What Should We Prepare for at AI's Inflection Point?
There is a strange time lag in conversations about AI. One person remembers the unreliable answers of a free chatbot from several years ago and concludes that the technology is still not useful. Another says that AI has already transformed the way they work. They appear to discuss the same technology while looking at entirely different moments.
Matt Shumer’s February 2026 essay “Something Big Is Happening” explains that gap in urgent terms. It argues that AI is moving fast enough to reshape knowledge work and society, and that people should begin adapting now. I agreed with parts of the essay and found other parts too certain. Its central question was nevertheless difficult to avoid: how much has the value of familiar work and expertise already changed?
The change reached my work first
The argument did not feel like distant speculation because I had experienced a similar shift as a software engineer.
Early coding assistants were closer to autocomplete. They suggested short functions, explained errors, and produced repetitive code. A developer still had to understand and repair the result. As models and tools improved, the unit of work that could be delegated grew. Given a clear requirement, they began changing multiple files, running tests, diagnosing failures, and revising their own work.
The unsettling part was not merely that code appeared faster. The balance of the work moved away from typing an implementation and toward defining what should be built and deciding whether the result met the requirement. Ideas that once required a larger budget or team became easier to test. At the same time, parts of the technical skill I had spent years developing became broadly accessible.
I saw the same dynamic while building Hereby. AI accelerated the design of document-processing pipelines and the operation of LLM infrastructure. But turning model output into a product still required human decisions about data quality, edge cases, cost, latency, privacy, and responsibility. It was more accurate to say that AI was moving the center of the work than simply eliminating it.
The unit of work is changing faster than the score
AI progress is often discussed through benchmark results. In practice, a more consequential change is the duration of work a system can carry independently. A tool that once completed a sentence begins writing a function, planning changes across a codebase, and testing and revising the result. A system that required guidance at every step takes responsibility for a longer segment of the process.
As that segment grows, the human role changes. Instead of checking every small output, people must define objectives and constraints in advance and create ways to verify a result without watching every intermediate step. The more autonomy an AI receives, the more important logs, tests, approvals, and clear lines of responsibility become.
Shumer argues that the same experience will spread quickly into other forms of knowledge work: law, finance, accounting, writing, and analysis. The timing is open to debate. A job is more than the sum of isolated tasks, and organizational accountability, regulation, and client trust do not change automatically when a model becomes more capable.
Yet it would also be a mistake to dismiss the change. An AI does not need to replace an occupation in one step to transform hiring, training, and the division of work. If it can perform a meaningful share of the underlying tasks faster and at lower cost, the occupation is already being reorganized. Asking only whether a profession will “be replaced” can hide the restructuring underway.
Possibility and limits in legal work
The issue became more concrete as I studied law. AI can summarize long judgments and contracts, identify possible issues, and produce an initial structure for a memorandum. It clearly reduces the time needed to find a starting point in a large body of material.
Legal work, however, does not end with plausible prose. A lawyer must confirm that a cited case exists, determine whether it applies to the facts, and decide which assumptions to make when the record conflicts. An argument favorable to a client is not necessarily advice a professional can responsibly give. Where the cost of error is high, verification and accountability come before speed.
Paradoxically, easier drafting may raise the standard expected of professionals. The value of routine organization and formulaic writing may fall, while the value of framing the right problem, choosing among incomplete facts, and accepting responsibility for the conclusion rises. The relevant question is not how lawyers compete with AI, but what they must learn to judge for themselves in a workflow that includes it.
Why I do not take the warning literally
The urgency of the original essay has a useful purpose. It asks people who are judging present AI from an old experience to try again. But I am cautious about treating every change as if it followed one predictable exponential curve.
Better model performance and real organizational delegation are different things. Data access, security, liability for errors, adoption cost, and integration with existing systems remain. High capability does not automatically create high trust. Reaching expert-level scores on an examination or benchmark does not prove that a system can handle the full complexity of real practice.
Claims that every computer-based job will soon disappear may create attention, but they can also obscure what an individual should do. Fear can push people toward either blind faith or total rejection. A more useful response is to apply the technology to real work and gather evidence about where it is dependable and where it is dangerous.
What can be done now
First, we should stop using AI only as a search box. Give it real work in a field you understand and observe both its quality and its failure modes. Expertise is what lets a person see where a model helps and where it quietly goes wrong. The goal is not one impressive answer, but a workflow that can be repeated.
Second, build verification into the process. Code needs tests and review. Legal research needs primary-source and citation checks. Data analysis needs reproducible inputs and calculations. As AI becomes faster, verification should not disappear; it should be automated where possible and concentrated at clearly defined human checkpoints.
Third, develop adaptability rather than loyalty to one tool. Today’s best model and interface will change. The durable skill is the ability to apply a new tool to a bounded task, compare it with an existing method, and discard it if it does not help.
Fourth, convert expertise from a means of production into a basis for judgment. An AI may generate a draft, but a person who understands the context still decides what matters, which risk to accept, and how to explain the result. Relationships, trust, licensed accountability, and organizational coordination may become more important, not less.
Finally, use the lower barrier to creation as an opportunity. Ideas that once required a large development budget can become small prototypes. An unfamiliar field can be approached through personalized instruction. The risk that existing jobs will change and the possibility that individuals can build more than before are both real.
The question that remains human
I do not agree with every prediction in “Something Big Is Happening.” No one can know exactly which occupations will contract, or by how much, over the next several years. The speed at which AI-assisted research feeds into the next generation of models is uncertain, as is the speed at which institutions will absorb new capability.
What we can already observe is significant enough. AI is moving from a small assistant toward a collaborator capable of longer tasks. As a result, the boundary of work people perform directly is changing. I have seen too much change in my own workflow to dismiss this as a temporary fashion.
Perhaps we do not need to answer the largest question—whether AI will ultimately replace human beings—in advance. It is more practical to ask which parts of today’s work can be delegated, which results must still be checked personally, and how faster production can be converted into a better outcome.
At an inflection point, the most dangerous position is neither optimism nor pessimism. It is evaluating the present through an obsolete experience. Use the systems, question them, verify their work, and retain human responsibility where it belongs. Preparation for an uncertain future will be built through that repeated practice.
Note: This is a personal response to Matt Shumer’s “Something Big Is Happening”, informed by my experience in software engineering and legal studies. It is not a full translation of the original essay.