Show HN: Hupreter – Create apps by describing them in English

[fersarr]

Hi HN! Fernando here. A few months ago [1], I shared Hupreter (https://hupreter.com) with you all and now I am finally opening it up so that everyone can try it for free. The goal is to let users create apps and process data effortlessly, by just describing what the computer should do, in spoken English. We have made a lot of progress since the first post, and even though it is far from perfect, I really want to see how people use it and get feedback.

In terms of creating apps, it supports persisting data (you can store/retrieve values), if statements, while loops, etc. For data processing, we currently support uploading tables, calculating the median/variance/etc., plotting, and more. And we will be improving all of those in the coming days.

For example, you can tell Hupreter: Given the table nba_players, calculate the average value in the column "points".

More examples are available here: https://hupreter.blog/

Thanks for your time! Let me know what you think, either in the comments below or via fersarr AT gmail

Fernando

[1] https://news.ycombinator.com/item?id=25945567

[andiareso]

I don't really know if I understand this. How is this not a programming language with a longer and more complex syntax structure? It would seem to be just as confusing to non-technical people as teaching them a language like Python.

[offtop5 → andiareso]

I can imagine this will end up creating more problems than it solves, for example depending on where you're from you might use various colloquialisms in speech which no AI will ever really get.

For example if I said something like hey I want all the NBA players but they got to be like a little bit taller than 6 ft 2. What's a little bit, once you get into any edge cases, or even have to clarify what you mean, you might as well just program it out.

[foxes → offtop5]

I mean what does a little bit taller than 6ft mean? A human would ask to clarify or supply a sane default. Using the Star Trek example characters would often ask the computer to refine something, eg “no not that tall”, so a system like this should allow for refinement from feedback.

[fersarr → offtop5]

yeah, if you are a programmer and it doesn't work quickly, you would just code it yourself. But if you are not...

[offtop5 → fersarr]

If you're not, you're going to endlessly frustrate yourself when this doesn't do exactly what you need it to.

I guess I'd be more interested if the tool generated python code or something, and then I can manually edit it, or one of my friends could send it to me so I can edit it. Don't get me wrong if this idea worked in practice I'd be all over it, but I don't think it will.

[fersarr → offtop5]

Hupreter generates python code, which you can see. Right now you cannot edit it, but we plan to allow that in the future.

[foxes → andiareso]

I think the end game should be basically Star Trek. You just declare to the computer what you want, and then it just generates correct code for you.

Declarative programming by itself is fairly powerful. I think In future this is actually how you want the system to work. Just saying what you want rather than how you manipulate the state to get it. The computer can generate provable correct code which does that for you.

[valgaze]

This is really interesting! The upload-a-table functionality is really neat

> On the more UI side, there's a guy named Sharif Shameem who released a somewhat-similar tool (during the early "holy cow!" days of GPT3's release) to build functioning UI's from natural language input: https://twitter.com/sharifshameem/status/1284095222939451393

> Jordan Singer made a Figma plugin too: https://twitter.com/jsngr/status/1284511080715362304

Slides 23/24 shows some basic examples of how this could work: https://docs.google.com/presentation/d/1m2pZfclAziqWrEJdebuf...

I don't know if people want it or how well this all would work in a non-toy scenario, but the notion of "conversationally" building an application-- even if just the scaffolding-- is really compelling for some reason

[fersarr → valgaze]

Really cool examples and thanks for sharing those slides, super interesting. Both the react app and the figma demo are super impressive. Like you said, I think that we will see a lot more of conversational tasks with machines in general in the future.

[twobitshifter]

Cool project, will I get syntax errors if I use “show me” or “display” instead of “print”. What if I actually want the computer to print on a piece of paper? “Print on the printer”?

It reminds me of HyperTalk some, but this seems to be designed to deal with the gray areas of language is that right?

[fersarr → twobitshifter]

Yea, exactly. It can handle different ways of saying something, like your example. Additionally it understands coreferences like he/she/it/the/etc (not perfectly though). But it gets better with time :)

[ddlsmurf]

As a developer this technology scares me not - for I have seen client specs and heard what they think they want.

[CommieBobDole → ddlsmurf]

I think eventually the role of 'developer' will be replaced by someone who takes business requirements and filters them for what's actually possible and desirable, then uses some sort of no-code software platform to turn them into instructions that a computer can use, then tunes those instructions based on testing and feedback until they effectively match the business requirements.

Oh wait...

[guenthert → ddlsmurf]

But isn't that where the profession ultimately heads towards? Less manual, more or less inspired coding, instead more dialogue with clients and translating their wishes into a specification which can be turned into code with ever more sophisticated tools?

[windmark → guenthert]

Definitely. Understanding the problem space and formulating the requirements is often overlooked when people claim "AI will take over programming".

Although off the shelf products will be able to do more and more tasks, I believe the demand for complicated solutions constantly will stay ahead of these.

This is exactly what we're seeing with Automation threatening the easy jobs, but as a whole, more specialized people are needed for the cases when that isn't enough.

[fersarr → ddlsmurf]

haha yea that is always tricky. To know what you want is something else :)

[amelius → ddlsmurf]

So, client iterates.

[ddlsmurf → amelius]

found the agile monkey

[gus_massa]

How does this compare with Wolfram Alpha? It looks like this project understand more sentences in English, but I guess the data that is preloaded is smaller.

[fersarr → gus_massa]

In terms of creating apps, the difference is that you can use typical programming blocks like if statements, while loops, DBs (persist a value), etc. For example: if three is prime, print "yes"

[johnisgood]

You would have more success with Lojban instead of a natural language, IMO, I think natural languages are limited by nature.

[grawprog]

I don't really understand this. I tried going at this as described with treating it as a natural language processor. I tried every iteration of 'calculate every prime number between 1 and 1000' i could think of. It was an error every time. I looked through the examples after and it seems to be a mix of natural language and formal programming constructs.

I don't really understand how to use this properly. It reminds me a bit of inform7 except less clear as to what's actually acceptable syntax.

[fersarr → grawprog]

Sorry to hear that! We will keep working to make it better, there are certainly many things that need to improve

[Detrus]

Maybe works in 10% of cases.

I remember seeing a natural language to SQL engine which worked in 50% of cases so the expectation was higher. https://blog.einstein.ai/talk-to-your-data-one-model-any-dat...

[fersarr → Detrus]

One of the hardest parts is guiding the user through what is currently possible, given that it's just a text input field. We'll continue to make it better.

[dennisy]

Is this built off the OpenAI API?

[fersarr → dennisy]

Hey Dennis. No it is not. We use offline open source models to create dependency trees and tags and our semantic engine takes off from there. After that the semantic structure gets transformed into code.