TL;DR
- There are three distinct AI research careers: industry research scientist, academic researcher, and research engineer. They share a name and almost nothing else.
- PhD: required for the academic track, strongly preferred for industry research scientist roles, not required for research engineer positions.
- Google DeepMind research scientist base salary is $210K-$264K+ (Rora / Glassdoor data). Research engineer roles at top labs pay comparably. Academic researchers typically earn around $100-120K.
- Path selection determines your entire prep stack. Building publication-track credentials for the research engineer path wastes years.
- Research engineer is the most accessible entry point and the most ignored by career guides.
Is AI research right for you?
AI research is right for people who want to spend most of their time generating new knowledge - designing experiments, writing papers, and moving the field forward - not building and shipping products. If your main goal is to deploy AI systems, build inference pipelines, or ship product features, you're probably looking at ML engineering, applied science, or data science. A data science mentor can help you map that path if you're closer to that end of the spectrum. Those careers have more open roles, faster feedback cycles, and no publication requirement.
When people come to us for AI and ML help, a pattern I keep seeing is someone who wants a research role but hasn't confirmed which version of it they mean. Or whether they want research at all. The ones who stall fastest are the ones who started building credentials for a path they hadn't chosen yet.
I've watched hundreds of career transitions through MentorCruise. The successful ones follow a pattern: they start with internal clarity - what do I actually want? - then move to skill mapping, and only then go external with networking and applications. Most people start with step three and wonder why they're stuck. For AI research, that internal clarity question is sharper than almost anywhere else, because the three paths inside "AI researcher" require such different preparation.
Here's a quick contrast to orient you before the detailed breakdown:
| AI research | ML engineering | Applied science / data science | |
|---|---|---|---|
| Primary output | Papers, new knowledge | Production systems | Models in products or analytical insights |
| Publication required? | Yes (academic/RS track) | No | Sometimes |
| Typical hiring bar | PhD + publications (RS) or strong SWE + ML (RE) | SWE + ML deployment experience | ML + product or domain fluency |
One path gets clearer from here. If the research column is where you see yourself, keep reading.
What AI researchers actually do
"AI research" covers three jobs that look almost nothing alike. An industry research scientist at DeepMind or Meta AI designs novel experiments and publishes papers while working in an engineering-rich environment. An academic researcher builds toward a faculty career with full autonomy over research direction. A research engineer builds the infrastructure and experimental tools that let researchers do the science - no papers required.
Here's how those three paths compare across the dimensions that matter for planning:
| Industry research scientist | Academic researcher | Research engineer | |
|---|---|---|---|
| PhD required? | Strongly preferred; rare exceptions exist | Required | Not required |
| Publications required? | Yes - NeurIPS/ICML/ICLR track record | Yes - venue strategy determines career | No |
| Day-to-day work | Design experiments, write papers, collaborate with engineers on implementation | Write papers, manage students, teach, review grants | Build training infrastructure, implement and scale novel models, maintain experiment tooling |
| Entry-level comp (top labs) | $210K-$264K+ base (Google DeepMind, Rora / Glassdoor) | \~$50-80K postdoc, \~$100-120K faculty | Comparable to RS at top labs |
| Path variance | Medium - PhD is the modal path, but verified non-PhD cases exist | High - advisor quality and venue placement matter enormously | Lower - clear SWE + ML fluency bar |
A Tuesday for an industry research scientist at a top lab might look like: reviewing a collaborator's experiment code in the morning, writing the methods section of a paper in the afternoon, and attending a lab meeting to discuss a new direction. The loop is experiment, analyze, write, and repeat.
For an academic researcher, Tuesday is closer to: meeting with a PhD student to review their experiments, three hours of writing, reviewing a paper submission for a conference, and fielding email about a grant deadline. The output that counts is the paper and the student - not code shipped.
For a research engineer, Tuesday is: debugging a distributed training job that failed halfway through overnight, reviewing a paper to understand what model architecture they need to implement next, writing unit tests for a new experiment pipeline. No papers on the to-do list.
Research scientist vs research engineer: what actually separates them
Research scientist and research engineer are two distinct roles at AI labs, not different names for the same job. The research scientist's output is knowledge - papers, proofs, trained models that advance the state of the art. The research engineer's output is infrastructure - the systems, pipelines, and tooling that let researchers run experiments at scale. One requires a publication record; the other requires strong engineering. They pay roughly the same at top labs.
The question that separates them for a career changer is not "which sounds more impressive?" It's "do I want to spend my time generating knowledge or building systems?" Both are technical. Both are at the frontier. But the credentials you need, the interview you prepare for, and the day you'll have at work are entirely different.
How to transition into AI researcher
The transition into AI research starts with path selection, not skill building. If you don't know whether you're targeting industry researcher, academic researcher, or research engineer, you don't have a roadmap - you have a reading list. Pick the path; the skills follow from it.
Transitioning into an industry research scientist role
Getting hired as a research scientist at DeepMind, Meta AI, or Google Research requires a publication record at top venues - NeurIPS, ICML, ICLR. That record usually comes from a PhD program, but a strong independent paper can open the same door in rare cases. The path is: strong ML foundations, then a PhD program or independent lab equivalent, then a first publication, then a competitive application.
"Strongly preferred" for PhD means: the exceptions exist, but they're rare enough that you should plan for the PhD unless you can realistically build a publication record through independent research. Competitive research scientist roles get hundreds of applicants per opening. The differentiator is not skills self-assessment - it's your publication record and the venues where your work appeared.
If you're not interested in producing papers - the actual work of generating novel research findings - this path isn't for you. The research scientist title at a top lab is fundamentally a publication-record credential.
David Fan got into Meta FAIR Research Engineer without a PhD via a single strong CVPR paper from undergraduate independent research. His account is published at davidfan.io. That case is real, but treat it as what it is: a verified exception in a field where the modal hire has a PhD from a strong program and multiple top-venue papers.
Milestone checkpoints:
- Milestone A: NeurIPS/ICML/ICLR submission accepted. Observable: acceptance notification.
- Milestone B: Research-adjacent role at a strong ML org - applied scientist or ML engineer. Observable: offer letter.
- Milestone C: First-author paper at a top venue. Observable: arxiv submission + acceptance email.
Transitioning into academic AI research
Academic AI research is a real career choice on its own terms - not what you do when you couldn't get an industry job. The trade is clear: lower pay, more autonomy, the ability to pursue decade-long questions companies won't fund. The path is defined by PhD program quality, advisor match, and venue strategy. Get the advisor wrong and the path gets much harder.
Advisor quality matters more than program ranking. A strong advisor at a second-tier program, actively publishing at NeurIPS, mentors students into faculty and national lab positions. A weak advisor at a top-five program leaves you with a PhD and a thin publication record. Ask which specific advisor you'd work with and look at where their recent graduates landed.
The venue strategy is simple: NeurIPS, ICML, and ICLR are the venues that move careers. A workshop paper at one of these is a step. A spotlight or oral at any of them is a signal. Most faculty job searches require at least two or three publications at these venues before you're competitive.
Postdoc is the mandatory step for the faculty track. Some national lab positions (Argonne, NREL, Sandia) hire directly from the PhD without a postdoc, but the faculty path almost always goes through one.
Compensation: academic median sits around $100-120K for faculty and senior researcher positions. Postdoc pay is typically $50-80K. The trade is autonomy and the freedom to pursue questions industry won't fund.
Andreas Madsen published an ICLR spotlight paper as an independent researcher without university affiliation before joining Mila. His account is on Medium - read it here. The path exists without institutional scaffolding, but it's harder and requires sustained self-direction.
If your primary concern is timeline certainty or compensation, academic research is the longest and most variable path. PhD 4-6 years, postdoc 2-3 years, then the faculty job market. Budget 8-10 years from starting a PhD to a stable senior position.
Milestone checkpoints:
- Milestone A: PhD applications submitted to 5+ programs with identified potential advisors. Observable: confirmation receipts, advisor contact logged.
- Milestone B: First co-authored paper submitted. Observable: arxiv ID.
- Milestone C: Acceptance at a named venue - NeurIPS/ICML/ICLR or workshop. Observable: acceptance notification.
- Milestone D: Postdoc offer aligned with your research direction. Observable: offer letter.
Transitioning into a research engineer role
Research engineer is the path the career guides forget. At DeepMind, Meta FAIR, and Google Research, research engineers build the systems that let research scientists run their experiments - and they earn comparable salaries without needing a single publication. The entry bar is strong software engineering plus deep ML fluency, not a publication record. A deep learning mentor can help you build the ML depth side of that bar if your SWE foundations are already solid.
RE roles require strong SWE foundations, deep ML fluency - the level where you can implement a paper from scratch - and systems thinking. Publications, a PhD, and a venue track record aren't on the list.
The salary point is worth stating clearly: at top labs, research engineers earn compensation comparable to research scientists. This is the information most career guides miss. If you want to work at the frontier of AI research and you have strong engineering foundations, the research engineer path is not a consolation prize - it's a direct route.
David Fan's case is relevant here too. He entered Meta FAIR as a Research Engineer, not a Research Scientist - that distinction matters for this section. He had a CVPR paper from undergraduate independent research, which is unusual, but the core credential for RE roles is engineering ability plus ML depth.
Note for readers with strong software engineering backgrounds: if you already have 3+ years of SWE experience and ML fluency, the research engineer path is likely your fastest realistic route into a top AI lab. You're building on existing strengths rather than starting a PhD from scratch.
Milestone checkpoints:
- Milestone A: ML foundations solid - can implement a transformer from scratch, can reproduce a NeurIPS paper result in PyTorch. Observable: working code on GitHub.
- Milestone B: Open-source ML contribution with meaningful adoption or a published independent paper replication. Observable: GitHub stars, merged PR to a known project.
- Milestone C: RE application submitted to at least one top lab. Observable: application logged, recruiter response.
Common roadblocks (and how to get past them)
The most common stall I see in people trying to break into AI research isn't a skill gap - it's a feedback gap. You don't know if your work is research-grade because you're not in a research environment yet. The fix is getting into an environment where research-grade work is evaluated: open-source ML projects, workshop submissions, or a mentor who can assess where you actually are.
"I don't know if my work is good enough to submit"
The only way to know if your work is submission-quality is to be around people who submit regularly - a research group, a workshop track, or a mentor with a track record in your target venues. The feeling that "it's not ready yet" is almost always feedback deprivation, not skill deprivation. Submit to a workshop. Get rejected. Learn from the reviews.
The milestone here is specific: submit to a workshop or lower-tier venue before targeting NeurIPS/ICML/ICLR. Observable: submission ID, reviewer feedback received. Reviewer feedback is the best data you'll get on whether the work is research-grade.
"I don't have a PhD or a university affiliation"
Independent research is harder but structurally possible. Andreas Madsen published an ICLR spotlight paper without university affiliation and joined Mila. David Fan got into Meta FAIR without a PhD via a strong CVPR paper from undergraduate independent research. Both are verified cases, not aspirational. The path exists - but it requires sustained self-direction and work that holds up to rigorous peer review.
If you're not willing to operate without feedback loops and institutional support for an extended period, independent research is not the right route. The cases that worked involved researchers who were extremely self-directed - no advisor, no lab, no guaranteed path to publication.
A few other common stalls worth naming:
If you've been out of ML work for 2+ years, the relevant benchmark isn't the size of the gap - it's whether you can still implement from papers. The verifiable milestone: reproduce a recent paper result from scratch.
For international readers: top US labs - DeepMind, Meta, Google, OpenAI - sponsor H-1B for research scientist and research engineer roles. It doesn't change the credential requirements, but it affects which labs are worth the application effort.
One pattern I keep seeing is using AI tools to learn faster but losing the depth along the way. In research, depth is the credential. You can't submit a paper you don't understand. Using Claude or ChatGPT to work through a concept is fine. Using it to shortcut the understanding itself means your paper will collapse at peer review.
Tools, mentors, and next steps
The fastest accelerant for an AI research transition isn't another course - it's feedback from someone already inside the field. A mentor who's made the academia-vs-industry decision themselves can assess whether your work is submission-quality, which labs are realistic given your background, and which of the three paths actually makes sense for your specific situation.
Resources that matter by path:
For industry research scientist and academic paths: Papers With Code and arxiv for staying current. NeurIPS/ICML/ICLR proceedings for understanding what publication-quality work looks like at those venues. Fast.ai's Practical Deep Learning course for ML foundations. Andrew Ng's deep learning specialization. ML reading groups - finding one locally or virtually - for getting feedback on your work from people who submit.
For research engineer: PyTorch documentation. Andrej Karpathy's nanoGPT on GitHub for building transformer intuition from scratch. Hugging Face contributor guides for understanding how to work on active ML projects. Open-source ML repositories where you can make meaningful contributions.
About 1 in 6 people who apply to MentorCruise are looking for help with AI and ML roles - it's our second-largest category. That's why we have researchers and engineers who've made exactly this decision on the platform. We accept fewer than 5% of mentor applicants, so the AI mentors on the platform have been through the hiring process at top labs or the academic path themselves.
If you're trying to figure out which of these three paths is realistic for your background - or how to move faster on the one you've chosen - a mentor who's already inside the field compresses that learning dramatically. You can try any mentor free for 7 days. Find an AI mentor.
FAQs
Do you need a PhD to become an AI researcher?
For two of the three paths, no. Academic research requires a PhD - it's the foundational credential. Industry research scientist roles at top labs strongly prefer one, though exceptional publication records from non-PhD contexts have opened doors in rare cases. Research engineer roles don't require a PhD at all and pay comparably to research scientist roles at top labs. The PhD question depends entirely on which kind of researcher you're trying to be.
How long does it take to become an AI researcher?
Academic path: 4-6 years for the PhD, then 2-3 years of postdoc before a stable faculty position - budget 8-10 years. Industry research scientist: PhD timeline plus 1-2 years of competitive publication building. Research engineer: the fastest route - 12-24 months of focused ML study plus open-source contributions if you already have strong software engineering foundations. The timeline is determined by which path you're on.
What does a research engineer do at an AI lab?
A research engineer builds the infrastructure and experimental systems that let research scientists run their work. At DeepMind or Meta FAIR, that means training infrastructure, distributed systems, experiment tracking, and sometimes implementing and scaling novel models from a paper. The output is infrastructure, not papers - but the work is closer to frontier AI research than most engineering roles outside an AI lab.
What salary do AI researchers earn at top labs?
Google DeepMind research scientist base salary is verified at $210K-$264K+ according to Rora and Glassdoor compensation data. Research engineer roles at top labs pay similarly. Academic researchers typically earn significantly less - around $100-120K at the median for faculty and senior researcher positions - in exchange for autonomy over research direction. The industry premium for top-lab positions vs academic is large: 50-100%+ for research scientist roles.
Can you become an AI researcher without a university affiliation?
Yes, with strong caveats. Andreas Madsen published an ICLR spotlight paper as an independent researcher without university affiliation before joining Mila. David Fan broke into Meta FAIR Research Engineer via a CVPR paper from undergraduate independent research. Both are verified cases. The path exists if you can produce research-grade work without institutional scaffolding. It's harder and higher-variance than the PhD route.
What's the difference between AI research and machine learning engineering?
Research generates new knowledge; ML engineering applies existing knowledge to build products. A research scientist asks "what's the most capable architecture for this task class?" and writes a paper. An ML engineer asks "how do I serve inference at under 100ms latency?" and ships a pipeline. If you want to build and deploy AI systems, ML engineering is the right path. If you want to extend what AI can do, research is.
