Walk into almost any machine learning team and you will hear some version of the same belief: the more options engineers have, the better the outcome. Pick whichever model family you like, whichever training library you prefer, whichever experiment tracker suits your mood. It sounds generous. In practice it means every project begins with a long string of small decisions, and each one drains a bit of the attention that should go into the problem itself. This article takes on six myths about unlimited choice in AI work, and explains why a short, pre-agreed menu usually beats an open field.
The pattern behind all six is the same. Every additional option has a cost, even if you never pick it, because you still have to look at it, weigh it and set it aside. Decide the common choices once, write them down as a short list of presets, and you keep your energy for the parts of the job that genuinely need judgment. Nothing here is a rule that fits every team; treat it as a practical way of thinking.
Myth one: good engineers should always pick the ideal tool for each task
It is a flattering idea, but the “ideal” tool is rarely knowable in advance, and the search for it is expensive. Comparing three plausible libraries for a data loader can eat an afternoon, and the measurable difference to the finished system is often tiny. Meanwhile the team now has three ways of loading data, three sets of quirks and three things to debug at two in the morning.
A better habit is a short approved list. Say the team agrees on one primary framework for training, one experiment tracker and one serving pattern. Engineers can still propose exceptions, but the default is settled, so a new project starts at the interesting part instead of at a shopping trip. The saved hours compound across every project that follows.
Myth two: more hyperparameter freedom means better models
Teams often treat the search space as a badge of seriousness: dozens of knobs, wide ranges, endless sweeps. Yet a huge space makes results harder to compare and easier to over-interpret. When ten things changed between two runs, nobody can say which one mattered, and the quiet temptation is to pick whichever number looks best and tell a story afterwards.
Config presets help here. A short set of named profiles, for example a quick smoke-test profile, a standard training profile and a careful final-run profile, fixes most of the knobs in advance. Engineers adjust only the two or three parameters that the problem truly calls for. The result is cleaner comparisons and fewer arguments about whether a gain was real.
Myth three: standard templates stifle creativity
This one is popular with people who have never lost a week to an unreproducible notebook. Creativity in AI work lives in the framing of the problem, the choice of features, the way errors are analysed and the questions asked of the data. It does not live in whether the repository layout has a folder called “src” or “code”.
When project structure, logging format and evaluation scripts are fixed, you actually free up room for creative thinking. Nobody burns attention deciding how to name an output directory. Reviewers know where to look. A newcomer can read a teammate’s work on the first day because the shape is familiar. Constraint on the boring parts is what pays for experimentation on the important ones.
Myth four: a bigger menu of models protects you from picking the wrong one
It feels logical: if you can choose from every architecture, surely you will land on the right one. In reality a wide menu mostly increases the chance of chasing novelty. A newly published architecture looks exciting, so a team spends weeks trying it, only to find that a well-understood baseline would have done the job.
A practical alternative is a ladder with just a few rungs. Start with a simple baseline. Move up only if the baseline’s errors show a specific weakness that a more complex model addresses. Keep the ladder short and written, so each step requires a reason. You can read about how organisations weigh machine learning against ordinary reasoning in this piece on how AI supports business decisions, and the same discipline of using a method only where it earns its place applies inside the team as well.
Myth five: limiting choices only matters for junior engineers
Senior people feel the pull of open-ended choice as strongly as anyone, and they tend to carry heavier decision loads: architecture, trade-offs, review calls, hiring. If a senior engineer spends the morning picking between logging styles, there is less left for the design review at lunchtime. A fixed menu is not a leash on juniors; it is a courtesy to everybody, because it removes trivial calls from every calendar.
It also makes reviews faster. When a pull request follows the standard layout and uses the approved libraries, the reviewer can concentrate on the logic. When it uses something unusual, that unusualness is itself a signal worth discussing, because it was a deliberate departure rather than a default.
Myth six: fixed presets make a team rigid
The worry is understandable, yet it confuses a default with a prohibition. A good short list has an exit door: a documented way to propose a change, plus a regular, scheduled slot for revising the list. Exceptions are allowed, but they carry a short written reason. That tiny bit of friction is useful, since it separates genuine needs from passing whims.
Rigid teams are the ones that never revisit their list. A team that reviews its approved tools every quarter, drops what no longer helps and adds what has proved itself is adaptive in the way that matters. It simply does not re-open the same question every Monday. The manufacturing world offers a nice parallel: lines that standardise settings still improve over time, as discussed in this look at optimising production with AI.
A sample short menu for an applied group
Picture a small applied AI group drafting its list. It is deliberately modest, and every team should make it their own.
- One approved training framework, one experiment tracker and one way to package models for serving.
- Three named config profiles: a fast smoke test, a standard run and a final careful run.
- A fixed evaluation script that every project uses, with the metric choices agreed up front for each problem type.
- A repository template with the same folders, the same readme headings and the same logging setup.
- A one-page note describing how to propose an exception and when the list is reviewed.
Pinning down these five items can take an afternoon. They then save time on every project for months.
A note from outside the lab
The same logic shows up in the way small organisations handle time away from work. The website Inoroff describes a tool for small and mid-sized teams that gathers check-ins, leave requests and manager approvals into a set flow rather than a scatter of messages. Whatever your view of any specific product, the instinct will ring true for anyone who has pruned a tool list: settle the routine, then stop renegotiating it.
How to start without a big reorganisation
You do not need a company-wide mandate. Begin with the next new project. Before the first commit, spend thirty minutes listing the choices you would normally make ad hoc: framework, tracker, config style, evaluation script, naming. For each, pick a default and write it in a short document. Use it on that project, note where it chafed and revise. By the third project the list will have settled into something your whole group recognises.
Resist the urge to build a grand platform around the list. A page of text and a repository template cover most of the value. The point is fewer decisions, not more tooling.
The practical takeaway
Open-ended choice is not freedom when it quietly spends the attention you need for hard problems. The six myths above all rest on the same assumption, that more options always help. They do not. A short approved menu of tools, presets and templates, reviewed on a schedule and open to reasoned exceptions, lets your team spend its best thinking on data, errors and results. Try it on one project, keep the list short, and see how much calmer the work feels.