Matthew Berman on AI Progress, Work, and Preparing for Change

Matt Shumer wrote an essay called “Something Big Is Happening,” and I think its message deserves attention. He compares the current moment to the early stages of the pandemic, when some people were already worried but many others had not yet changed their behavior. The analogy is an argument about awareness, not proof that the two events will unfold alike.

The essay argues that AI is improving quickly while many people are still making plans as if work will remain much the same. A small group of companies and decision-makers has enormous influence over how these systems are built and deployed. I want to explain why I think the essay’s warnings are worth considering.

Shumer describes a personal change in his own coding work. Tasks that once meant writing code himself can now be delegated to an agent, with him explaining the goal and checking what comes back. That is one person’s experience, but I recognize the shift from my own work.

I use AI tools to build and automate pieces of my video workflow. I can describe an application, inspect the result, and keep asking for changes. The process still requires me to know what I want, notice mistakes, and decide whether the result actually works.

What makes this different from an autocomplete tool is the amount of a task the agent can attempt before it needs another instruction. It can explore files, make changes, run commands and iterate. It is not infallible, and its performance depends on the task, the tools available and the supervision around it.

I encourage people to try current tools on a real problem in their own field rather than rely only on old impressions. I think paid models are useful for seeing what the strongest offerings can do, although a subscription is not a prerequisite for learning or evaluating whether a tool fits your work.

The pace of recent improvements makes it hard to project the future from a single older demonstration. At the same time, a compelling demo should not be mistaken for evidence that every task in a profession can be done reliably or economically by an agent.

The METR chart looks at AI performance on tasks grouped by the time a human expert needs to finish them. Its 50% horizon estimates the human task duration at which a model succeeds half the time on the benchmark. It is a measure on a particular task set, not a clock showing how long an agent can autonomously run any job.

Anthropic described using multiple agents to build a C compiler over roughly two weeks, at a reported API cost around $20,000. The researchers designed the environment, tests and feedback the agents used. The experiment demonstrates a significant project under those conditions; it does not mean a model independently replaced a software team.

I also discuss Cursor’s browser-building experiment, in which agents generated a large codebase and produced something that worked to some degree. Line counts are an imperfect proxy for usefulness or quality, but the examples show why coding agents are drawing attention.

OpenAI said an early version of GPT-5.3-Codex helped with work involved in its own development and debugging. I see that as an example of AI contributing to AI research. That is different from demonstrating an unsupervised, self-sustaining loop in which a model independently improves itself.

Shumer and other AI leaders warn that rapid improvements could disrupt white-collar employment. Those warnings are forecasts, not settled estimates of how many jobs will disappear or when. The effect will depend on reliability, cost, regulation, organizations and the choices people make.

A job is a collection of tasks, relationships and responsibilities. Automating one part can change the rest without necessarily eliminating the entire role, while some organizations may need fewer people for certain work. I think the possibility of significant disruption is serious enough to consider before it arrives.

Human judgment becomes important when a system produces plausible work that still needs to be checked. Understanding the problem, setting a goal, assessing consequences and taking responsibility for an outcome are not the same as accepting an answer because it sounds confident.

My first practical suggestion is to experiment. Choose a task you know well, ask an AI tool to help, and compare the result with what you would have done. This makes it easier to discover both useful capabilities and the mistakes that glossy demonstrations can hide.

Learn how to give context and evaluate the output. Do not put sensitive information into a service without understanding its handling of that information, and do not treat a model’s answer as a substitute for independent review when the stakes are high.

A McKinsey chart I discuss says 7% of surveyed organizations using AI had fully scaled it across their enterprises. That figure describes a subset of survey respondents, not all companies. It suggests that even where people report using AI, organization-wide deployment can lag behind experimentation.

I also suggest thinking about financial flexibility and avoiding assumptions that current employment conditions will remain unchanged. People’s circumstances differ, so the appropriate steps depend on their resources, obligations and tolerance for risk.

For families, the question is how to help children use new tools while retaining the ability to think, learn and act independently. Fundamentals and curiosity still matter: a tool is more valuable when its user can tell whether the result makes sense.

I see opportunities as well as risks. An individual or small team may be able to test ideas and build products that previously required more resources. The opportunity does not guarantee success, but it is a reason to learn by making something concrete.

The main point is not that anyone can predict the precise timetable. It is that waiting for perfect certainty may leave people less prepared. Try the tools available to you, keep your judgment engaged, and adjust as the evidence changes.

If you can afford to experiment with a paid service, I think it can be worthwhile; free options and shared resources can also provide a starting point. The goal is to understand the technology and its limits through experience.