2025 is a wild time. More people from academia are heading into industry: professors do sabbaticals, senior PhD students graduate early, and junior PhD students drop out. On a recent visit to my office at Berkeley, I spoke with several second- and third-year PhD students. Shockingly, more than half1 of them were considering either dropping out or finishing early2, to join startups like OpenAI, Anthropic, or Cursor. This situation confuses incoming and early PhD students: why continue a five-year PhD program when so many senior students are heading into industry ASAP?
This blog will present (1) well-known reasons why industry is compelling, (2) common but misguided arguments for doing a PhD, and then (3) share a few reasons for why PhD training is still important (with caveats).
Spoiler: I am still passionate about faculty positions3, and the 2019 version of myself would still start and finish a PhD today.
Disclaimer: By mentioning the option of dropping out, I do not necessarily endorse it—but rather sharing relevant information. Talk to your advisor to hear opposite arguments. I believe in the readers’ ability to aggregate information and make their own decisions. Of course, this blog is biased by my experience, so treat this as just one datapoint.
Well-known reasons for joining the industry
Resources: Industry offers better compute, data (e.g., user feedback), and infrastructures (e.g. RL code that uses 1K GPUs).
Direct impact: Industry research is more likely to be impactful through products.
Compensation: Total compensation for a junior researcher can range from $400k to $1M, or more.
Common but misguided reasons for doing a PhD
There are a few common reasons academic researchers cite when advising students against going into industry. However, many of these reasons are misguided.
Myth 1: industry research is more applied and less innovative.
One common stereotype is that industry simply applies existing ideas from academia to improve its products, rather than innovating independently. In reality, however, industry has driven many of the major paradigm shifts in language models. Examples include scaling laws (OpenAI), in-context learning (OpenAI), CoT (Google), Constitutional AI (Anthropic), and more.
Caveats:
The degree of innovation varies widely across companies and roles. Make sure to understand your actual responsibilities before accepting an offer.
Many paradigm shifts are still led by researchers with PhDs — though they tend to happen in industry after the PhD, not during it.
Myth 2: “Industry focuses on short-term impact, while academia optimizes long term impact. ”
In reality, some industry labs are very forward-looking. For example, the (super)alignment teams from OpenAI/Anthropic work on mitigating risks of superhuman AI systems, which are likely more than a year away.4
Caveat: Many industry longer-term programs/startups are unstable, e.g. the OpenAI super-alignment team has vanished. One might need to move between companies frequently.
Myth 3: “Finishing a PhD is the most effective strategy to become a research scientist in industry.”
Many companies have their AI residency program (e.g. Google/Meta) or other sources of recruiting talents (e.g. Anthropic’s alignment team recruits from MATS). Finally, you can always just apply, since the PhD degree is not a hard requirement. Finishing a PhD is not the shortest path for many industry positions.
(P.S. Many industry research leads do not have a PhD in CS)
Myth 4: “Industry does not care about compute-efficient methods”.
Many academics claim this because “industry researchers have a lot of compute already so they don’t have the incentive to be more efficient”. In practice, however, they try extremely hard to improve efficiency (e.g. LoRA, GQA, better scaling laws, etc). After all, cheaper training/inference is directly tied to their profit!5
Why PhD training is still valuable (for me)
Still, there are many upsides of PhD training. I will list a few below, but some come with important caveats.
Building my brand and taking credits for my own work.
As a PhD student, I generally have full freedom to publish my research. These publications can help me establish my reputation and allow others (e.g. hiring managers or faculty search committee) to evaluate me.
It is harder to build a personal brand in industry: papers or reports might have hundreds of authors, making credit assignment unclear. They may not even be published at all due to competitive pressure.
Cultivating soft skills and being well-rounded.
PhD programs also cultivate soft skills in writing papers, giving presentations for external audiences, or quickly introducing your research agenda at a conference. These skills are often less needed/emphasized if you are an individual contributor in a large research team.
Learning in a safe environment.
PhD programs are, at their core, educational. They provide advising resources and the opportunities to explore with lower risk. For example, I can regularly pursue random ideas outside my expertise (e.g., the impact of AI on the labor economy) without worrying about delivering results. In contrast, in industry, a failed project might impact performance evaluations or lead to concerns about “not meeting expectations.”
Caveat: That said, LLM research does not have a steep learning curve, so 5 years of PhD training can sometimes be overkill. In many cases, 2-4 years are indeed sufficient to learn independence and publish representative works.
Freedom to explore undervalued directions.
Sometimes I have an idea I strongly believe in, but no one in industry is pursuing it. In such cases, I need academic freedom to explore it.
Caveat: Many directions don’t actually require academic freedom to pursue. Three years ago, some academic researchers assumed that industry cared only about scaling, neglecting directions like retrieval, personalization, or coding agents. In hindsight, these predictions were flawed because they were based on visible products (e.g., ChatGPT 3.5) and assumed no further innovation.
Rule of Thumb: An idea will probably be tried in industry if some of my labmates can come up with it too, and most of my peers think it’s likely to work well.
Alignment with the public interest.
Some AI research has broader societal impact, shaping public discussions about AI and even influencing policies. For example, in discussions around SB-1047 (a bill that involves regulating open-weight models), some suspect that closed-weight companies are exaggerating misuse risks of LLMs, while companies relying on open-weight models are downplaying them. Academia can contribute significantly by conducting research independent of these corporate interests.
Caveat: Beyond academia, many other organizations operate independently: e.g., UK-AISI, US-AISI, Transluce, and METR. While unconventional, they are sometimes more impactful than academia because they have stronger focus.
There are many other reasons to be a PhD student, such as supporting open-science, simply enjoying the PhD lifestyle, or becoming a professor, etc. I will not list all of them here.
Closing Thoughts
The option of industry research is growing increasingly compelling. By discussing the idea of dropping out or not pursuing a PhD, I do not necessarily endorse it—but rather sharing relevant information. No matter what you choose, I hope you are confident with your decision —knowing that the grass isn’t greener on the other side.
Acknowledgement
I would like to thank Jiayi Pan, Nicholas Tomlin, Zhiyuan Zeng, Lisa Dunlap, and Jiaxin Wen for their useful feedback.
sample size = 9, probably not representative of Berkeley AI Research as a whole.
Graduating in 3-4 years, while the standard is 5 years.
I don’t know whether I can get the faculty positions though
We don’t need to agree on whether their agenda is reasonable, but they are forward-looking.
Overheard some conversations from my industry friend: “we should hire this PhD because they re-invented a technique we have invented ~3 years ago that we did not publish.”
Thanks for the great write up! It does seem like this is specific to a very small set of labs. The number of students/professionals wanting to get into AI research is much higher than the number of jobs available at labs like OpenAI. In addition to this, most Big Tech companies that hire for research generally have PhD requirements. Doesn’t this mean that a PhD would still be useful in 2025 even if you want to pursue research in the industry?