A Dozen Top Conference Papers and Still Struggling in Fall Recruiting? What Do Companies Look For?

I've talked with quite a few new graduates lately, and I've been struck by how sharply divided this year's job market feels.

Companies are offering sky-high salaries to compete for a tiny pool of candidates, because people who fit what they need are so hard to find. Meanwhile, many graduates are having a miserable time looking for work. I've met candidates with dazzling résumés and a dozen or more papers at conferences in the China Computer Federation's top tier (CCF-A) who still couldn't get through an interview. Others came in with plenty of internships on their résumés and struggled once the fall recruiting season got underway.

Why such a divide?

I think the rise of AI is rapidly eroding the value of what I'd call “instrumental rationality,” the ability to apply a set of techniques to a goal someone has already defined.

At university, students often learn concrete skills such as writing code, tuning parameters, and running experiments. Today, if those skills are your whole pitch, the advantage they give you in the job market will drop sharply.

So how can new graduates make themselves competitive in the AI era? And what do we, as employers, really need from the people we hire?

PART 01 | When AI takes over more of the hands-on work, what do you bring?

I see three levels of work in the day-to-day running of a company.

• Setting goals means understanding where things stand and deciding what problem to tackle next.
• Making things happen means bringing together the people and resources you need, clearing obstacles, and seeing a complex goal through.
• Doing the hands-on work means completing a concrete task, such as writing a feature and delivering it yourself from start to finish.

For a long time, doing hands-on work quickly and well was what most technical people could count on. AI is already taking over a substantial share of that work. As AI's efficiency at writing code and running routine experiments improves exponentially, the value of relying on those tasks alone will keep falling.

To get ahead in the AI era, graduates need to build on strong hands-on skills and learn to set goals and make things happen.

PART 02 | Setting goals takes insight and the nerve to aim high

One of the first difficulties many graduates face at work is deciding what to do without an adviser or a more senior student to guide them. Some can't see where to start. Others have a sense of what needs doing, but the problem looks too hard or they're afraid of failing.

That's also how someone can have a dozen top conference papers and still struggle in an interview. In school, your adviser often gives you the research question and your job is to solve it. In a company, there is no exam paper waiting for you. Finding the problem is as important as solving it.

Setting a good goal takes two things: the insight to see what matters and the nerve to aim high and follow through.

Start by finding a real problem that's hard and worth solving. That takes a clear view of where the industry stands, what's holding the technology back, and where things are headed. Many students get used to making incremental improvements within someone else's framework or chasing state-of-the-art benchmark scores. How often do they ask where the technology is really stuck or what need is most urgent for the business? Insight lets you see past the noise and identify a problem with lasting value on your own, before you spend your time competing over marginal gains.

Once you see the direction, you still have to aim high and act on it. The bar you set shapes every plan and decision that follows, and it puts a ceiling on what you can achieve. At our model lab, for example, if we set out to build the best of the second-tier models, every plan, compromise, and decision would take that tier as its reference point. That's probably where we'd end up.

If we aim to build a model that ranks among the best in the world, every decision forces us to ask what it would take to get there. The odds may be daunting, but aiming there gives us a real shot at reaching that level.

Going from good to great takes judgment and ambition together. You need the insight to choose a difficult goal worth pursuing and the confidence to go after it.

PART 03 | Making things happen means going beyond what you can do alone

Completing a task and delivering a complex outcome are different skills.

You can usually write a piece of code yourself. Training a world-leading model in a particular domain at our lab takes a much broader effort. Along with deep technical skills in areas such as algorithms, you need real support from the infrastructure and data engineering teams.

In the AI era, people in execution roles need qualities that used to be expected of leaders. Four abilities matter here: managing work, collaborating, getting past roadblocks, and understanding what causes the outcome you want.

1. Manage AI agents

When you start a job today, you have AI agents available around the clock. How do you break up the work and design a workflow so they can complete well-defined tasks as efficiently as possible? That's a basic management skill now.

2. Plan ahead when you need another team's help

Suppose the data team has to build a whole new data collection pipeline for you. How do you get their support when the request will take considerable effort?

• Showing up and saying, “I need this pipeline, ideally in three days,” is effectively asking them to drop everything and make your work their priority. That's rarely going to happen.
• If you see the need a month in advance, you can talk through the business value and agree on a timeline with the team. If necessary, you can get both managers to agree on why the work matters. That gives the data team room to plan and makes sure their manager recognizes what they've delivered. You're much more likely to get the support you need.

3. Take responsibility for getting past roadblocks

Here's an example from the day-to-day work of training models. Say you need to finish a critical experiment as soon as possible, and a GPU driver crashes in the middle of the night, bringing the run to a halt.

People react very differently. Some think, “That's the infrastructure team's problem,” and go to sleep. Others report the issue and plan to restart the run the next day once it's fixed. Some simply wait.

Someone who takes responsibility for the result starts by checking the infrastructure team's on-call schedule and contacting whoever is on duty. If they can fix it that night, the experiment can resume immediately. If it's a thornier problem, they look for a way to keep going on another machine. What if every GPU in the cluster is occupied? They might use AI to quickly review which running experiments are less urgent, then ask a colleague if they can borrow those GPUs long enough to finish the critical run. They can take the colleague out for a good meal afterward to say thanks.

Getting the run back on track calls for communication across teams, flexibility, and persistence. That's the kind of initiative AI can't replace.

4. Work backward from the business outcome

Every outcome has causes. Start with the result you want, work backward through everything that needs to happen first, and turn those prerequisites into intermediate results you can deliver one by one.

Then test whether each link in that chain is realistic. As we saw earlier, asking people to drop their current work and come to your rescue usually won't get you very far. If the situation is truly urgent, look for what would make it possible. A manager might need to explicitly make the work the top priority. Securing that decision is how you get the support the result depends on.

Take it one step further. After we train a leading model, we still need people in the industry to hear about it and actually use it. Understanding what leads to adoption requires thinking about the market and the competition as well as the technology.

• What's the main objective of this release? Do we want to reach a broad audience, or is recognition from the field's top experts enough?
• Where do our competitors stand at the time of launch, and how should that shape the way we announce it?
• How might publishing a paper or releasing model weights change the competitive landscape across the industry?

Questions like these are how technical people learn to shape the outcome beyond their own piece of the work. If you have big ambitions, start practicing that broader way of thinking now.

PART 04 | Stay steady under uncertainty

I want to end with a personal quality. Alongside the ambition to act and the insight to find a hard problem worth solving, there's something I value a great deal and keep hearing about from industry leaders. It's resilience, the ability to stay steady under pressure and uncertainty.

The AI era will be turbulent and hard to predict. Our aversion to uncertainty is ingrained. Under sudden change and intense pressure, it's natural to get anxious, lose our composure, make hasty choices, or want to escape the situation altogether.

The person who can withstand the pressure, give their emotions a little distance, and look at the situation calmly is in the best position to make a good decision with what they know.

Where does that composure come from? Part of it comes from reflecting on what matters to you and having a place to stand when things get rough. As The Great Learning puts it, “Knowing where to stop, one can become settled. Once settled, one can be still.” It also comes from deliberately developing perspective and generosity of spirit through experience. Once you can see where the longer journey leads, today's setback starts to look much smaller.

As Nietzsche put it, “What does not kill me makes me stronger.”

What AI makes obsolete is the mindset of treating yourself as a tool. In this period of enormous technological change, I hope everyone about to enter the workforce builds strong technical skills and becomes someone who can see where to go, set ambitious goals, hold steady under pressure, and get things done.