The AI cosmos these days seems like a very curious place to be: The funding hype of the past years has sobered somewhat, and companies and start-ups can’t just walk around anymore and collect venture capital for their gimmicky AI-powered app without showing the economic benefit in cold, hard numbers. Nevertheless, AI still feels like one of the most quickly progressing research areas at the moment (I might of course be biased here), especially if you include the ever-growing wave of new application domains. Models and learning approaches, alongside compute hardware, have gotten crazy enough to regularly produce “breakthrough” results in a boatload of problem settings that have previously been thought reliant on human intuition, such as Go, protein folding, language and image processing, and most lately generative tasks for text/speech, images, program code and video.
Two very prominent examples of this year constitute Stable Diffusion and ChatGPT. With these two, the usual things happened: They produce high-quality solutions with a level of intricacy in a problem domain/for a specific problem that has been previously thought to be “solvable” by humans only. As the word gets spread from the inventors over to the research community, tech magazines and eventually all the way to “mainstream” media, the details and outlines of the actual attained improvement get blurred and the models gets hailed as the next “step towards true AGI” or something like that. In addition to this, the first commercial applications or creative acts using this piece of tech become viral and prompt reactions/ opinions from many more people than are informed enough about this topic. The reactions range from statements of awe, fear and distrust and concrete ethical, moral, legal, practical and/or other concerns, all the way to some crazy sci-fi shit that could be made into a mediocre wannabe blockbuster. Many of the raised points are valid, many others are not, some get discussed, many get overheard, and slowly the hype dies down until the next model/architecture/big thing hits the news. Well, to be honest, every next model changes things a tiny bit more, and we’ll never know in advance just how much these breakthroughs will shape our lives (I’d recommend this read by MIT technology review). Anyway, I’d say that a “singularity”-like moment still is very unlikely any time soon.
This periodic appearance of hype however also has slowly changed the way the research community communicates their findings. In line with the increasing frequency of hyped models and applications, the AI (and by this I mean mostly the deep learning subfield) research landscape has acquired what I’d call a “social-media-like” vibe when it comes to their researchers’ communication patterns. I feel like AI researchers have gotten quite active on Twitter, podcast platforms, YouTube channels, books and blog posts, and thus new research is increasingly being promoted just like people are promoting their private identities (or their facades) on social media. This undoubtedly has some big advantages for the research community, e.g. increased exposure to very interesting related work and quicker propagation of interesting research insights and ideas, or networking effects. Just like with “regular” social media content, it can motivate, inspire and inform people, and staying informed and inspired is a core skill of any researcher. However, just as with “regular” social media content, content often carries more than just the objective information about the thing because it’s all about visibility, and the constant exposure to the presentation wrapping and “glitter” does something with the consumer (Here’s an article about a prime example for the ugly side effects of social media). Myself, I’m not one of the fortunate few that work on the next big thing at {OpenAI, Google Brain, etc.}, and I did not yet receive a {CVPR, NeurIPS, ICML, etc.} {best paper, test of time, etc.} award for my work. And so I have noticed that at times I feel intimidated by the groundbreaking innovations made by seemingly superior researchers which seemingly happen at an incomprehensible speed, to the point where I doubt that my work is needed at all anymore. Of course I have formulated my observation in quite drastic words, but I’m sure that there’s fellow researchers that have felt similarly.
For research prodigies and everyday researchers alike, it is part of our job to stay informed about what is the state of play, what is possible and what are open problems. Given the potential benefits of the new communication channels, it’s practically a no-brainer to make good use of them. However, this also means that it is even more important for us to learn to navigate these social spaces healthily, especially for us “ordinary” researchers that are not VIP guests at top-tier conferences or in tech magazines and podcasts. I probably do not have the wisdom nor competence to comment with authority on how to do this, but I feel like the solution is very similar to what people recommend for social media use in general (see here and here). Some of the rules proposed in those two articles, in our AI research context, could be read as follows:
- Evaluate your “why”: Make sure you’re aware of what precisely you’re looking for when browsing such communication channels e.g. something like latent diffusion model architectures for text-to-video generation and their evaluation on this and that dataset. If reading/watching a concrete piece of content, check carefully for a possible “layer of glitter” on top of the actual content: what is the actual result/improvement, and what is just empty phrases, missing/omitted/fishy details and marketing?
- Follow people and things that bring you joy: Avoid immersing yourself aimlessly in “cool AI news in general” too much, as you’ll end up being tossed around between the most glittery posts currently surfing the hype wave. Restrict your follows/search spaces to the few researchers that actually are relevant to you.
- Link instead of compare: Just like for general social media personalities, avoid comparing yourself and your work too much to the very shiny/hyped examples because you’ll just downplay the worth of your own work. Don’t worry: If you have done your research well, you will have found an interesting and – above all – open problem to work, and thus your work will not be worthless at all. Rather, use the opportunity and your knowledge to critically assess the communicated findings and connect with the corresponding researchers if you have further questions, comments or ideas for collaboration.
- Live in the moment: Try to not paralyze yourself and your own research activities by constantly scrolling through all the awesome research others have done. You are the one that makes the awesome research in your own domain happen, so remember to take long enough brakes from social media exposure to go out there and make it happen!
Research already is an activity characterized by uncertainty, so it’s not always easy on the mental well-being of its practicians. Let’s not make it worse by mis-using our newly acquired communication platforms, and instead use them to our advantage.