Comment by pbronez | Hacker News Reader

Comment by pbronez | Hacker News Reader

This is an interesting interface for a sophisticated machine learning model. The user input is a moderately structured prompt (a line-delimited list of facts) and the output is text crafted for a particular medium (email).

This could be extended to support output for other mediums, such as Blog posts or social media updates.

It does drive me slightly crazy that this system JUST INJECTS NOISE. You provide it with concise, readable information. It dilutes that information with additional words and scaffolds it over complex sentence structure. The resulting email has a flourish, but takes more effort to read. It's an inverse document summarizer!

Perhaps there are benefits to this. Sometimes you want to make your reader invest a bit of effort in your message. I believe that can improve information retention. The email is also a generally-accepted human API. People might be more receptive to information presented as a letter than as a list.

The biggest benefit is probably on the author's side. This forces you to actually think about what you want to say. The input is basically an outline!

Paradoxically, I think Flowrite could improve communication by making authors and readers do MORE cognitive work rather than less.

> Paradoxically, I think Flowrite could improve communication by making authors and readers do MORE cognitive work rather than less.

I would love it if I could get recipients to actually devote cognitive work to my communications. Unfortunately, increasing cognitive load is a great way to ensure fewer people read what you write.

Every recipient of a message thinks differently than the writer. For communication to be effective, it must be designed for the recipients, to trigger the right sequence of ideas that will get the message across. I’ve written up the exact same content in 3 or more ways to target specific teams - bullet lists, email, slide deck, etc. It’s always worth my time. Before I did this, I would get several elaborate responses of “why didn’t you write this in X format”, arguing that X was a superior format. But everyone provided different versions of X. You may think conciseness is paramount; not only is it not, the definition of what is concise varies from person to person.

This

> It does drive me slightly crazy that this system JUST INJECTS NOISE.

is actually an important service. What is noise to the sender is generally not noise to the reader. It would be awesome if at some point in the next few decades we only have to provide information that we need to think about and the computer could find a way to compile it into what the reader is looking for. In other words, if the computer could do what most every secretary was doing in the 1960s.

If the example is correct, it's injecting more than fluff and flourish. It injected "After weeks of hard work", which doesn't seem supported by anything in the input, and probably sets the wrong tone. Certainly it was at least many months of work prior to launch.

Scroll up or go to https://www.flowrite.com/?ref=lbixer#examples to see more examples of what it can do.

The additional examples seem to integrate external knowledge. How does that work?

It expands "follow up" to "in person or via Skype / phone call" -- where did those platform preferences come from?

when asked "How about a Zoom call on the following Tuesday at 2PM?" it replies "I'll be in a Zoom meeting on Tuesday at 2PM, so let me know if that works for you." Does this mean Flowrite looked at a calendar and detected a conflicting zoom meeting? Or is it trying to accept the proposed meeting time & venue?

I do really like the Blog Outline example, where it expands a general question ("How to align your Marketing, Sales and Customer Success teams to maximize revenue?") to a list of six other questions to dig into the problem more. That might be a powerful way to push people to think more deeply and question assumptions.

> where did those platform preferences come from?

GPT-3. Which I think is reasonable - not a lot of work to read it through and replace it with "on Zoom, or next time you're in the bay area".

Good point on the calendar though.

Or is it? Huggingface model fine tuned in 3 weeks? Maybe the model knows what it's saying. /s

I think text written by models should be properly marked as such. It's too cheap to generate and doesn't have a reputation to defend.

So the first thing I would do with the message would be to put it into a model to detect artificial text, before deciding if I want to see it.

Isn’t this just GPT-3 under the hood? Other similar things do exist (eg: copy.ai).

Not sure whether there’s much of a difference between all these GPT-3 powered services when all that distinguishes you from competition is some (slick) UI and the 500-1000 extra words of “training” you give to GPT-3.

Gwern [1], who has spent quite some time with GPT-3 and previous models, seems to think that coming up with the right 500-1000 words can be a subtle business.

Now, what's the prompt that will get GPT-3 to generate good prompts for us?

We'll call this technique Promptception.

[1] https://www.gwern.net/GPT-3#prompts-as-programming

https://arxiv.org/abs/2102.07350 already calls it "metaprompt" :) I gave it a quick stab a while ago but I think prompt programming is too new, and you can't easily cram demonstrations into an existing prompt, for it to really work well. It's more promising to train models on examples of tasks from instructions, or work on directly optimizing prompts for a goal (https://arxiv.org/abs/2101.00190) - it's a differentiable model and a whitebox, so use that power!

Put it is way, it might make it easier to write the reverse system that convert a (machine generated) noisy input text to a succinct list of facts :).

Sounds like a way to make a text based GAN :)

Reply on news.ycombinator.com