How credible is it that some algorithm will be able to write code at e4 or even e5 level? That will probably happen. How credible is it that some algorithm will replace e4 engineer in the foreseeable future? Very very unlikely.\
\
Current algorithms are unable to properly reason, they are unable to solve novel problems, they are unable to prioritize. There is nothing yet in the works that overcomes those... not sure deep learning as it is currently is able to do so. 

Engineering is not about being a coding monkey but about solving abstract problems where sometimes the real solution is "do nothing".

Silicon Valley leaders have an agenda. What they say about "AI" needs to be looked at through that filter.

Guiding ML Engineers through their career journey. ex-Head of ML, ex-Meta Staff SWE, ex-Adobe

Ilya Reznik

Congratulations on your offer! I agree with Rahul's assessment. While every organization has its strengths and limitations, here's my perspective on working in Ads at Meta:

The workload in Ads tends to be substantial, with most people working 60+ hours weekly, though some smaller teams may have lighter schedules. However, this intensity comes with notable benefits: you'll work alongside exceptional colleagues, and promotion opportunities are more plentiful (up to E7). During the 2022 major layoffs we lost only one person out of 30, which was significantly lower than other orgs, making Ads a relatively stable place to work.

As an E4, your primary focus should be advancing to E5, since Meta is up-or-out system at that level. Keep in mind that pursuing promotion typically requires more effort than maintaining your current level. 80 hours for a few months is expected.

One bonus piece of advice regarding PSC (Meta's evaluation process): while it's important for you to understand how it works, it's even more critical that your manager does.

Based on my experience, I would strongly recommend against joining any team where the manager has less than 12 months of tenure at Meta. Even with the best intentions, newer managers simply lack the deep understanding of PSC's nuances that comes with time at the company.

Historically, many SWEs came from Data Science. They had a pretty steep learning curve to pick up coding. You are going to have coding down and have a steeper learning curve learning ML, honestly **yours is the easier path**.\
\
I am working on a course about transitioning to MLE career that should be available on Taro in Q1 2025, but the answer to your question about Masters is: usually no.

I have onboarded hundreds of MLEs, the ones coming out of even the best masters (and PhD) programs cannot do anything real world requires of them. It helps get your foundations right, but honestly **the industry evolves fast enough to where it is hard for traditional schooling to keep up**. 

If you don't have a masters, it may help you clear the education requirement, but if you have one in another field there is no reason for it. 

**Best MLEs learn by building** and use papers/courses/chatgpt/forums/... to fill in the gaps. I can explain overfitting 50 times, but you will never be a good MLE until you overfit a model and see the results.

Other than knowledge, the main paradigm shift for MLEs is that *there is no unit test*. There is a probabilistic assessment that your system will work under certain conditions. 

Also you live with near-certainty that something better will come out tomorrow and the kids coming into the field will know the relevant tech better than you, so learning is not an afterthought but a main feature. But **you have to balance true learning with jumping on every hype bandwagon**. Fun, right?

Some excellent answers here already.\
\
I'll weigh in as "AI" practitioner. 

AI is a marketing term that is not precise enough for what we need to answer your question, but I will roll with it so this doesn't turn into a 10 page essay.

Currently "AI" is mostly applied to the transformer-based models. These are completely devoid of true reasoning potential or ability to make complex connections. Those are major roles for engineers, so you are not competing with them and any company would be foolish to cut engineers because they have a more advanced auto-complete. 

Adding reasoning to these models is something that we are actively working on as an industry, so some day it will be solved. Working "with AI" will be a very important skill. 

To do that effectively we need two elements:

1. AI should be built with understanding of humans
2. You need to understand how AI works

Unless you are an ML researcher of engineer you really can't change much about 1, but 2 is within your control. 

So to your question of should you be integrating AI into your workflows, the answer is *absolutely yes*. 

This is a great question.\
\
The first thing to ask yourself is **why are you interested in ML**? From your question it sounds like income potential is the biggest driver. In this case I would advise you not to make the switch for several reasons:

1. ML salaries are actually not much higher than comparable backend SWE levels. There are some outliers and there are more opportunities in ML today, but your career is a 30 year proposition and most of these are temporary. Full stack SWEs  make about 15% less for similar experience (I'm not sure what the delta with BE is, but it's smaller) and even then it is not worth it...
2. As someone who is already knocking on the door of E5 promo, you built up a lot of capital in your area. Going to ML is going to be challenging and likely come with a step back in your career.
3. ML is a very hard area to be in, especially if you are trying to be good. The main paradigm shifts, best practices are only now being written, you are in the public eye without having best practices... There is always some demo (which was likely faked) that your work (which is real) gets compared to.

This doesn't mean that people should not be going into ML careers, I actually want to encourage more engineers to do so. I'm currently working on a course for Taro about how to make that transition that should be out in Q1 of 2025, but the reason why you are doing it is important and compensation is the least compelling one of all.

Just to quickly add to this:\
\
If you are driving the interview you can pick where to show tradeoffs. For example, in modern models with multiple modalities (text and image) early vs late fusion between modalities is a tradeoff that comes up often. Here you can demonstrate both:

- The technical signal (modularity and flexibility of late fusion vs better cross-modal representations of early fusion) and
- The leadership component (which complexity is harder for the team to manage, which will depend on the problem here)

The other quick clarification is level signaling and for this you need to understand the roles:

- Jr Engineer's job is to do the work they are provided. This means you would naturally have more questions, need more guidance, and will likely propose one idea that you think to be the best instead of discussing tradeoffs of multiple approaches.
- Staff Engineer's job is to "find scope": determine what is worth working on and how to set up the team for success. If you spend a lot of time on things that matter very little, that's a negative signal at this level. Knowing what to prioritize and not being stuck on one solution is key here.

Head of Machine Learning, ex-Meta Staff SWE, ex-Adobe

Great points, thanks for sharing!\
\
As someone who helps a couple dozen MLEs every month, I have some additional insights I can share:

- To your point #5, I find that those interviews are usually harder and much more variable. Big Tech has very standardized interviews with people who are (usually) really knowledgable about ML. 
- Depth and breadth both matter a lot. The **best way to show depth is by talking about tradeoffs**. I was a staff MLE at Meta and my whole job was to understand tradeoffs of various ML and data techniques. 
- Relative importance of various rounds depends a lot on your level with behavioral and system design becoming much more important at higher levels. You still need to do DSA/ML Fundamentals, but they are less important. Having said that, failing any round will result in not getting an offer.

While the patterns of interviews are different among different companies here is what you will generally be asked for MLE:

- **DSA** (very common everywhere)
- **ML System Design** (very common everywhere)
- **Behavioral** (very common everywhere. This can take form of "Project Deep Dive" too, think of that as a technically-focused behavioral, and like you mentioned know the why behind every choice)
- **ML Fundamentals** (very common at smaller companies)
- **ML Coding** (starting to become common in tier 2-3 like Cruise and Reddit)
- **Distributed system design** (becoming less and less common for ML roles)

To end on a hopeful note, I am seeing a lot more hiring since October from multiple companies for ML roles.