10 Key Questions on the PhD Path in the AI Era | A Conversation with 李磊: HKU PhD, Xiaomi MiMo Researcher

Summary

• A data company’s long-term value depends on its ability to keep filling new gaps in models. 李磊 expects demand to persist. If data companies cannot keep producing high-quality data, models’ own data-generation capabilities will drive costs very low and intensify competition. Pipelines that can identify capability gaps, organize experts and deliver at scale may see their advantage widen. When model companies buy data, they are also buying time to push the frontier of intelligence; as he puts it, “time is an irreversible resource” (“时间是不可逆的资源”).
• 李磊 expects rapid progress in RSI on well-defined tasks, but deciding what research to pursue remains the key gap. Better parameters, algorithms and synthetic data can all be found by scaling compute. But researchers still largely decide “which problems matter” and what to tackle next. He has yet to see an RSI benchmark that excites him or commands broad consensus.
• Xiaomi’s post-training livestream brought the training dynamics and costs of large-scale RL into public view. 李磊 says the team saw broad capability gains from large-scale RL early in training and wanted MiMo V2.6’s Pro and Flash models to give the industry confidence in RL. Chinese viewers focused on cost; overseas viewers paid more attention to transparency. The near-public TensorBoard also gave researchers a “gold mine” for analyzing training dynamics.
• In the hiring 李磊 has been involved in, papers have slipped from a key screening criterion to a “nice to have.” He pays more attention to whether code and data are open, whether issues are maintained, and whether anyone uses the model. A model claiming SOTA with only a few dozen downloads a month “might” be addressing a less important problem. Asked by Koji whether these metrics can be gamed, he agreed they can: evaluation systems always face reward hacking.
• As paper volume surges, research communication itself has become part of building influence. One conference has grown from a few hundred papers to 4,000–5,000, leaving readers increasingly reliant on recommendations and social media. 李磊 sees this as “attention is all you need” made visible. He has 24,000 Xiaohongshu followers, but his advice to PhD students is to be helpful, persistent and authentic—and to find people with similar research taste by publishing consistently.
• Whether a PhD is worth doing starts with what you want to use that time to figure out. 李磊 does not think a degree is a prerequisite for joining a big company: “He was good, so he got in.” What he values is a relatively complete research journey that leaves room for curiosity. When choosing a program, differences between advisors and research groups are far more worth investigating than the school’s Top 30 or Top 50 label.
• Research efficiency comes from choosing carefully, validating early and then moving fast on problems that are “hard but right.” 李磊 once spent more than half a year tuning an algorithm before discovering the underlying assumption was wrong; now he isolates a minimal hypothesis and tests it with a small experiment first. When Koji questioned whether hard problems necessarily mean slow progress, 李磊’s answer was that even a small advance on an important problem can have an outsized impact—the key is finding a reliable angle.

Deep dive

1. The value of a PhD depends on the questions you want to answer

• 李磊 is both a PhD student at HKU and a post-training researcher at Xiaomi. Two months into his program, in November 2023, he attended his first conference. He had a paper of his own, and a junior labmate’s paper won Best Paper. It felt like a “peak start”—and left him wondering whether he had peaked at the outset.
• What pushed him toward a PhD was imagining himself 10 years later, lying in bed and possibly regretting that he had missed the experience. He acknowledges that industry compensation raises the opportunity cost. By their final year, his classmates talked about packages as much as academia, but he still wants time to study what interests him.
• 李磊 sees no need to do a PhD just to join a big company; researchers such as Jason Wei show that “he was good, so he got in.” A different motivation is worth pursuing: having a relatively complete stretch of time to satisfy your curiosity, build knowledge, and develop research and communication skills. “People live 80 years”; 5 of them can be a distinctive journey.

2. Take your time choosing a problem, then validate and execute quickly

• 李磊 frames problem selection around difficulty and importance. If a problem is not difficult enough, there will be many competitors and it is easier to get beaten to the punch. If it matters enough, even making “a little bit of progress” can carry significant influence. The goal is to find an angle and offer a reliable partial explanation or experimental analysis.
• Koji’s follow-up is worth keeping: if hard problems take a long time, how do you get timely positive feedback? 李磊 says the two can coexist. When everyone knows a problem is hard, the bar for solving it completely is lower; a small, solid advance can become an important milestone for those who follow.
• He sees the “slow research” initiative launched by more than 100 professors and researchers as a call not to use publication as the only screening criterion. Students should build strong foundations and explain clearly what they did, why they did it and how. The point is not to stop doing research, but to balance coursework and research, focus on important work and make progress step by step.
• He once spent more than half a year tuning an algorithm, only to discover that the assumption was wrong. Now he first isolates a minimal hypothesis and tests it with a small experiment. Before settling on a direction, he reads papers intensively and talks to researchers; once the direction is set, he raises GPU utilization and speeds up data processing, using AI to move work forward in parallel. “Fast and slow are relative.”

3. Models’ “jagged” capabilities leave room for research

• 李磊 sees 3 layers of PhD anxiety: first meeting publication and graduation requirements, then worrying whether the work matters, and finally wondering whether the transition into the workforce will go smoothly. When systems such as Astra emerge, some researchers in embodied AI feel that their entire domain has been upended.
• His response is to return to the experiments: “Actually sit down at the computer,” write code, design a harness, run the data and see exactly how the model performs. Large models have “jagged” intelligence: they are exceptionally strong in some areas and entirely incapable in others. Social media can magnify visible achievements; the remaining questions need to be checked firsthand.
• Koji speculated that OpenAI’s continued purchases of embodied data might explain where some of these capabilities come from. 李磊 responded only that “we suspect that may be the case.” He says he is not an embodied AI expert, but questions about how capabilities generalize, and whether there is a gap between simulators and real-world systems, are still worth studying.

4. The paper system is under strain, and acceptance decisions cannot capture the full picture

• As a reviewer and Area Chair, 李磊 has seen his workload rise from 2–3 papers to about 10; deadline-driven judgments may therefore become more random. He notes that NeurIPS has 3 venues. The community is also trying to limit submissions, make senior reviewers more accountable and introduce AI assistance. For example, CHAIR-AI requires 2 senior SPCs to discuss a paper before making a decision.
• Koji asked about using an SOTA model to write 100 papers and relying on the odds to get some accepted. 李磊 called that a “DDoS attack” on the system. Papers are meant to share interesting research; the key question is, “Did the author actually figure this stuff out?” Is there experimental or theoretical support for the hypothesis?
• He leaves room for another possibility: if a human working with AI really comes up with 100 valuable questions and gives solid answers, it does not matter whether they are submitted for publication. But bulk production does not eliminate the cost of verification; confirming the results still takes substantial human effort.
• In candidate screening, a paper is now closer to a “nice to have”; open-source contributions can also be dressed up with AI-generated PRs. So he continues to check whether code and data are open and whether issues are maintained. “People shouldn’t treat acceptance or rejection as the final reward.”

5. Influence comes from real-world use—and requires active distribution

• 李磊 looks at Hugging Face downloads: if a model claims SOTA but had only a few dozen downloads last month, it might not be solving a very important problem. Koji asked whether download counts can be gamed. 李磊 agreed they can, but says the cost of downloads, storage and similar tactics makes manipulation more expensive.
• Academic distribution is changing too. ACL used to accept only a few hundred papers, which readers could filter by keyword or familiar professors. Today, a conference can have 4,000–5,000 papers; no one can read them all, so people increasingly rely on recommendations. Earning visibility in those systems can “amplify the value of your work.”
• Koji noted that he has 24,000 Xiaohongshu followers, and 李磊 has also shared posts with more than 10,000 citations. Social media’s immediate value, in his view, is fast feedback and a way to find peers with similar research taste. The “big-follower theory”—that followers and likes outnumbering citations—is “attention is all you need” made visible.
• His advice on building an audience is to be helpful, persistent and yourself. He posted paper notes consistently early on, with little growth; a later post about the relationship between citations and followers drove a jump. His notes on research such as On-Policy Distillation also reflect genuine interest. Controversial takes can attract comments, but they do not necessarily make people feel the content is connected to its author.

6. Doctoral training should progress from independent research to leadership and industry collaboration

• 李磊’s first stage is mastering the fundamentals of research. Even with AI assistance, it is worth implementing basic algorithms from scratch. He cites implementing SVMs with NumPy and MATLAB, and taking 吴恩达’s Coursera courses; concepts such as generalization and overfitting still serve him today.
• Independent research also means writing, designing experiments, justifying hypotheses and communicating knowledge. AI can act as a reviewer or teacher to help refine hypotheses; open source, blogs, software and sharing help maximize impact. Being able to produce a strong paper independently and build some influence is what it means to have run the full pipeline.
• The second stage is leadership: moving from “generator” to “critic,” teaching methods to more junior students and adjusting the level of guidance case by case. Self-motivated students may need only light direction; others need closer supervision. The third stage is entering industry, where competition is more practical and has more dimensions. He says he is still learning too.

7. Getting started depends on building a body of work and connecting with key nodes

• In the summer between his sophomore and junior years, 李磊 sent about 10 cold emails and got 1 reply. Professor 孙旭 gave him the chance to intern in his group, which led to top-conference work and a recommendation for graduate school. Professors’ inboxes are more crowded now, while compute and headcount constraints weigh on industry; proactive outreach still matters.
• With limited compute, you can build a GitHub project, a harness or an interesting application. For a research-oriented path, get the fundamentals straight, write a strong research blog and use small experiments to make the case. He points to 苏剑林’s early, derivation-focused blog posts and his later proposal of RoPE and move to a research role at Kimi as evidence that work can come before resources.
• No reply is common, but producing work is itself a learning process. Turn what you want to learn and create into a portfolio. He likens the process to “bubble sort”: keep building, and people will gradually notice. His own first internship at Kingsoft came after someone saw his blog.
• He recounts 2 findings from The Formula: The Universal Laws of Success: even Nobel laureates struggle to predict which of their papers will become their signature work, so they should keep creating; and artwork placed in a relatively obscure municipal museum is more likely to be loaned by major museums because that institution is a key node. Koji said, “If you just keep building, good things will happen.” 李磊 added: “Build smart.”

8. Advisor fit needs more firsthand investigation than a school’s label

• 李磊 puts choosing an advisor far ahead of choosing a school. He sees little fundamental difference between Top 30 and Top 50 schools, and acknowledges that an elite university’s name brings recognition, but says the variation between groups is greater. A PhD still looks a lot like an apprenticeship; the day-to-day relationship has a direct effect on how you grow.
• Google Scholar, CSRankings and public talks provide only public information. He recommends contacting current students and alumni privately, and looking at both their work and how they are doing: some people have published extensively and are deeply unhappy, so output alone cannot tell you whether a group is a good fit. He cites HKU’s NLP group and Peking University’s Institute of Computational Linguistics as groups with a relaxed, enjoyable atmosphere.
• He met his current co-supervisor, 孔令鹏, while interning at Shanghai AI Lab. It took about 6 months of collaboration before he felt their styles matched. That is why he sees a research assistantship as a good route: it lets you experience the collaboration firsthand, join the student community and learn about the group. Advisor selection is a “two-way match”; some differences take 6 months or longer to become clear.
• Geography matters, too, but with caveats. Mainland China makes it easier to connect with industry, but students should look for groups that balance project pressure with student development. Hong Kong has close ties to the mainland; in his view, faculty profiles there are more dynamic, though the cost of living is a pressure. He says he does not know the US well, and only relays friends’ reports of visa, CPT and internship difficulties, while acknowledging that the US is closer to Silicon Valley and the American academic community.

9. Professors’ startups can multiply resources—and pass investor demands down the line

• 李磊 believes a funding environment oriented toward technology has pushed frontier researchers and professors to the forefront of entrepreneurship. A company can buy chips more flexibly, hire engineers and purchase data, bypassing some university funding constraints and assembling the ingredients for larger projects more quickly. It acts as a “leverage accelerator.”
• The other side is that money comes with expectations. The resulting pressure can travel down to students and create “dirty work” beyond academia. He sees it as two sides of the same coin: more resources come at a cost, and students should weigh their needs against their career stage rather than focus only on the upside of additional resources.

10. Agents and large-scale RL were his 2 big shocks of 2026

• Looking back on 2026, 李磊 puts the rise of the agent paradigm front and center. His earliest research asked whether a sentence was positive or negative; later, models could translate and write essays. Now, with tools and scaffolding, they can complete many tasks autonomously. He hopes these capabilities next move from the digital world into the physical world, where new sparks can emerge.
• His other aha moment came during the team’s early training. 李磊 thought it was somewhat like R1 and believes large-scale RL can deliver broad capability gains. Livestreaming post-training was also a positive signal for the industry amid rumors that were damaging perceptions of Chinese LLMs: this can be done, and it works well.
• Koji noted that the livestream covered MiMo V2.6’s Pro and Flash models. 李磊 emphasized that RL has high infrastructure and data requirements. The team wanted to build confidence through transparency, open what could be open-sourced and use the livestream to gather feedback, address weaknesses and scale what worked.
• Chinese viewers cared most about the cost; the amount was displayed directly on the page, giving the impression of “RMB burning.” 李磊 says overseas commenters were more struck by the transparency, with some calling it the first livestream at industrial scale. Showing TensorBoard almost in public made the training dynamics a “gold mine” for outside researchers to analyze.

11. Data companies must keep moving with the model frontier

• On the high valuations of data companies, 李磊 first notes that he does not know the valuation landscape especially well, but does not think the valuations are obviously excessive. Compute, data and talent are all expensive; the complexity of data production and the quality bar have both risen by many orders of magnitude. He sees good data as the source of intelligence, while stressing that solid algorithms and good infrastructure are also needed.
• Demand persists because models have “jagged” capability boundaries. Data companies need to find the gaps, organize experts, design incentives and deliver at scale. Model companies can do this themselves, but it would divert focus and compute. “Time is an irreversible resource,” so model companies are willing to pay for data.
• “High quality” is always relative to a specific model: data that is too difficult cannot be learned by a model that lacks the capability. Suppliers need to judge difficulty, evaluate models and source experts. They also need to balance 2 kinds of value: PR value from making a demo look impressive, and economic value from the model truly learning the capability.
• He gives a clear condition for sustainability: companies that can keep producing high-quality “fuel” will continue to create new value; those that cannot will not stay in business. As models become better at generating their own data, costs will fall very low. Only a scalable pipeline that keeps upgrading may see its advantage compound.

12. RSI is advancing execution, while people still set the research agenda

• 李磊 says the signal for RSI is getting “stronger and stronger,” but warns that the term is broad. In the broad sense, using model-generated data to improve the model already counts as self-improvement. More specific discussions include models training themselves, exploring post-training branches, and finding better parameters and algorithms.
• He expects rapid progress on these tasks because they are well-defined and models are good at search; scaling compute can also help find better paths. During training, a model’s PPL keeps falling, and researchers can use models to help improve data, training and other parts of the process.
• The gap comes further upstream: “Which problems matter?” What should be solved in the next stage? Researchers can hand over end-to-end research only if models acquire that ability; he has not yet seen strong progress there. Asked to name the most influential RSI benchmark, he had no recommendation. The field is still early and has yet to reach a strong consensus.

13. A good benchmark adds a real dimension of capability

• 李磊 uses Video-MME, which he worked on, as an example. When others were evaluating 6-second video understanding, his team extended the duration to more than an hour. As agents become stronger, understanding movies and analyzing long videos will matter; moving from 6 seconds to an hour measures a “new dimension,” not just a longer test.
• Second, the problem should come from a use case and the data quality should be high. Overly artificial tasks are easy for models to pick up, creating the sense that “it knows you’re testing it” and ultimately a large gap with the real user experience.
• Third, a benchmark needs impact and usability: make it “one-click to launch, one-click to evaluate,” support mainstream evaluation suites and let others reproduce the score. Ecosystem integration matters because it directly saves adopters time and makes them more likely to use the benchmark.

14. Professional advantages shift; keep learning and protect your health

• Koji noted that mechanical engineering has become the most popular major among MIT freshmen, while enrollment in US 4-year undergraduate CS programs fell 8.4% year over year. 李磊 first allowed that “there may be many other factors,” then said a decline this large might partly reflect AI taking some junior coding work: “Every generation has its own version of the game.”
• He still thinks people should take 1 or 2 courses to build a general understanding of CS, even if they do not major in it, and learn to code with AI assistance. Asked what major he would recommend to a younger brother in his final year of high school, he would not pick a specific field. Instead, he would focus on comparing yourself with your past self and improving, staying in the top 20%–30% of a field rather than fixating on the top 1%.
• On the divide between people who use AI and those who do not, he acknowledges that he has been fortunate, but also sees many “ladders” across the gap. ChatGPT, DeepSeek and MiMo offer a wealth of knowledge; people in every field can use these tools to change their productivity and working methods, and contribute their expertise to improving models. The key is to recognize the change and embrace it.
• Through the long PhD journey, he says physical and mental health should always come first. He also mentions the “boxing theory” and the value of changing things up. His closing advice is straightforward: “The paper will get accepted sooner or later; you still have to eat and sleep.” It is not the end of the world.