the equivalent of 700 customer service agents. Duolingo cut 10% of contractors. IBM paused hiring for routine roles. JPMorgan told managers to avoid hiring as AI deployed. The message from every boardroom is identical: prove a human is necessary, or the position disappears.

Only 9.7% of U.S. firms use AI in production today, and only 15% have achieved enterprise-wide implementation. Goldman Sachs' chief economist said AI had basically zero effect on U.S. GDP growth in 2025. That's both the optimistic case and the alarming one. If this much disruption is happening when adoption is still in single digits, the question is what happens when it reaches 30 or 50%.

What happened to the artists

The destruction of the freelance illustration market is the most important case study for understanding what's coming to other knowledge work. It happened fast, it was decisive, and the dynamics transfer directly.

In China's gaming sector, illustrator jobs declined 70% in a single year. Artists who previously created one scene per day found themselves producing 40 or more with AI assistance, which meant far fewer artists were needed. Freelance commission rates for Chinese illustrators dropped from 3,000 to 7,000 yuan per poster to about one-tenth of that. On Western freelance platforms, researchers at Imperial College Business School found graphic design jobs fell 18.5% and writing jobs fell 30% within eight months of ChatGPT's launch. And here's the counterintuitive finding: experienced freelancers offering premium services were hit harder than cheaper alternatives, because AI competed most directly with quality work priced above AI's near-zero marginal cost. That flips the usual assumption that skill protects you.

The stock photography industry tells the same story in corporate form. Getty Images' creative segment declined about 5% year-over-year. Shutterstock's per-download earnings fell 37% for videos in 2024. The two companies announced a defensive merger in January 2025, stripping $150 to $200 million in overhead. In a particularly grim twist, Shutterstock earned $104 million licensing its own images to the AI companies that were destroying demand for its archive.

The testimonials from working artists are hard to read. A children's book illustrator said flatly that illustrator, as a job, does not exist anymore. A comics artist with more than 20 years of experience said that in 2023 it seemed like overnight all those jobs disappeared. These are not people on the margins. They are experienced professionals whose entire market evaporated in months.

The lesson that matters for ML engineers is this: artists weren't replaced because AI was better than them. They were replaced because AI was good enough and essentially free. Think about what that means concretely. A game studio that used to hire 10 concept artists for a project now hires two, gives them AI tools, and gets roughly comparable output. The eight artists who lost their seats didn't lose them because their work was bad. They lost them because the studio could get most of the quality for a fraction of the cost, and for concept art that gets iterated on anyway, that was sufficient.

That principle — adequacy at near-zero marginal cost — applies directly to entry-level code, legal research, marketing copy, customer service scripts, and basic data analysis. A junior developer writing boilerplate CRUD endpoints is doing work that AI coding tools can now produce in seconds. A paralegal summarizing case law is doing work that fits squarely in the sweet spot of what language models handle well. The quality bar for displacement is not excellence. It's sufficiency at a fraction of what a human costs. If your work can be described as taking a standard input and producing a standard output, the timeline is short regardless of how good you are at producing that output.

The artists who survived did so through a narrow set of strategies: extreme specialization, moving from execution to creative direction, building personal brands around distinctive styles, and — this is the important one — mastering AI tools themselves to stay ahead of the productivity curve. The ones who waited to see how things developed were devastated. The adaptation window was shorter than anyone expected.

The trades problem

Whenever AI displacement comes up, someone suggests learning plumbing or electrical work. The surface-level numbers support this. The U.S. construction industry is short approximately 500,000 skilled workers. There are 2.9 million annual job openings in trades versus only 1.25 million qualified graduates, leaving an annual shortfall of 1.7 million. Wages are up 20 to 30% since 2020. Geoffrey Hinton himself has said plumbers are less at risk than paralegals.

But this advice has three serious problems that rarely get discussed.

The first is a bottleneck. Becoming a licensed plumber or electrician requires four to five years of apprenticeship. Training capacity is constrained because fewer experienced tradespeople means fewer people available to mentor apprentices. If even a fraction of the roughly 10 million American workers Goldman Sachs estimates could be displaced by AI pivot to trades simultaneously, the system simply cannot absorb them. Trade school enrollment is already growing at 6.6% compound annual growth, and search traffic for trade schools rose 27% in 2024. The shortage is real today, but mass entry could flip the market to saturation within five to seven years.

The second problem is the physical reality of the work, which knowledge workers tend to romanticize. Construction accounts for one in five workplace deaths in the U.S. One in 10 construction workers is injured every year. The career fatality risk is 1 in 200. Musculoskeletal disorders dominate long-term health outcomes. Construction workers have the highest suicide rate of any U.S. occupation, at 56 per 100,000 males. Career changers entering at 35 or older face higher initial injury rates and shorter productive careers. For someone who's spent their working life at a keyboard, this is a fundamentally different kind of risk calculation.

The third problem is that the robots are coming for physical work too, just on a longer timeline. Venture capital put $2.5 billion into humanoid robotics in 2024 alone. China invested $7 billion in the first nine months of 2025, up 250% year-over-year. Unitree's R1 humanoid launched at just $5,900. China accounted for more than 80% of all 16,000 global humanoid installations in 2025. Goldman Sachs projects a $38 billion humanoid robot market by 2035. Figure AI's robots already operate daily 10-hour shifts at BMW, having loaded over 90,000 parts across more than 1,250 runtime hours.

Current robots work in structured environments like warehouses and factories, with roughly two-hour battery life. They can't crawl under houses or diagnose electrical faults in century-old wiring. Bain & Company estimates meaningful construction automation is 10 to 15 years away. But partial automation — reducing a crew of 10 to six — could begin much sooner. Someone entering a four-year apprenticeship today faces the risk of graduating into a market that has already started shifting. Trades are a window, not a permanent refuge.

The orchestrator treadmill

The most popular career advice in the AI era is to move up one level of abstraction: stop writing code, start directing AI agents. Become the orchestrator. The progression sounds clean — coder to conductor to strategist. But the evidence suggests this is a treadmill where the belt keeps accelerating underneath you.

The prompt engineering career arc is the cautionary tale that everyone in this field should study. In 2023, Anthropic advertised prompt engineering roles at $375,000. By 2025, Microsoft's survey of 31,000 workers ranked Prompt Engineer second to last among new roles companies were considering. Job postings for the role on Indeed went to minimal. Intel Labs demonstrated that automated prompt engineering already outperforms human prompt engineering. The entire lifecycle — emergence, hype, enormous salaries, democratization, obsolescence — took roughly two years.

AI coding agents are already demonstrating the ability to handle orchestration tasks autonomously. Claude Code resolves 72% of medium-complexity GitHub issues in under eight minutes. Devin handles planning-to-deployment cycles end-to-end and is being piloted at Goldman Sachs alongside 12,000 human developers. Gartner reported a 1,445% surge in multi-agent system inquiries from Q1 2024 to Q2 2025. The pattern in multi-agent architectures is already clear: a capable model creates a strategy, cheaper models execute it, and costs drop by 90%. That pattern pushes the human orchestrator further and further from the actual work.

There's a deeper problem with the orchestrator strategy that rarely gets discussed. If AI handles the entry-level work that trains future experts, who becomes tomorrow's senior engineer? As Johns Hopkins researchers noted in February 2026, junior workers are being replaced by AI, and they're not able to manage AI because they don't have the experience. You can't orchestrate what you don't understand. The industry is producing what you might call fragile expertise — people who can direct AI tools but can't spot when those tools are wrong, because they never did the underlying work themselves. Picture a future team lead who learned to code entirely through AI assistance, reviewing AI-generated pull requests for a system they've never debugged manually. They can tell the AI to write a retry mechanism, but they can't evaluate whether the retry logic will cause a thundering herd problem under load, because they've never been paged at 3 AM for one.

Each abstraction layer in computing history — assembly to compiled languages to frameworks to AI-assisted coding — has eventually been commoditized. The orchestrator role is real, but it's a transitional phase, not a destination. Optimistically, the window is three to five years before the treadmill catches up.

The demand problem nobody wants to discuss

If AI displaces enough workers, who buys the products? This isn't an abstract concern — it's the logical terminus of every AI cost-cutting strategy being pursued simultaneously across the economy. Consumer spending accounts for roughly 70% of U.S. GDP. White-collar workers drive about 75% of discretionary spending. A widely discussed February 2026 memo coined the term Ghost GDP: AI-inflated national accounts that never circulate through the real economy because machines spend zero dollars on discretionary goods.

The serious economists are divided but uniformly worried. Daron Acemoglu, who won the Nobel Prize in Economics in 2024, warns of what he calls the worst of all possible worlds: none of the transformative productivity gains and all of the displacement. He estimates AI will increase total factor productivity by only about 0.66% over 10 years, far below the hype. He's coined the term "so-so automation" for tools that save money for individual corporations but fail to raise aggregate productivity — self-checkout kiosks being the canonical example, having eliminated around 300,000 cashier jobs without measurably improving anything for consumers. Acemoglu has explicitly warned that if the economy goes down this path of destroying jobs and increasing inequality, democratic institutions are at risk.

David Autor at MIT offers a more nuanced view. His finding that 60% of current occupations didn't exist 80 years ago is genuinely grounds for optimism — new technology has always created categories of work that were unimaginable before the technology existed. But Autor draws a sharp distinction. AI could restore the middle class by enabling mid-skill workers to perform expert-level tasks: a nurse practitioner handling more complex diagnostics, a paralegal doing more sophisticated legal work. Or it could destroy the middle class by simply replacing those workers entirely. The outcome depends on whether AI is deployed for augmentation or automation, and right now, market incentives overwhelmingly favor automation.

Erik Brynjolfsson offers probably the most empirically grounded optimism. His call center study showed 34% productivity improvement for novice workers with minimal impact on the most skilled, suggesting AI could reduce inequality by spreading expertise downward. But his own 2025 ADP data showing 13 to 16% employment decline for young workers tells the other side. Both findings are true at the same time: AI helps the workers who remain while reducing how many workers are needed.

On the policy front, the responses are fragmented. UBI experiments consistently show that recipients don't stop working — Stockton's SEED program actually saw full-time employment increase — but funding mechanisms remain unsolved. Various robot tax proposals and sovereign wealth fund concepts exist as serious policy ideas but face enormous political obstacles. The EU AI Act mandates worker notification and human oversight for HR AI systems. The U.S. under the current administration is actively deregulating. China is dual-tracking, accelerating AI development while expanding retraining subsidies and social safety nets. Nobody has a credible plan for what happens if the accelerationist timeline turns out to be right.

Where the experts actually disagree

The expert landscape splits into three camps, and which one you believe has radically different implications for career planning.

The accelerationists — Dario Amodei, Sam Altman, Mustafa Suleyman — believe major disruption is one to five years away. Amodei predicts AI systems broadly better than all humans at almost all things by 2026 or 2027, and warns of a white-collar unemployment wave potentially reaching 10 to 20%. Suleyman has said white-collar work has a year to 18 months until AI completely automates it. Altman says certain categories of jobs like customer support are completely going away, but believes society will adapt through new job creation. These are the people building the systems. Their timelines may be self-serving, but they also have the best visibility into what's in the pipeline.

The skeptics — Yann LeCun, Gary Marcus, Karl Friston — argue that current AI architectures are fundamentally limited. LeCun, who left Meta in late 2025 to found Advanced Machine Intelligence Labs, insists that large language models are a dead end that lack world models, causal reasoning, and genuine understanding. He predicts meaningful adoption will take 10 to 15 years. Marcus has bet 10-to-1 against AGI by end of 2027, arguing that hallucination problems are architecturally unsolvable. These aren't Luddites — LeCun is a Turing Award winner, Marcus is a respected cognitive scientist — but they may be underestimating the practical impact of AI that is imperfect but cheap.

The pragmatists — Andrew Ng, Brynjolfsson, Autor — focus on actionable advice rather than timeline predictions. Ng's December 2025 Stanford lecture is probably the single most useful resource for ML engineers right now. His core argument is that as AI makes code cheap, the bottleneck shifts from implementation to specification: knowing what to build becomes more valuable than knowing how to build it. The "product engineer" who combines technical depth with user empathy is the most valuable person in the industry. Traditional ratios of product managers to engineers, which used to run 1-to-4 or 1-to-8, are collapsing toward 1-to-1. Ng's practical observation is that tools change every three to six months, and falling even half a generation behind creates a meaningful productivity gap. The common denominator he's observed among his most successful students is simply extraordinary work ethic.

The honest answer about timelines is that nobody knows. Expert AGI estimates range from 2026 to 2047 and beyond, with a weighted median somewhere around 2035 to 2040. The field has consistently shortened its estimates in recent years. Pre-training scaling does show diminishing returns — knowledge benchmarks scale at roughly N-to-the-0.3, with plateaus emerging at 30 billion parameters and above. Ilya Sutskever has declared the age of pure scaling over. But new scaling dimensions keep appearing: test-time compute, reasoning chains, multi-agent collaboration. The most intellectually honest position is to plan for a range of scenarios: optimize for the three-to-five year window where human ML expertise remains highly valuable, while building skills that hold up even if general AI arrives later than that.

Moravec's paradox — the observation that physical and embodied intelligence is harder for AI than abstract reasoning — still broadly holds. AI solves International Mathematical Olympiad problems but can't write the answers with a pencil. That suggests knowledge work is actually more vulnerable than physical work, which is the opposite of what most people assume. But it also means that roles requiring physical-world interaction, embodied understanding, or navigation of genuinely unstructured environments have longer runways.

What ML engineers should do now

A caveat before we get specific: the tools and platforms mentioned here are current as of early 2026. Some of them will be obsolete or superseded within a year. That's actually part of the point — the ability to pick up new tools quickly matters more than loyalty to any particular one. But you still need to start somewhere, and vague advice to "stay current" isn't actionable. So here's what the concrete version looks like, with the understanding that the specifics have a shelf life.

78% of AI and ML positions now target professionals with five or more years of experience. Pure entry-level ML jobs are highly competitive and shrinking because AI is automating routine data preparation, model tuning, and experimentation. AutoML platforms let business analysts with no coding experience build predictive models that used to require weeks of data scientist work. If you're early-career, the on-ramp is narrowing fast. If you're mid-career or senior, you have leverage, but the clock is still running.

The near-term priority list

The single highest-value skill shift right now is from model training to model operations. Most ML engineers spent their formative years optimizing architectures and tuning hyperparameters. The market has moved. What companies are desperate for is people who can take a model from "works on my laptop" to "runs reliably in production at scale, and we know when it's broken."

Concretely, that means getting hands-on with deployment infrastructure. If you haven't already, spend serious time with serving frameworks — vLLM for high-throughput LLM inference, or TensorRT for latency-sensitive applications. Set up a model behind an API with proper load balancing, autoscaling, and graceful degradation. Understand the difference between batch inference and real-time serving and when each is appropriate. Build monitoring that tracks not just uptime but output quality: drift detection on input distributions, automated evaluation of model outputs against ground truth, alerting when confidence scores shift in ways that suggest the world changed but the model didn't. Tools like Weights & Biases, MLflow, and Arize exist for this; pick one and learn it well enough to set up a complete monitoring pipeline from scratch.

LLM evaluation is its own emerging discipline, and if you can do it well, you're immediately more valuable than someone who can only fine-tune. The problem is that traditional ML metrics don't transfer cleanly. Accuracy on a benchmark tells you very little about whether a language model will hallucinate dangerous medical advice or leak PII from its context window. Learn to build evaluation harnesses that test for the failure modes that actually matter in production: factual grounding, instruction following, refusal behavior, consistency under paraphrase, and adversarial robustness. Frameworks like DeepEval, RAGAS for retrieval-augmented systems, and Anthropic's own evaluation tooling are good starting points. The ability to design an eval suite that catches the problems your specific application cares about — not just run someone else's benchmark — is what separates a senior ML engineer from someone who can follow a tutorial.

Fine-tuning is another area where the gap between "I've done it" and "I can do it well for production" is enormous. Most ML engineers have fine-tuned a model on a Hugging Face dataset and posted the loss curve. Far fewer have done the hard work of curating a high-quality dataset for a specific business use case, choosing the right fine-tuning approach for the constraints (full fine-tune, LoRA, QLoRA, depending on compute budget and how much behavioral change you need), evaluating whether the fine-tuned model actually improved on the specific task without regressing on everything else, and then deploying the result with proper A/B testing against the base model. If you can walk into an interview and describe that full loop — data curation, method selection, evaluation design, production deployment, measurement — for a real project, you're in strong shape.

Agentic AI workflow design is where a lot of the new engineering work is going. This means building systems where language models don't just respond to prompts but take actions: calling APIs, reading databases, making decisions, executing multi-step plans. The engineering challenges here are different from traditional ML. You need to think about error handling when an agent takes a wrong step three actions into a chain. You need retry logic that doesn't compound costs. You need human-in-the-loop checkpoints for high-stakes decisions. You need observability into what the agent did and why, so you can debug failures after the fact. If you haven't built a multi-step agent system that handles real-world messiness — network failures, ambiguous instructions, conflicting information from different tools — that's a high-priority project to take on. Frameworks like LangGraph, CrewAI, and the tool-use capabilities built into the major model APIs are the current options, though this space is moving fast enough that the dominant framework a year from now may not exist yet.

AI risk mitigation deserves its own paragraph because it's becoming a genuine career specialization. Andrew Ng's guest speaker Laurence Moroney identified the number-one interview differentiator right now as the ability to identify and manage deployment risks. That means understanding prompt injection attacks and how to defend against them. It means knowing how to implement content filtering that catches harmful outputs without creating so many false positives that the product becomes useless. It means understanding the legal landscape around AI-generated content well enough to advise product teams on what they can and can't ship. Companies want people who can say "here's where this will fail and here's how we mitigate it" — not just people who make the demo work.

On tooling: fluency with AI coding agents is no longer optional. If you're not using Claude Code, Cursor, Windsurf, or a comparable tool as part of your daily workflow, you're leaving significant productivity on the table, and the gap between you and someone who does use them is already visible in output volume. Ng's observation that his preferred tool changes every three to six months is consistent with what most practitioners report. The meta-skill is the ability to evaluate and adopt new tools quickly, not mastery of any particular one.

The longer game

For positioning over the next three to ten years, the durable specializations cluster around things AI does poorly, and the key is to get specific about which flavor of "things AI does poorly" matches your strengths and interests.

AI safety and alignment is growing rapidly as frontier models become more capable, and the supply of qualified people is far below demand. If you have the inclination, this means studying interpretability techniques — mechanistic interpretability work on understanding what's happening inside neural networks at the circuit level, activation patching, sparse autoencoders for feature discovery. The field is small enough that meaningful contributions are possible without being at a top lab, and the career trajectory is strong because regulatory pressure is only going to increase demand.

Domain expertise combined with ML is where 57.7% of ML job postings are heading — they now prefer domain specialists over generalists. The actionable version of this advice is: pick a domain and commit to learning it deeply enough that you can identify problems worth solving, not just implement solutions someone else specified. Healthcare ML means understanding clinical workflows, HIPAA constraints, the FDA approval process for AI-based diagnostics, and why clinicians distrust black-box models. Climate ML means understanding atmospheric modeling, satellite data pipelines, carbon accounting standards, and what policy decisions your models might inform. Finance ML means understanding market microstructure, regulatory requirements, model risk management frameworks, and why your risk team will never let you deploy a model they can't explain to an auditor. In each case, the domain knowledge takes years to develop, which is exactly why it's a moat.

The "small AI" opportunity — optimizing models for on-device, self-hosted, privacy-preserving deployment — is a genuine and growing niche. Moroney predicts the industry bifurcating into centralized AGI players running massive cloud models and a parallel ecosystem of self-hosted edge AI. Many industries will never send proprietary data to cloud models. A law firm isn't going to pipe confidential case documents through a third-party API. A film studio isn't going to upload unreleased footage to someone else's servers. Defense contractors are legally prohibited from doing so. If this interests you, the concrete skills are model distillation and quantization (taking a 70B parameter model and compressing it to something that runs on a single GPU or even a phone), efficient inference optimization, and understanding the tradeoffs between model size, latency, and task-specific performance. Engineers who can take a customer's problem, select or fine-tune an appropriately sized model, optimize it for their hardware constraints, and deploy it in an air-gapped environment — those people have an uncrowded market with real pricing power.

The MLOps specialization deserves a specific mention. LinkedIn data shows 9.8 times growth over five years, and the role sits at the intersection of development and production in ways that resist full automation. Production environments are messy, unpredictable, and specific to each organization. Every company's data pipeline has its own peculiar failure modes, its own legacy systems that can't be migrated, its own compliance requirements that don't match the textbook. That organizational specificity is a real moat, at least for now.

More broadly, the career moves with the most durable advantage combine technical depth with something AI cannot easily replicate. Deep domain knowledge in fields where the cost of errors is high. Systems architecture at scale where failures have real-world consequences — the kind of judgment that comes from having been paged at 3 AM when a pipeline went down and understanding viscerally what "reliability" means in practice. Cross-functional communication, translating business problems into ML solutions and back again in language that both sides can act on. And governance and policy work, where regulatory expertise intersects with technical understanding. None of these are bulletproof, but they have more staying power than pure implementation skills.

The optimistic case and its limits

The strongest argument against the pessimistic reading is empirical: the macro disruption hasn't happened yet. Yale Budget Lab, Oxford Economics, and Census data all converge on this. Goldman Sachs' chief economist says the GDP impact has been basically zero. Enterprise adoption barriers remain substantial — 60% of AI leaders cite legacy system integration and compliance as primary challenges, only about 25% of AI initiatives achieve expected ROI, and fewer than 20% have been fully scaled. 42% of C-suite executives report that AI adoption is tearing their company apart. Klarna's reversal, where they cut customer service staff and then had to rehire after quality collapsed, is likely a pattern that will repeat. Gartner predicts 50% of companies that cut jobs for AI will rehire by 2027.

Autor's finding about 60% of occupations being new is genuinely powerful. AI is already creating roles that had no analog five years ago: AI safety researcher, alignment engineer, AI governance specialist. Technology has always generated categories of work that were unimaginable before the technology existed.

But the optimistic case has an important weakness, which is the difference between things working out eventually and things working out for you personally. The Industrial Revolution's Engels' Pause lasted 40 to 60 years, during which productivity grew substantially but wages stagnated and human suffering was extreme. Even if AI creates net new jobs over decades, the transition for people working right now can be brutal. The ATM story is instructive: bank teller employment actually rose after ATMs first appeared, because cheaper branch operations meant banks opened more branches. But then mobile banking eliminated the need for branches entirely, and teller employment collapsed. The first chapter was optimistic. The second wasn't.

The most honest assessment is that AI displacement fears are premature but not unfounded. The macro signal is calm. The micro signals — entry-level hiring freezes, freelancer income collapse, corporate headcount mandates — are alarming. And capabilities are improving at a rate that could shift the macro data within one to three years.

What this adds up to

ML engineering is among the best careers in the economy right now. The 3.2-to-1 talent deficit means you have leverage, and you should use it.

The role is transforming underneath you. The ML engineer of 2028 will look very different from the ML engineer of 2023. Production systems, safety, domain expertise, and product instinct are consuming what was once primarily a research-and-implementation job. Adapt continuously, or risk following the trajectory of prompt engineering — a role that went from $375K salaries to near-obsolescence in about 24 months.

Entry-level knowledge work is being systematically eliminated. If you're junior, the on-ramp is narrowing. If you're senior, the pipeline that produces your replacements and future colleagues is being hollowed out, which creates long-term fragility in the entire profession.

Every supposed safe harbor — trades, orchestration, product management, creative direction — is temporary to varying degrees. Trades have perhaps a 10-to-15 year window before partial robotics automation. Orchestration may have three to five years. The only durable strategy is building capabilities at the intersection of technical depth and something AI structurally struggles with: physical-world reasoning, deep domain expertise, organizational trust, regulatory judgment, or the ability to generate genuinely novel approaches.

The policy response is not keeping pace. The EU is regulating. The U.S. is deregulating. China is doing both. Nobody has a credible plan for what happens if the accelerationists are right about timelines. The gap between the speed of AI capability development and the speed of institutional response is the single largest risk factor for everyone in this field.

We covered the specific skills and moves earlier, so I won't repeat them. The overarching principle is simple: get to the intersection of technical depth and something AI structurally can't do yet, and keep moving as that intersection shifts. Ng is right that falling half a generation behind on tools is already costly. And maintain the intellectual honesty to recognize that if general AI arrives on the accelerationist timeline, all career advice becomes temporary, and the only question that matters is whether we built the institutions to handle it.

The artists didn't get this warning. You just did.