Read on Twitter
Thread about what's wrong with machine learning tools today.
1/ It's almost every week that some new "no-code" machine learning tool is released. They usually have a pretty sleek UI, boast about how quickly you can build a model, and say they aren't sacrificing performance. What's weird is that their value props are all the same.
2/ The issue is that the industry as a whole isn't realizing that this these "no-code" tools aren't what people want. If the steady introduction of pretty similar tools was working, the growth curve we'd see would look *atleast* something like this:
3/ Instead we're seeing high churn in these AutoML businesses, where they're only able to capture very fragmented parts of the market. People don't like these drag and drop softwares, with so much abstraction that we're forced to just "trust" and proceed with inference.
4/ You can argue that it's because of performance just isn't there. But the way to solve that isn't the "Apple" approach, where you focus on design. Even if you could produce the best performing models, you still wouldn't have resolving the core reasons why people don't use it.
5/ We need to focus on making these models we're building iterable. After you've trained a model in any sort of ML tool, you're forced to stick with the result. You can't *change* anything, even if you might have some insight on what would work better. Iteration is key.
6/ Next, there's a push towards building tools that non-technical users can use. While this is wonderful, we need to accept that the space of ML/AI is only going grow as time passes. Using these tools only leads to a neglect of the field, which will only hurt in the future.
7/ My philosophy is that: let's build ML tools that are easy to use in the beginning for anyone, but decrease the "time to insight." Give the user the ability to apply new chunks of knowledge where they choose to do so.
8/ Why not only automate what they don't know how to do. Let the user pick and choose where they start learning, and where they get involved first. This is the future. This opens the doors for non-technical users who can "not get involved at all" and hard-core ML engineers.
9/ If somehow, you've read this whole thread then what I just described is exactly what I'm building with Paraglide. I'd love to have you check it out and join the waitlist at http://paraglide.ai !
