Akira Yoshiyama
About Blog Admirations Photos
An Ode to Alchemy
How lucky we are to work in a field in which we have created machines and processes we do not fully understand. Processes which require poking and prodding with the tools of mathematics and engineering. With few exactitudes we have the opportunity to drive innovation through empiricism, tricks and intuition. How lucky we are to have little precedent to guide us, little rigor in our protocols and so very much to explore.
Advice:
- Ideas can be scooped, but large enough problems will remain partially unsolved. Choose problems with large action spaces.
- Write about your intuitions and make them self-consistent. Keep a notebook of these thoughts.
- Alik Khilazhev's Advice. Sleep is important.
- Focus on one thing at a time. Context switching is costly and stressful. Projects will never end if you work on 10 simulteanously.
- You want as many tools in your toolbox as possible. Implement many metrics to really understand your experiments.
- Matthias Endler's Advice
- Regulate your emotions. What will you do when everything is failing for weeks? Will you stay principled, intelligent, strategic and systematic, or will you let your emotions take control of you?
- Read the logs and raw data, read the nitty gritty outputs. Be meticulous and you will discover things those who just look at the plots will miss.
- Richard Hamming's Advice
- Reason from first principles to come up with your own understanding.
- Many people have lots of ideas, the really great researchers can very effectively filter the promising from the misguided.
- John Schulman's Advice
- John Baez's Advice
- Read a lot of other peoples' code, look at how they do things.
- Ask for code reviews!
- Practice your software engineering skills even if you're a researcher. Try not to let yourself be bottlenecked by implementation speed.
- Prototype fast! Make things work, then make them nice once you have time.
- Andy Jones' Advice
- Be very principled about hyperparameter tuning. Make hypotheses about how your training dynamics should look and tweak your hyperparameters accordingly.
- Work with amazing people. People make or break your experience.
- Sometimes be lazy and sometimes definitely don't be lazy. Learn when to be which.
- Read science fiction.
- Work hard, play hard. Enjoy yourself during work and after work.
- Ethan Perez's Advice
- The most important decision you make is what you will work on. Ask yourself what will be the impact in the best case scenario. If big, then ask yourself if you're the right person to work on it.
- Aim to get experimental results at least once a day for good iteration speed.
- Terry Tao's Thoughts
- Take advantage of the connections you already have and grow from there.
- Publish usable artifacts. Along with your paper, release an (importantly) easy to use codebase, package, dataset, huggingface model, etc. People should want to use your work.
Scenii:
Motion Pictures:
- Monty Python and the Holy Grail
- The Princess Bride
- Indiana Jones
- Snowpiercer
- Schindler's List
- Parasite
- Fantastic Mr. Fox
- The Phoenician Scheme
- The Irishman
- Shutter Island
- Back to the Future
- The Big Short
- Three Idiots
- Chernobyl
- Seven Samurai
- Neon Genesis Evangelion
- Gremlins
- Rambdo (only the first one)
- The Untouchables
- Ghostbusters
- Wizard of Oz
- Edward Scissorhands
- Sweeney Todd
- Star Wars
- PK
- Black Mirror
- Forrest Gump
- Letters From Iwo Jima
- Inglourious Basterds
- The Death of Stalin
- To Live
- Hotel Rwanda
- The Notebook
- Silence of the Lambs
- Life is Beautiful
- The NeverEnding Story
- Top Gun
- Coming to America
- Argo
- BlackBerry
- Karate Kid
- It's a Wonderful Life
- Catch me if you can
Music:
- Basically everything (but as of writing this I'm listening to lots of Weezer)