Speed is Everything for AI research.
I will state something very obvious in this post. To excel in AI research, whether in academia or industry, the key is to move quickly and deliver results promptly. I will refer to this as the speed-first principle. Although this principle may seem self-evident, not everyone acts upon it. Even I, at times, fail to follow this principle. Therefore, it is worth explaining why this principle is essential and why we should advocate for it more. Moreover, I will argue what the speed-first principle is not, what it is, its benefits, and how to practice it more.
I would like to provide some context before presenting my arguments. I conduct AI research in academia, where I primarily interact with MS and PhD graduate students in computer science. Although these students are generally super smart, capable, and motivated to conduct research, not all subscribe to the speed-first principle. Especially among new students, there are many reasons why they may not believe they can deliver quickly and take the initiative. However, I will make the case that speed is everything in this post.
Why speed-first principles are essential.
- Rapid Evolution of AI field: AI is a field that is advancing at an extraordinary pace. Speed-first principles are essential because they allow researchers to keep up with these rapid developments. By executing research projects quickly, researchers can stay ahead of the competition. By doing so, they can contribute to the frontier of the research field, rather than being followers.
- Rapid Feedback: AI research is heavily reliant on data. Speed-first principles facilitate a quicker iteration of the data collection-analysis-feedback cycle in your own research practice. This accelerated feedback loop is crucial for refining AI models and algorithms, leading to more accurate and efficient outcomes. But also the rapid feedback can help you find the right problems that matter most instead of wasting time on something that is not going to make a difference.
- Global Competitive Environment: The AI field is highly competitive, with many people striving to invent new methods, algorithms, systems, applications, and datasets across the globe. Talents from all over the world are trying to innovate and compete in the AI field today. Being fast is absolutely crucial if you want to be a player in this arena. If you’re not faster than most people, your work will never be published or used by customers. Or your patent will not be granted, and your startup will not gain customers.
What speed-first principles are NOT
Not Sacrificing Quality for Speed: Emphasize that speed-first does not mean cutting corners or compromising the integrity and quality of the research. It’s about efficiency, not haste. In fact, I argue if you take time and work slowly, it often leads to a worse quality product whether it is a research paper, software, or a domain application. This is because practice enhances quality. Frequent practice and rapid iteration typically lead to better products than a slow and prolonged approach.
Not Constant Rush: Speed-first is not about keeping the team in a perpetual state of urgency. Instead, it’s about streamlining processes to minimize delays. I have to say that the most challenging aspect of implementing the speed-first principles is avoiding rushing and burnout. Personally, and I believe many others would agree, this is an area where we all need to improve. The best way to prevent rushing is to start early and complete tasks efficiently and confidently at ease. As the old saying, it’s not only about doing things quickly but making them look easy. By doing so, we avoid the need for last-minute urgency and can iterate enough time to produce a high-quality product. We all know that rushing at the last minute leads to poor-quality work, so starte early.
Not Short-Term Focused: Speed-first principles do not ignore the importance of long-term goals and sustainability. It is about making steady progress while keeping a watchful eye on the long-term vision. However, I believe a long-term vision only matters if one can successfully execute their short-term goals with speed and confidence. If someone struggles to achieve their immediate targets, discussing long-term vision is a futile exercise. First, the person may not be qualified to make long-term judgments on AI research direction. Second, no one would respect or believe in a person who cannot deliver their short-term goals. Third, I have never seen anyone who cannot deliver short-term objectives provide any valuable insight into long-term vision. I agree and encourage individuals who excel at delivering their short-term targets to take time periodically to assess and plan for their long-term goals. It may be beneficial for them to explore emerging directions, make necessary connections to promote their work, understand what is happening in other areas outside their expertise (e.g., attend seminars in other fields, talk to colleagues or domain experts, attend new conferences they’ve never been to), visit different places (e.g., visit industries to understand their practices or priorities), and find exciting datasets. Lastly, it is important to write down the long-term goals you have for yourself and others and gradually convert them into short-term targets.
What speed-first principles are.
The speed-first principles are about creating a product of the highest quality within a fixed amount of time. The speed-first principle requires 4 strategies: iteration, adaptation, prioritization, and determination.
Iteration: I am a firm believer in the Scrum methodology. The focus on speed allows individuals to treat time as a valuable and limited resource and plan accordingly, avoiding waste. The key is to iterate consistently and obtain feedback as quickly as possible. For instance, instead of choosing a topic, and spending as much time as needed to write a paper, why not choose a deadline first (e.g., AI conference deadline), select a topic that can be completed within the timeframe, and receive feedback? In fact, consider every possible target venue as an invaluable and free way to gain feedback. By choosing these venues upfront and earlier, you will have a better scope and target, and you will know what papers to read, what the target paper should look like, and when it should be completed. The predefined target simplifies the plan and provides extra motivation to move quickly to meet the deadline. Even in the worst case, you are not suitable to do AI research. By rapid iteration, you will find out much sooner, gain valuable skills and insights along the way, and make better decisions about your future sooner.
Adaptation: Speed-first is a strategy that allows you to constantly adjust your goals and methods, which is more flexible than having a long and fixed plan. In the field of AI research, new advancements are made daily, making it impossible for anyone to read everything on any topic. Therefore, it’s important to quickly adapt and build confidence in pivoting when necessary changes arise, such as someone publishing the same or better idea as yours. It can be discouraging to find out that someone else has already published a similar idea to yours, but this is common. The best way to deal with this is to adapt and improve rapidly based on the current situation. Sometimes, the topic you are working on may no longer be relevant in the current context of AI research. When this happens, it is advisable to wrap up (i.e., publish) your work quickly and move on to other relevant areas. Lingering too long on an irrelevant topic just because it is the only thing you are good at is not a healthy strategy. This can limit your growth and impact on the world.
Prioritization: To prioritize speed, you will need to arrange multiple tasks in an optimal sequence for execution. For instance, if you have to submit a paper by a specific deadline in a month’s time, you will need to create a list of all the tasks that need to be done. You will then estimate how long each task will take and prioritize them to minimize the total required time. We should also optimize the process by 1) removing non-essential tasks from the list, 2) finding time by cutting other activities outside the list, and 3) doing tasks faster and better. Over time, task 3) will become more efficient, while we have to balance tasks 1) and 2). By prioritizing speed, you will have a better chance of completing the right tasks and avoiding unnecessary ones, thus reducing waste.
Determination: The first three speed-first strategies are tactics, while the last one is about mindset. In AI research or life, unexpected issues can arise, both small and large. For example, a compute job might crash due to server issues, or someone might get sick due to COVID-19 or face a personal issue at home. It’s easy to push back the original goal and say, ‘I’ll just submit to the next conference or journal deadline when I’m ready.’ This mentality of failing a deadline as long as there are sufficient justifications is the most common reason for failure in life. True researchers find a way to deliver and deliver well and on time, regardless of the barriers and unexpected hiccups. They will be relentlessly resourceful to meet the original target because they can’t sleep or live normally if they miss a deadline. They won’t stop until the current goal is achieved. They never crush under pressure. This mindset to deliver on time with high quality is what speed-first is truly about. I’m not sure if this determination attitude can be taught or if it’s innate to individuals, as I haven’t been successful in converting anyone from one camp to the other. However, I know that I would love to work with anyone who has this trait because I know they will succeed quickly in the long-run.
How to practice more speed-first principles
Now that you are convinced about the speed-first principles, how can you and your team become more speed-first? What concrete steps can you take to improve?
Individual-level
Time management: To be speed-first, efficient time management is essential. Various techniques such as the Pomodoro technique can help improve time management. However, I find that maintaining a regular routine and tracking activities along with their outcomes is key. Personally, waking up early and exercising in the morning can provide me with the necessary energy to have a productive day. I prefer reviewing my daily to-do list quickly and completing a productive task early in the day, such as writing this post. But it’s important not to make your to-do list unrealistically long. There are many things that require your attention, but you only have limited time. Therefore, it’s better to keep the to-do list short/concrete and clear the list every day.
Rapid Learning: To be speed-first, people all need to be rapid learners. There are so many resources online, it is really a matter of willingness to learn quickly (e.g., taking online courses, watching technical Youtube videos, reading papers and books, using LLM to help to learn, and attending workshops and seminars). Rapid learners also know when to focus on a deep dive into a specific area. Moreover, it’s important to approach new subjects confidently and quickly, without fear. In reality, most people don’t know much more than you do, and if you can learn quickly, you can quickly catch up and even surpass them.
Agile Methodology: To be speed-first, everyone should adopt agile practices when working and studying, such as breaking down tasks into smaller, more manageable parts, setting short-term goals for quick and measurable wins, and being adaptive.
Technology Leverage: To be speed-first, we should use state-of-the-art software tools to be productive. This includes all the online resources, and AI assistance such as ChatGPT or other LLMs, co-pilot for coding, and writing assistance such as Grammarly, Paperpile for reference management, Overleaf for paper writing, calendar apps, Trello board for to-do and scrum, slack for team communication, and many more. Don’t be a guy stuck in the past doing it the hard way. AI research is still hard even if we use all the tools. Let’s embrace technology while inventing new ones.
Frequent Reflection: To be speed-first, it is essential to be self-aware of your work process, outcomes, and your strengths & weaknesses. Regularly assessing yourself can help you identify bottlenecks and areas for improvement, which can help you streamline your workflow. It is also helpful to benchmark against your peers to see how you compare. Don’t be afraid to acknowledge when someone is more productive than you, as it is an important step towards continuous improvement.
Team Level:
Effective Communication: To be a speed-first team, it is crucial to establish clear, concise communication channels. Use message apps like slack or Teams to encourage effective/real-time communication.
Collaborative Environment: To be a speed-first team, you need to foster a collaborative team culture where members can freely share ideas and feedback. It is important to utilize collaborative tools for efficient teamwork, e.g., Trello board, shared drive, slack channel, and GitHub repo. It is not to make everyone within a team be friends with each other. It is more to establish accountability and encourage the culture to deliver.
Goal Setting and Monitoring: The team should set clear, measurable, and time-bound goals. Regularly monitor progress against these goals to ensure the team stays on track. This kind of meeting can be a short group meeting so everyone is aware of each other’s progress.
Celebrating achievements & milestones: Acknowledge and celebrate milestones and wins to maintain motivation and momentum.
Identify and address underperformance: Speed-first team should consist only of speed-first individuals, and the nonperforming members need to be managed out periodically.