The Machine Learning Lessons I’ve Learned Last Month

Source: Towards Data Science

Introduction

as the years before: fireworks across the globe. People greeted the new year with new resolutions and new goals. Someone, somewhere, surely said: “2026 is going to be THE year.”

Then January happened.

As the weeks went on, the end of the month came closer. And for people in ML research, late January has a very specific flavor: it’s deadline season. The ICML rolls around, and suddenly the calm end‑of‑year downtime—when offices were empty and inboxes were quiet for long times—turns into 110 % energy.

That’s a stark contrast: from slow days to sprint days. But perhaps that contrast is not a flaw, but part of the rhythm? After a proper recharge, people can come back stronger, full of energy.

Looking back (delayed) on January, I found three themes that belong together: deadlines, downtimes, and flow times. The first two are obvious. The third falls in between the two and consists of long stretches of focused work; where something is just challenging enough to require concentration, and you can stay with it for hours.

Deadlines

Deadlines come in all sizes, and in all areas.

In private life, it might be the deadline for an insurance claim, an interest payment, or some official letter that you really should not ignore for too long. In work life, it’s project milestones, feature releases, and — for researchers — paper deadlines.

If the last minute didn’t exist, nothing would ever get finished.

We all know the downsides of stress, and chronic stress is bad. Over the years I’ve come to appreciate that a pointed dose of stress, over a limited time, can be good. For me, those deadlines are essentially a mechanism to create that dose. Suddenly, everything else pales in comparison because this one thing must be done—now.

That’s also why, rather strangely, I often enjoy getting closer to deadlines. Not because I like the panic, but because I like the clarity. A deadline creates priorities in a way that normal days don’t.

In daily ML work, things often feel more continuous: experiments run on the cluster, pipelines are adjusted, bugs are fixed. No big dramatic finish line, only minor mishaps. But once in a while, the pipeline has to ship, a feature needs to be added, an evaluation needs to be stable. Then, briefly, the whole team locks in—less chatter, fewer side quests, more alignment. Deadlines can be stressful, but they give you focus.

If you don’t have deadlines in your life, create small ones: a weekly internal deadline for a prototype, a Friday cutoff for an ablation set, or any checkpoint that forces clarity without burning you out.

Downtimes

After the deadline is before the deadline. Yes, yes. But first, a downtime.

After a stressful phase, it is genuinely good to do nothing for a while—or at least to do less, or to do things slower. After a stressful year, it’s good to take an extended break and recharge the batteries.

If you track time at work and collect overtime, this is the perfect moment to use it. If you don’t track time, taking one paid day off—or two—to balance the intensity is still a good idea. Leaving work earlier isn’t laziness; it’s making sure you can do the things you enjoy for the long term.

I used to underestimate downtime because it can feel unproductive—you don’t do anything. But that’s the point: downtime is productive, just in a different currency called future readiness. It restores your ability to focus later and prevents the slow decline where attention wanes, patience shortens, and you need more effort for the same output.

Plan downtime like work time. Put it on the calendar, especially after intense stretches. If you “wait until you feel like it,” you might never feel like you’ve earned it.

Flow times

These three ideas are connected.

After the deadline, you have earned your downtime. Once recovered, you return with new energy for new projects—helping you reach the next deadline. The interesting part is what happens between downtime and deadline: the flow time.

Flow time is when you’re working on something that is just challenging enough to require real concentration, and you stay with it long enough that your mind fully enters the task. Good work happens here.

Flow can look like:

• Implementing a feature (adding attention mechanisms, handling missing values properly, getting evaluation right, integrating an “agentic” principle carefully rather than slapping it on).
• Moving a whole project forward, such as the steady march toward a submission deadline.

Either way, flow requires one thing that modern work life often attacks: uninterrupted time.

The concept of flow was coined by Mihály Csíkszentmihályi, describing the sweet spot of high engagement, high focus, and low self‑consciousness. You don’t “try” to concentrate; concentration happens as a side effect of being in the right zone.

When you’re in the flow, things get done.

Protect flow time explicitly:

• Schedule a regularly recurring blocker for a few hours.
• Turn off messaging notifications (or close applications entirely).
• Make one task the only task.

Even one or two flow blocks per week can dramatically increase productivity and fulfillment.

Closing thoughts

January reminded me that a good work rhythm is not about always pushing hard. It’s about working in cycles:

• Deadlines – create focus and priorities.
• Downtimes – restore energy and prevent burnout.
• Flow times – enable meaningful, high‑quality work.

For the rest of the year, I’m treating this as a deliberate loop: earn the downtime, use the downtime, then invest in flow—and let deadlines do what they’re meant to do: bring things to completion.