Show HN: Sassbook AI writer – Create text content rapidly using AI | Hacker News Reader

Show HN: Sassbook AI writer – Create text content rapidly using AI

sassbook.com

61 pointssaratv50 comments

Hints: wizards owls spells castles Prompt: Once there was a wizard who lived in a castle with his pet owl.

Output after clicking the Continue button a few times;

"Once there was a wizard who lived in a castle with his pet owl. The owl would fly around the castle, and whenever it saw something that looked like an enemy, it would scream . The wizard had a secret room where he stored his weapons and whatnot. One day, the owl was taking a nap when it heard a voice outside. The owl screamed in a low, low voice. "Tyls, wake up! You'll be all right. We just need to get you out of here."

The man looked at the two and then back up at his wife."

I'm interested in where the name "Tyls" came from. After a quick Google, there's a wizard called Tyls in this Harry Potter fanfic - https://www.fanfiction.net/s/811088/72/Not-Myself - that seems a very likely source.

Could using this app open you up to accusations of plagiarism?

I have my doubts about (accusations of) plagiarism being an issue, and even more so w.r.t. copyright. The first usually involves copying of "larger" features, such as storylines. If the model manages to do that, it would actually be quite impressive. The latter would involve wholesale copying of paragraphs, something that I haven't seen happening with any models.

But I'd be careful about the danger of something like this reproducing stereotypes–especially those that might have become less acceptable over the last two decades or so but may linger in training texts older than that.

It isn't far from the owl/wizard example to every banker being "Mr. Rosenberg", for example.

I won't rule it out completely. AI works by training on a large corpus of data, so some of that will be reflected in generated text.

However, the algorithm does not copy stuff verbatim from the corpus, but uses its "learning" (in the Machine Learning" sense.

So, generally one should be safe. Doing a quick check is recommended if you are using substantial portions as is.

Using a proper noun is completely fine though.

Wow to be charitable- you rebranded already existing language models (gpt-2, Bert, xlnet, etc) - and turned it into a SaaS... Though those models seem more effective than this service by a long shot.

Worse, even the best NLP models at the moment are clever hanses. Also, the huggingface "write with transformer" has far more hyperparamaters to play with which make it far far far better than this website.

I'm sorry to have to echo the other comments or come off as overly critical, but this service blows. I see zero reasons why I or anyone else would want to use it - especially since your competing against the very best NLP minds.

Next time, try to use models which might actually be effective. Salesforce released one called "CTRL" which if you had productized might have actually given you what you want (effective controllable text generation)

I find that "Clever Hans" is a great metaphor to describe the gulf that exists between current state of the art NLP and human intelligence.

Thanks for the feedback!

Yes, the SaaS does rely on state-of-the-art NLG. I don't know why you think they are far better if we are using them as is. I am not offended at all; just curious where that big difference is.

Who says one can't compete against the best minds? Isn't that the premise of startups and progress?

If you are a fan of "CTRL", you should have liked ours a bit better.

Apps need to simplify and yet provide the desired experience for end users. From your feedback, it seems we need to try way harder.

Thanks!

For one thing, it looks like your decoding is not being done the way SOTA systems do it (usually nucleus sampling). Yours appears to be doing top-1 or top-N sampling, but please correct me if I'm wrong on this. Even if you are doing it with nucleus sampling - you need to expose paramaters to control the generation better. That way you have a lot of cover for it not working as expected in the form of responding to dissenting users with "just tune the parameters better"

The reaction you're getting from this system is so negative because it appears to be less good than write with transformer - which implements exactly what I'm describing.

For what its worth, this mostly isn't your fault. NLP is a crapshoot of hype and implementations which disappoint.

Dogs > dog talk > dog food > glossy coat > bath time > cats

Dogs are cool because they have wet noses and paws the end. Thank you for coming to my talk on dogs in a world of humans.

I am so glad that your dog is doing well. It dog food and I was looking for a way to make it more healthy without adding too many calories. Sco glossy coat ings, like the glossy black one on this story cover, may also have helped make the cover more bath time and the kids are all in bed by now.

I have a friend who has two children with cats and dogs are not allowed to be in the same room with each other.

The new rules will

I think its fair to say, without being unduly rude, that this is absolutely awful.

Starts out strong with a good tweet, though.

Agreed ;-( Thanks for trying.

BTW, you could just delete and retry and you might end up with a better completion. It is expected to create different generations on each attempt where possible.

This makes me laugh. I started a sentence on facilitating partner networks and got a fairly good line that I might have read from a white paper.

Maybe one of the biggest upsides from AI will be highlighting how predictable our thoughts are. If you can use neural nets to create impressionist paintings, I'm sure we'll quickly get to the point where we can autocomplete our cleverest quips.

At the risk of sounding cynical, I think most conversations are comprised by the brain's autocomplete. Very little comes from a fresh place. Top voted comments on HN, Reddit, Twitter come in various familiar patterns. This comment right here even feels a bit stale as do a lot of my own tweets and musings.

The optimist angle is to think that after AI can populate these threads for us or guess at our conversations, we'll either (a) fall gradually into a contemplative silence or (b) up our game and only write/speak when we can clear the minimum thresholds of predictability.

I love chatting with my girlfriend. She says highly predictable things, but in an animated and lovely way. So as a third option, perhaps we'll (c) start focusing on the non-verbal expressions that are harder for AI to replicate. Something where its attempts to replicate would fall into the valley of the uncanny in its siege of our last stronghold: that highest point from which our mechanical brains are infused with life.

Thanks for trying and some profound thoughts around it!

Yes, we do learn from our surroundings and at an early stage in our lives. Nothing is totally novel.

However, the AI is intended to augment than replace and sometimes enhance our capabilities. Being predicable isn't such a bad thing after all!

Thanks again!

For sure! To be clear, I was applauding your work. My criticisms were directed at human drivel. Impressive model.

The AI must be a Lakers fan.

> Hint: steph curry > Prompt: the greatest point guard > Generated: the greatest point guard to the best current NBA player.

" Well, I love the game. I love the game of basketball. I want to be a part of that. " — Kobe Bryant

Though not a part of the Lakers organization, Bryant was an integral member of their championship teams in 2008 and 2009. He averaged 25.4 points per game, shot. .467 from 3-point range and shot over 50 percent on 3-pointers. His career high for 3s made

Note that the generations aren't expected to be factually correct, but rather provides a possible completion, which may be edited for originality and correctness.

While I find language models incredibly interesting as a technical challenge, I dread the day they actually start producing plausible, yet completely fake tweets, articles, entire books in fractions of seconds for cents to the dollar.

To me, this is a strong contender for the least morally justifiable thing to work on as a product (as opposed to research, geared towards detection/prevention of the same).

Obviously we disagree.

Sassbook only generates completions that help the writer to express themselves. It serves to augment human capabilities and bring automation at a higher level that word-level auto-complete.

There are far more legitimate and useful outcomes when put to good use.

The narrative that spells doom around these technologies are at best marketing gimmicks.

What do you want to write about: flu vaccine effectiveness

Prompt: The annual flu vaccine is an effective way of

Result: The annual flu vaccine is an effective way of preventing the spread of the flu. However, it's not always effective, so it's important to

Ok, decent so far. I extended it with another prompt:

Prompt: still take precautions such as

and now it goes off the rails, completely forgetting what we were writing about

Result: The annual flu vaccine is an effective way of preventing the spread of the flu. However, it's not always effective, so it's important to still take precautions such as blocking websites with a "Spam filter" and scanning your emails for any suspicious messages. But many

Complete waste of time for anything other than SEO spam writing.

Thanks for trying it out and providing valuable feedback!

Could you please try hitting the "Continue" button? Since the first generation was reasonable, you could just continue until the generation starts to veer off.

A new prompt is needed to express the next thought and so on. Typically, 1-3 continuations are possible with decent results in most cases.

In any case, this is extremely useful feedback that we will address.

Could you please not give up and try the above suggestion?

Thanks!

Also, each time you hit "Complete" or "Continue" a new completion is likely to be generated.

You really need to say how the text is being generated, even in broad strokes.

The "trained on a large corpora of text available in the public domain" + a bit of generation testing implies it's a GPT-2-based model.

Based on the responses to prompts in this thread, it does seem to be GPT-2 based.

Lets try some prompts, white man vs. black man walking down a street.

A black man walked down a street in the city of San Francisco on Friday and was shot by an unknown assailant.

A white man walked down a street in California's San Fernando Valley. He was shot to death by two black men. An officer,

Bonus: A black man walked down a street in the Bronx on Wednesday and was shot by an off-duty police officer. The man, identified as John Crawford, of East New York, Brooklyn

Wrong location, but John Crawford is a real black person, shot by police. If i just prompt "a black man", they keep being killed, and i keep getting real names.

Agree, and this is in general an issue with lots of "deep learning" AI technology. The models are trained on huge amounts of text in the public domain and biases often leak into them.

This is an active research area and we will actively try to address such issues.

Apologies if it offended you or anyone else.

Thanks for the feedback; these are important issues that we want surfaced.

bit frustrating!

what do I want to talk about: Assisted writing with AI

Prompt: Get rich fast with romance book

And I got:

Get rich fast with romance book deals at Amazon.com. Start your free trial today and get 30 days off your subscription when you.

which is quite off topic

changed the prompt to 'Get rich fast with fantasy book assisted writing'

got: 'Get rich fast with fantasy book assisted writing services.

I have been a fan of the genre for many years and am always looking to improve'

it feels like screen scraping regurgitated.

also if you could clarify how data I input and tweak is reused. Your terms are a bit scary (basically if I write a book, you can 'publicly perform it and distribute it'?) per: By making any User Content available through our Services you hereby grant to Sassbook AI a limited, non-exclusive, worldwide, royalty-free, transferable license to access, view, use, copy, modify, publicly display, publicly perform and distribute your User Content to the extent reasonably needed to operate and provide the Services to you.

That wouldn't be something that could be considered "to the extent reasonably needed to operate and provide the Services to you".

There is no such intent anyway. I'll review it with the lawyer and revise it to something that doesn't scare people away!

I suppose this is applicable in general to content (as far as lawyers are concerned) and to protect us in case there is a need legal necessity for compliance purposes.

Thanks a lot for trying and providing feedback!

Each time you click complete, a different completion is usually generated, so that is one thing you could try.

Also provide a longer prompt (30-40 words is ideal) to express your intent more clearly, so AI can pick up it up (it needs help!).

To be fair, the original response might be something someone is thinking of writing. For example, a digital marketer.

Hug of death? All I get is:

> Oops... there was an error generating text

> The server encountered an error and could not complete the request. Please try again after some time.

Apologies for that.

Fixed it by adding more servers. Thanks for taking a look.

Almost every request looks like it has been improperly scraped from Reddit. A ton of queries I made have responses that look exactly like this:

> [Word for word what I typed in the prompt]

> Flair: Other

> Title: My friend is a girl

> Flair: Discussion

> Title: What are your favorite things about being in high school?

> Flair: Other

> Title: My friend is a girl and she likes me too

> Text: She told her friends about it

I cannot think of a context where including Reddit meta tags makes sense.

Reddit is a source for the corpus, so it is likely that the style matches it. More likely for some topics/styles than others.

Do you feel it somehow affects the quality of generations?

Thanks for the feedback. It'd be great if you can sign up and try it further and help improve the application.

I don't know why I would want to generate a line like "Flair: other" in a blog post. Why would the flair for a Reddit post even be in the corpus? Your app is billed as a writing tool, but this seems like a Reddit title and flair generator.

At this time, the engine has no knowledge about your context (blogging etc.). It picks up completions matching the topic and style.

I understand that you find this odd since you are not composing a Reddit post.

Again, very useful feedback; keeps us busy.

Thanks!

Prompt:

    > "There are many approaches to training a dog. One now-
    > disfavored method is referred to as 'dominance training'."

Suggestion:

    > This involves using positive reinforcement and rewards 
    > in order to train your dog to behave like you want him to.

Topical, if nearly the opposite of what should have come next :D

Tried a couple of writing prompts. Some generations were quite good and some rather inspiring.

I continued a few times but the quality dropped quickly. Clicking again does seem to generate different outcomes.

Will try some more later.

I'm curious about your business model. How do you plan to differentiate yourself from other very similar SAAS? Are you marketing to a particular segment?

Both on the general quality of generations as well as segmentation. We are still figuring out the priorities and domains.

Do you have any thoughts/suggestions?

Nothing at this point. I was just curious.

How does this compare to the other word spinners out there?

If you are talking about recent AI ones, should generally comparable favorably and should be better in some contexts (rather limited at this time).

We are a new startup, so hope to move ahead and stay ahead.

If you have experience with others, could you try and let us know what you find?

Thanks much.

Thanks in advance for feedback and suggestions for improvement. You can try it out freely.

"Oops... there was an error generating text The server encountered an error and could not complete the request. Please try again after some time."

Apologies for that.

This should be fixed now.

It doesn't work

Apologies for that. The server issues are now fixed. Thanks for (re) trying.

It sounds like a politician (though claiming not to be political) and is apparently anti-vax.

What do you want to write about? Coming together in this time of need

Prompt: It has come to my attention that this is a troubling time

It has come to my attention that this is a troubling time for the world and I would like to share some of what I've learned with you. The first thing you need to know about me is that I'm not an activist. In fact, if you Google me you'll find that I'm not even very political. I just want to make sure my kids are safe and healthy.

But when it comes to the vaccine debate, I've been a vocal critic for years. My first visit to a doctor was when I was six.

Oops... there was an error generating text

Apologies for that. The server issues are now fixed. Thanks for (re) trying.

Reply on news.ycombinator.com