Skip to main content

We quietly entered the age of abundance

Insights

AI can make implementation easier. Turning that output into useful software and capable machines still takes judgment, evidence, and accountable decisions.

By Emal Alwis

6 min read
Share

AI agents changed how I think about building a company. When more implementation can be delegated, the interesting questions move toward what to build, which constraints matter, and how to know whether the result works. That is the possibility I mean by abundance: more people able to turn useful ideas into working tools.

It is a direction to pursue, not a claim that software is free or that automation benefits everyone automatically. Producing code is one part of engineering. Reviewing it, operating it, protecting the people who depend on it, and correcting mistakes remain part of the work.

Measure the result, not the activity

More commits can mean more progress. They can also mean smaller changes, more retries, or more repair. A contribution graph cannot distinguish those cases. The useful question is whether a team delivers a working result with less total effort, without increasing failures or pushing work onto reviewers.

There is good reason to measure rather than assume. In a randomized study of early-2025 coding tools, METR researchers assigned 246 tasks across 16 experienced open-source developers to allow or disallow AI assistance. In that setting, access to AI increased completion time by 19%, despite participants believing it had made them faster. The study concerns particular tools, experienced maintainers, and mature repositories. It does not establish how every team or later model performs.

For a team evaluating its own workflow, I would track accepted changes, review effort, rework, escaped defects, and time to a usable result. Keep the task mix and quality expectations visible. A faster first draft is useful only if the rest of the process can turn it into something dependable.

Delegation needs a clear contract

Directing an agent resembles technical delegation: define the outcome, supply relevant context, inspect the result, and decide whether it meets the requirement. It does not make alignment automatic. An agent can implement the wrong interpretation, miss a dependency, or produce tests that confirm its own assumptions.

The practical unit of work should therefore include:

  • A concrete user or caller and the outcome they need.
  • Boundaries on data, permissions, and systems the change may affect.
  • Acceptance evidence drawn from the requirement, not just the implementation.
  • A reviewer who can challenge the result and a recovery path if it fails.

This is where judgment matters. Breaking a large request into many tasks does not help if the tasks converge on the wrong product. More parallel activity makes shared assumptions and integration review more important.

More people can participate without removing expertise

The possibility that excites me is a domain expert being able to explore an idea directly: an operations specialist trying a scheduling tool, a designer testing an interaction, or a researcher inspecting a dataset. A prototype can make a conversation concrete before an engineering team commits to a production design.

That does not erase the distinction between a prototype and a service people rely on. Security, accessibility, reliability, and maintenance still need owners. Technical expertise helps people recognize failures that a convincing demonstration can conceal.

I would hire for clear reasoning, curiosity, communication, and the ability to investigate a result that looks right but behaves incorrectly. I would also preserve ways for junior engineers to learn implementation and debugging. Reviewing generated work requires understanding it; skipping that foundation is not a shortcut to senior judgment.

Physical AI has to close the loop in the world

The same distinction applies to robotics data. A proposed annotation is not yet an accepted label. A technically valid recording is not proof that a robot completed its task. A reviewed dataset is not, by itself, evidence that a model improved.

Automation can assist parts of that workflow, but its value needs to be evaluated on the actual sensors, taxonomy, conditions, and acceptance criteria. Human review also needs checks: reviewers can disagree, share a mistaken interpretation, or miss rare failures. The aim is to make those errors detectable and correctable.

For a labeling program, compare the total effort required to produce accepted outputs, including review and rework. For a learning experiment, evaluate the resulting model on an appropriate held-out set. Keep those measurements separate. A software productivity anecdote cannot establish a labeling speedup or a gain in robot performance.

Avala's fleet learning workflow connects the stages around that distinction. Teams can scope a labeling program around their input, output contract, and review requirements. The evaluation should establish what works for that program before broader claims are made.

Abundance is a choice about who benefits

Lower implementation costs can open opportunities. They do not guarantee better jobs, fair compensation, or an easier transition for people whose work changes. Those outcomes depend on decisions about training, access, ownership, and how the benefits are shared.

People who prefer hands-on implementation still bring valuable skills. People entering the field need opportunities to develop them. Domain knowledge and technical knowledge can reinforce each other; treating either as obsolete makes the work weaker.

For Avala, putting people at the center means treating human judgment as work worth supporting, rather than an unlimited resource that appears when automation becomes uncertain. It also means being willing to automate tasks when evidence shows that doing so improves the result. The mission does not require every current task to remain manual forever.

What we are betting on

Our bet is that making tools easier to build can expand what people are able to do. Realizing that possibility takes more than generating more output. It takes a reliable connection between an intention, the work performed, and evidence that the result helped.

That is the ambition behind Avala's vision: machines that learn from real-world experience and make work safer and life easier, with people at the center. The opportunity is large. The next step is still concrete: choose a useful problem, define success, and demonstrate progress.


Emal Alwis is the Founder and CEO of Avala, which builds fleet learning infrastructure for Physical AI.

  • AI Productivity
  • Engineering
  • Founder Insights