nutanc

It is a well explored territory and no it can't. Especially in summarisation, it can sometimes fail suddenly and unexpectedly.

You are right. Humans will move on to building more high quality images. But for regular run of the mill stock images, AI is already there. I had done a small experiment to create stock images[1] [1] https://medium.com/ozonetel-ai/generating-a-landing-page-wit...

"girl with a cap standing next to a shadow man having a speech bubble, digital art"

And I just used it to create cover art for a book published in Amazon :)

https://twitter.com/nutanc/status/1549798460290764801?s=20&t...

Have been thinking about this idea. Are there any open source implementations of diffusion models which take an image as an input?

The embeddings provide the guidance for the diffusion models. So they have a direct influence. Similarity of the sentences means that they are around the same area in the multi dimensional space. And when they are used to guide the diffusion models, the images come out wrong. One thing is the experiments I have done are on sentence transformer models. Dalle-2 uses its own sentence embedding afaik.

Dalle2 is really good for stock marketing content. Your use case is certainly feasible. Especially with the edit feature we can make dogs eat your product. If you can DM me a product image I can show a POC :)

The text is picked up from https://storyweaver.org.in/stories/98-rani-s-first-day-at-sc... for this example. But I have experimented with GPT3 generating content and Dalle generating the image. For example here, https://twitter.com/nutanc/status/1538372995764617216

I tried this. Every time a new girl pops up :)

I tried a lot, but thats one shortcoming. You can try editing and redrawing. Some success is there, but scope for improvement is there.

I would associate a bird with a bird. I wouldn't associate birmongle with bird just because both start with the token "bir". What Dalle is doing here is what it has been trained to do. Just find the closest token match and try to draw something. It does not understand something. And that's absolutely fine. I am just saying it does not have a secret language or something. It's just regular language and it's closest matches.

I don't think it's a secret language per se. It's just that the tokens generated for these sentences are for some reason coming close to a bird latent space. Maybe if we can dig deep and do a google search for kinds of birds we can find the connection. Tokens from OpenAI below.

https://t.co/Of8CBGdGAE.

Found this answer:

https://twitter.com/BarneyFlames/status/1531736708903051265?...

Slack was supposed to cure email. Instead it created a new "mania"?

As someone said in the comments, the results have to be cherry picked, but the cherry picking is at a pretty low level. I run an AI guru bit at https://twitter.com/guru_ai and I manually select the output. But most of the times, the AI does a pretty good job.

That said, it will be a reach to say that the AI has a sense of humour. Because fundamentally the AI does not know how to laugh.

For example, the AI sticks to the script. Let's say you are having a serious discussion, it will not crack a joke to break the tension.

Now someone just train a model using these questions and answers and we will let the model take all future interviews.

A small experiment that I am trying out. Follow and get your spiritual and startup gyan from an AI :)

This is now popular enough that the owner should probably create an NFT out of this :)

It's not just that. As a developer I wouldn't want Apple to know who my paying customers are. Right now, there is no option. With this hopefully there will be more options.

Trying out an experiment, creating a substack with purely AI generated content. GPT3 is pretty good for buzzfeed kind of content.

https://aifeed.substack.com/

I have used GPT-3 and it works for most of the time. But it fails some of the time too. And thats the problem for use cases like programming or generating config files. Because if you cant trust the output 100% you are pretty much reading the output every time.

So, the only time save is that GPT-3 makes you type less.

In any case I don't type much anyway now a days. Its mostly copy paste from stack overflow update parameters etc.

GPT-3 will be useful, maybe a year from now.

So is this the equivalent of saying,"My AI program is a bunch of if else statements."

Yes.

I have recently started an experiment to generate an AI generated newsletter[1]. All posts are generated by GPT-3. I work as the editor. It works well for some topics and not so well for some topics. Since I curate the content, I dont publish topics which are not done well. For example, I tried to make it generate a nice article on the Suez canal crises. But it was harder than I thought it would be.

It generates buzzfeed kind of stories very well though :)

[1] https://aifeed.substack.com/

I am actually from India. The bank mobile websites are good enough. But the apps have some good user flow for P2P payments etc. But yeah, I can live with a good mobile website.

I want a Linux phone. I want to buy something which I can just own and do what I want with it without some overlord deciding that I have violated some terms and conditions. All I need is for some basic apps like email, Maps(OSM is fine),Spotify, Twitter, WhatsApp/Telegram and my banking app to work. It covers 90% of my use cases of my phone.

In any case we all have more than one phone in our house. An extra Linux phone will give me and my family options. The junk stuff and games etc can go on my Android.

I applied online and got access at the initial stages.

Launching the world's first AI generated news letter(https://aifeed.substack.com/). All the content is generated by AI. Subscribe and get original AI content regularly. Watch out Buzzfeed, here come the AI overlords :).

This is powered by GPT-3. I have made some custom prompts to generate the article content. But this is more of an experiment to see what kind of content can be generated and will it keep people occupied. I will be regularly updating the newsletter with content.

All the articles on the site have been generated by AI. I just gave the first line maybe and in some cases,adjusted a sentence here or there. Overall, for all the articles on the site now it took me around 20 minutes to generate.

How much for an NFT for this comment :)

Multiple things to think about here:

1. For summarizing, in my experience GPT-3 still has some ways to go. It gets it right a lot of the times, but when it misses, it misses bad.

2. Assuming that after scraping we feed all the scrapped data as a prompt from which GPT-3 generates bullet points, that will be a very big prompt. Since prompts are also counted as tokens, it might end up costing 10cents minimum to generate one summary.

3. I think the core USP from the process the OP has detailed is in steps 1,2 and 3. Step 4 is a good hook to get people to try this out, but have to test it properly and check the costs also.

If anyone is interested :) Output from GPT-3:

Detritus and Carrot walked into a dwarf bar.

"How do we get in?" asked Carrot, looking around.

"That's a good question," said a dwarf. "You look like you're from far away."

"We are. We're from Arbouretum."

"That's a long way away."

"That's where the Arbouretum are from."

"The Arbouretum?"

"Yeah."

"What does that mean?"

"It's a long story."

"I've got time."

"I've got no time."