Cool project! Something to consider: "Transformer"[1] is already used to refer to a popular element of state of the art deep neural networks, especially ones that are used on natural language, i.e. text. That might make this name a little confusing for people who have a foot in that world, especially since "Universal" and "Text" are also words thrown around in similar contexts.
Good to know, thanks! I was originally considering a name more like "universal text finagler" but then I realised that there's already a well-known "UTF" (:
The term "transformer" is used throughout mathematics and computing. Machine learning hardly has the monopoly on it. Or should, for example, Haskell rename their monad transformers library?
Well, the really ironic part of it is that the machine learning models using transformers were literally conceived for converting between languages. So this would be quite the valid use case. Unfortunately, it makes the name more confusing.
It doesn’t look that messed up to me. The vector word is used both as an element of a vector space and the coordinates. Tensor was so far mostly used only for the multilinear relationship, but its numerical description is a natural extension of the vector word.
Making one oversimplification means you need to be very particular? The point stands that some people that look at this tool will likely expect something very different than what it is. If this tool gets enough prominence, it won't be a problem. If it doesn't, it will be a misleadingly named software stack. Worse things have definitely happened.
Almost every Show HN includes detractors complaining about namespace collisions or like parent, general name critiques. This is fine, I just wish it'd be towards the tail end of the thread since as a rule of thumb, such discussions are intellectually low-value.
It would be nice, albeit a bit tricky to implement, for downvotes to be optionally augmented with a choice from capped list of reasons. Among them, "it would be better if..." and "i have a different opinion", in addition to "this is badthink" and "this is offensive to me".
To take this idea one step further, each of those options could be attributed to a colour value. Then, instead of "greying out" the resulting comment would be an amalgam of the received votes, giving an indication of exactly what it was in particular that caused the comment to be downvoted. The stronger hues would then indicate a degree of consensus among the audience.
Of course, the same should be applied the other way, so when upvoting, one could also choose to assign a value for "this is good but could be improved", "I have a different opinion, but the above comment is worthy of consideration", "this is simply a correct take" and "this comment is pleasing to me".
And for the sake of readability, maybe instead of the whole comment being coloured, a marker could be applied before or after unadulterated text so that it remained accessible during the course of this process!
The challenge would be in deciding which colors represent what sentiment, as these days even our visible light spectrum seems to be corrupted by political/moral/tribal prejudgements. So perhaps it would do more harm than good. It may make for an interesting experiment, all the same.
I would recommend highlighting what are the supported input and output formats or at least highlights of what is supported. Another key question I have is why use this tool instead of jq?
Interesting project and seems very useful. Reminds me of Pandoc, maybe you could make a diagram similar to how [1] Pandoc shows conversions possibilities.
Not sure how long it takes today to start up and load classes on a JVM. The idea of using a java program regularly in the shell immediately makes me worry about slowness.
Also, the eventual goal of supporting more formats means slower startup time as the feature set grows.
42 comments
[ 3.2 ms ] story [ 95.5 ms ] thread[1]: https://en.wikipedia.org/wiki/Transformer_(machine_learning_...
this could be called tiny, huge, text transformer: thtt
tiny because functionality, and huge for java
Transformation is also used in enterprise IT to mean modernising.
Transformer is also a mathematics term, a kids toy and many other things.
In short, ML doesn’t have a monopoly on it.
(That you can represent a tensor by a multidimensional array does not mean that a tensor is a multidimensional array).
https://en.wikipedia.org/wiki/Tensor
Of course, the same should be applied the other way, so when upvoting, one could also choose to assign a value for "this is good but could be improved", "I have a different opinion, but the above comment is worthy of consideration", "this is simply a correct take" and "this comment is pleasing to me".
And for the sake of readability, maybe instead of the whole comment being coloured, a marker could be applied before or after unadulterated text so that it remained accessible during the course of this process!
The challenge would be in deciding which colors represent what sentiment, as these days even our visible light spectrum seems to be corrupted by political/moral/tribal prejudgements. So perhaps it would do more harm than good. It may make for an interesting experiment, all the same.
> For example, utt does not process data in a streaming manner, but rather loads the entire dataset into memory before processing
[1] https://pandoc.org/diagram.svgz?v=20220210130556
Not sure how long it takes today to start up and load classes on a JVM. The idea of using a java program regularly in the shell immediately makes me worry about slowness.
Also, the eventual goal of supporting more formats means slower startup time as the feature set grows.