Hacker News Reader

Comment by motoxpro | Hacker News Reader

Interesting that the large company AI products (this, Microsoft copilot, etc) are so much more compelling than any of the startups I’ve seen.

Makes me think that we haven’t seen what true innovation looks like in this space. Right now, AI is a feature, not a product.

Edit: Not talking about the models themselves (stability, mid journey, GPT-x), talking about what is built on those models.

View on HN

I would consider both Stable Diffusion and MidJourney to be startups and both are better by far than the established companies like OpenAI, and there are dozens of LLMs soon catching up with GPT3/4. As you say, it will be a very interesting next 6-12 months..

I played around (till free account expired) with MidJourney and what one can produce out of blue is mind-boggling. Indie devs can with a bit of effort generate much of their art via this for peanuts (the only problem may be consistency from what I've seen, basically every image is like from another artist, even in same batch)

The biggest problem is that you can't do much if not conditioning the generation with your prompt and maybe image embeddings, if you are an indie dev you would probably find much more useful Stable Diffusion.

If you control the --chaos --stylize and --seed then you can get more consistent results.

Stable Diffusion is a thing you can run on your computer for free and train models for. It's from a research university. Is this one of those more expansive definitions of startup?

Presumably he means stability.ai is a startup.

Stability AI who trained Stable Diffusion is a $100M startup. See https://www.crunchbase.com/organization/stability-ai

Dall-E is still ahead of the competition for some tasks. Midjourney has a very polished look but it lacks the depth of understanding that Dall-E can manage. I regularly hit prompts that I need to jump into Dall-E for.

For playing around, sure. For serious work though you probably need control-net in some way, otherwise you end up with a bunch of images which are great on their own but make no sense together.

Yeah, but text understanding is just one part of the "product" in this space nowadays. If you look at all the modes and features a GUI around SD can do now (like AUTOMATIC1111's) and compare to what you can do at the DALLE2 webpage it's night and day.

I wouldnt consider stable diffusion, mid journey and Open AI products. Little bit too low level, they seem to be more platforms that the products are built on. Not to say they aren’t amazing, just that the productized versions (one level up) are being executed really well by the big companies.

> AI is a feature, not a product

Part of the reason for this is that very often the best applications of AI are really invisible to the user, making everything a little bit better, quietly behind the scenes.

The same is true for software in general. The best software products automate loads of stuff behind the scenes so from the user perspective they just click a button and the thing they want to happen, happens.

Perhaps the most successful and important AI product is a great example of this: the spam filter. I bet most younger email users don't even realize how much value this old AI tool provides them unseen.

But the trouble is making effective AI products is not "sexy" right now, and if your team ships one it won't get any credit in a big company. I've had my PM instantly shoot down plenty of interesting applications of AI that wouldn't be very visible to the customer or larger product-org.

I don't fundamentally disagree, but if spam filters are "AI" then my bloody toaster is also artificially intelligent.

You must have a smarter toaster than I do. My toaster is a completely deterministic mechanism that knows nothing about my past toasting experience, nor makes any inference about future toast.

A spam filter:

- Is provided historic information about the problem

- Learns from this information to construct a basic model of language

- Is then provided unseen information

- Then makes a probabilistic decision based on a degree of uncertainty.

The exact model behind the spam filtering can be extremely simple (naive Bayes) but could easily (and probably will without you even realizing it) expand to include things like GPT.

A spam filter is making decisions under uncertainty with new information based on patterns it learned from previous information. If this doesn't fit your definition of "AI" then I think if you understood what was happening under the hood I don't think you would consider GPT to be AI either.

If you toaster does learn your toast preference over time from your toasting behavior (unlike mine) then I would consider that AI as well.

>A spam filter is making decisions under uncertainty with new information based on patterns it learned from previous information. If this doesn't fit your definition of "AI"

No, I don't. I would generously call it an "algorithm", but at the end of the day it's just a script with lots of if statements getting processed by an interpreter.

>I don't think you would consider GPT to be AI either.

I don't. "AI" in this context stands for Automated Interpolation. Nothing about this is intelligent.

Isn’t almost everything on this page concept stuff though? Seems like the only thing they’re shipping is text to image and the results they show look pretty underwhelming.

Seems like all these big companies are pushing amazing looking concept videos more than anything.

Yeah, I thought it was funny that most of the page is "in the future...". Guess they're trying to get attention for the product before it's ready and gauge interest/response.

Microsoft is just licensing their technology from OpenAI, which is still a small company, if not still a startup.

ChatGPT is not nearly as good of a product as what was demoed by Microsoft if you want to make that comparison.

Not talking about LLMs/underlying model. I don’t think adobe makes their own either. Talking about the interface to it.

What MS demo are you referring to? I like chatgpt way more than new Bing’s GPT interface.

https://www.youtube.com/watch?v=Bf-dbS9CcRU

Thank you. I hadn’t seen that. That is indeed pretty impressive.

Without the underlying model what's the point of the facade?

GPT-3 had been out for almost a year before ChatGPT and no one cared

Almost two years actually

> Interesting that the large company AI products (this, Microsoft copilot, etc) are so much more compelling than any of the startups I’ve seen.

This is an indicative signal. It's amazing how easily it is ignored.

> Makes me think that we haven’t seen what true innovation looks like in this space.

It's interesting that a bunch of product add-ons could be considered "innovation" in the first place.

In Adobe's case, it absolutely makes sense it would be a feature of Photoshop, Illustrator, and Premiere. They've been adding various "smart" masking features and the like over time already.

Reply on news.ycombinator.com