Using this post to think through all assumptions, strategy and technical decisions made so far about DenseLayers. Deploy date is tomorrow.
Today is 01May2020, and DenseLayers is ready to release minus a few last-minute changes that I keep making. I am in a state where I keep adding one more feature to make it easy to maintain the website once it is out, so that I can focus exclusively on content creation and marketing for the next several weeks.
It is classic scope creep. But in the midst of all the excitement, I need to recollect my thoughts and ideas and get a snapshot of where I’m at.
1. What is DenseLayers and what problem is it solving?
– DenseLayers is a website where people can discuss research papers.
– Most scientists and scholars have to spend a significant chunk of their time crunching papers. I know for a fact my friend in the lab at Cornell said he had to read 200 papers for his project.
– Assumption: reading oscillates between ‘skimming’ and being stuck, given the nature of these papers.
– The vision is that reading papers should be a collaborative exercise, done along with everyone else in the world, in the language of your choice, free of cost.
There are some other tools/platforms out there for doing this, such as Fermat’s Library etc, but they don’t solve the same problem or even solve it effectively.
2. How is it different from just using another tool like a subreddit or Google Docs? Why build a custom website?
– I like to control my destiny.
– I can build custom features and innovate out of the box, iterating quickly.
– I can manage the culture of the community by seeding it from the beginning.
3. What is the content and growth strategy?
– I will first begin with Deep Reinforcement Learning as the field I pick papers from. It is a small niche field but with a lot of exciting potential, and a lot of eyeballs on it. DRL is mainly done by a few key groups like DeepMind and OpenAI etc, and they only publish so many papers per year. Growing the site to become the de facto place for high-quality discussions around DRL research is much easier than doing so for a broader field.
– Papers should ideally be only those which are published on arXiv or other open-source journals. I do not wish to endorse or encourage publication in the major paid journals. Open-access is a good selective criteria that also reduces potential legal hurdles, and it makes the appearance of a paper on DenseLayers to be even more of a symbol of quality/novelty. Moreover, open-access journals used to be often considered to have questionable quality of peer review (the perception due to which Springer and Elsevier exist till today). Having a publicly open, transparent discussion platform improves their perception.
– If/once the site is successful at doing this, I can begin to offer a side service like job posts etc to begin generating just enough revenue to sustain the site.
– After that, I can start to expand into other adjacent fields which are interdisciplinary in nature, such as Deep Learning in Neuroscience etc.
– I’d like to focus on fast growth in terms of the quality and depth of discussions, slower growth in terms of papers added from each particular domain, and even slower growth in the number of research domains on the website. Quality is the goose, growth is just eggs. If you let the goose die, the eggs will stop coming. Wow I just came up with a good adage.
4. Talk more about the revenue model and long term plans
– Not entirely sure yet how the site will make money, but so far a job portal sounds very enticing. It’s very much like the StackOverflow model, which was actually an early inspiration for the project. I’ve also studied the models of Reddit, Wikipedia, Quora, Yahoo Answers, Urban Dictionary, Justin.TV among others.
– Finding niche talent from a research domain is hard and expensive, so companies probably need help. I intend to offer it for as cheap as I can as long as the website can stay in business.
– The job posts thing can only happen if the website grows enough.
– But the long term prize is to build the world’s greatest open-access publication platform. Scientific publishing has been staggeringly profitable for way too long. It is time for the giants to fall.
5. Where will I find users?
– I happen to have a relevant following on Medium, and I will soon begin to publish more essays there. Every time I publish a paper on DenseLayers, I will write its breakdown essay there to generate interest – if anyone would like to join the discussion, they can do so.
6. Thoughts on alternatives/competition
– The basic necessity for the solution to grow in usage, is to inherently be useful from day one. Most platforms have a chicken-egg problem. What I can do that others aren’t, is to seed extremely-high quality content in the beginning.
7. Potential problems, Achilles’ heels, etc
– Deepmind and OpenAI themselves also publish long blog posts that clearly explain what they did in each research paper, in a fairly easy to understand way. So do most researchers. So there may not be a big enough need for DenseLayers to attract readers.
1. Well, they don’t really offer a consolidated platform where anyone can add their own thoughts, and read multiple papers across companies without having to search around.
2. People will still read research papers.
3. My writing can rival theirs. I’ll try my best.
1. Much more beautiful than I intended at first, due to a friend remarking that the design is “terrible”.
2. Decided to use memcached, and then good friend Patrick said I shouldn’t, to avoid unnecessary complications for an MVP that isn’t meant to scale yet. So not using memcached.
3. Will deploy on Heroku, and use SendGrid for email support.
4. Added user moderation. Admin can delete posts that break rules.
5. Decided to show user names on paper page with Gravatars. I guess the friction of not being able to find your own comment is more of a nuisance than anonymity. Will need to add anonymity controls in next update as well as upvotes and downvotes.