Part 3 of 3: Finish the Research Story, Keep Senior Researchers Hands-On

Write toward a finished research story

Keeping up with papers and finishing research require different habits. At the opening of Part 3 of the conversation, one researcher describes mostly skimming papers to stay current. For their own projects, the practice is more deliberate: start writing while doing the research, imagine the eventual paper, blog post or launch story, and work backward to identify the work it requires. Sometimes a title comes before the experiments have been chosen.

The researcher keeps many Overleaf drafts, revisiting titles and presentation to make the final product tangible. In this personal account, visualizing the finished paper improves motivation and helps prevent experimentation from continuing without a finished research story. Writing becomes part of getting the work done.

Research taste develops in an environment

The researcher's career began with curiosity and a desire to remain in Singapore. They entered a PhD without carefully vetting an advisor, discovered an aptitude for research during the program, and learned much of the work independently. The path into international research comes without a reproducible recipe: looking back, they could not confidently tell a mentee how to repeat it.

Moving to Google in Mountain View brought a major culture shock. It changed the researcher's research taste—judgment about which work is worth pursuing and what makes a contribution valuable. They describe Singapore's research culture as focused on publishing papers and contrast it with an emphasis on real impact in the US.

That comparison draws an immediate challenge: - Institution changes alongside geography. The interviewer points out that moving from a university to an industrial lab also changes incentives, complicating the national comparison. - Both settings still pursue science. The researcher acknowledges the incentive differences but argues that scientists in both environments publish papers and seek to advance science.

The researcher's central point is that environments shape judgment, and accumulated judgment is hard to transfer quickly. Returning to Singapore, attempts to communicate changed standards could sound condescending to colleagues. Mentorship offered a more sustained way to work through those differences: while at Google Brain's Singapore office, the researcher took on an intern whose approach initially clashed with their own, then recalls being pleased with the intern's growth.

Respect talent; be careful what policy accelerates

Asked how governments could develop more AI research talent, the researcher favors restraint. They see ecosystem growth as organic and doubt how much governments can directly control it. Their recommendation is to respect talent and avoid accelerating in a bad direction: intervention can make conditions worse when it pushes the ecosystem toward the wrong goals.

The interviewer raises a separate concern about AI sovereignty—countries' ability to shape the values expressed by the AI they use. They worry that US decisions about model behavior can transmit US values through models used worldwide, and acknowledge having no clear solution. This remains the interviewer's concern, without a developed policy answer from the responding researcher.

The coda: senior researchers should stay hands-on

A coda moves to a Tech in Asia meetup featuring Yi, where Managing Editor Terence Lee asked how leading AI labs could emerge outside Silicon Valley. The responding speaker argues that senior individual contributors drive frontier AI research. Individual contributors, or ICs, do technical work directly; the speaker cites leading labs, including OpenAI, as examples and argues that governments may underestimate experienced researchers who remain hands-on.

Another participant suggests this advice applies beyond AI. The speaker accepts its broader usefulness but argues that AI breakthroughs are harder than ordinary software engineering, making experienced technical contributors especially consequential. In this view, seniority should keep experience connected to implementation.

The speaker contrasts two ways to organize research: - Seniority through delegation: a senior person directs large groups of interns while spending their own time in meetings. - Seniority through technical contribution: experienced staff continue writing code and participating directly in research.

The argument briefly becomes a demand that everyone code, but the speaker immediately calls that rule extreme and recommends incorporating only part of the advice. Keep senior staff technically engaged is the practical recommendation; everyone must code is explicitly qualified.

Who will actually train the model?

A hypothetical conference makes the organizational argument concrete. To establish Singapore as an AI hub, the speaker recommends inviting technically capable practitioners instead of centering policy speakers discussing AI safety. The concern is that an event can feel productive without supplying people who can do the work. Convening a discussion and building technical capability are different accomplishments.

The speaker then describes hearing secondhand accounts of territorial disputes around AI efforts. Their response is practical: who will do the work and train the models? Claims of seniority should face a corresponding test of coding ability. The question brings the discussion back from institutional status to the experienced people capable of carrying out an AI initiative.

The exchange ends with a curry joke about how spicy the discussion has become.