Midjourney is easier, its not better. The low barrier to entry has it popular, but it isnt as realistic, doesnt follow the prompt as well, and has almost no customization.
SD is the holy grail of AI art, if you can afford a computer or server to run SD + have the ability to figure out how to install python, clone Automatic1111 from git, and run the installer, its the best. Those 3 steps are too much for most people, so they default to something more like an app. Maybe it is too soon, but it seems SD has already won. MJ is like using MS paint, where SD is like photoshop.
Does being easier not influence whether it's better? I mean this in that for many of the ways AI art would be used, MJ already seems to be "good enough" at a lot of it.
Secondarily: doesn't Midjourney's increased user base and increased ratings they get from users help it refine its model, thus meaning that "ease of use" creates a feedback loop with "quality of output" because more users are engaged?
I'm asking real questions, not making a statement I believe in and just adding a question mark.
As mentioned, what is better, MS Paint or photoshop? If MJ ignores your prompt and spits out a half related picture, are you going to continue using it?
If anything MJ is a stepping stone to SD. You get a taste of AI art, but want to do something specific that MJ cannot do. You learn about control-net, alternative models, inpainting, etc... and you decide you need to move on from MS Paint to Photoshop.
I personally used free AI art(cant remember which), it was super cool, but quickly I wanted to use different models and generate thousands of pictures at a time. I wanted to make gifs, img2img, etc... and the only people doing that were on SD.
The question is not "which is easier?", but rather "which is easier to use to produce high-quality output". In your analogy, I'd argue the answer to that question is actually Photoshop. Likewise the answer in the MJ/SD case is MJ.
Metaphor is useful, but this feels overly-reductive. The gap between the amount of effort it takes to make something great or approaching the vision you had is massive between MS Paint and Photoshop. Not so for SD and MJ.
However, I am appreciating that SD seems to be clearly better if you need something more specific / precise. I don't think I'm convinced (yet) that because it can get more precise inherently makes it a better tool.
There are definitely some professionalized paradigms emerging in the use of SD: one video tutorial I saw this morning covering basic photobash + img2img and use of controlnet had a commenter saying that they preferred using the lineart function in controlnet to get more control and leverage their drawing skills.
When you see that kind of thing it's a huge signal for professionalization, because someone suggesting the optimal first step is "learn how to draw" deviates so completely from the original context of prompt-based image generation: "just type some words and the image magically appears".
Midjourney let's you type in a thing and get a result that will look great, which is no small accomplishment. If you want "incidental art" like blog post heros there is no competition. But it's really hard to use if you want to get exactly what you want.
Alternatively, if you don't know what you want, it's really good for inspiration.
In my opinion, it is tedious to discuss whether Midjourney or Stable Diffusion is better.
Both are tools with their own strengths and weaknesses.
Midjourney
- Comparable to a ready-made consumer product.
- Easy to use and you get excellent results quickly.
- The feature set is limited. So you can't influence so well how the result should look like.
- Restrictions on what you can generate with it (gore, nudes).
- Fee required.
Stable Diffusion
- Comparable to a development kit. Technical, Difficult to use, different tools that you have to configure and use correctly.
- For excellent results you have to invest much more time.
- Extensive feature set (inpaining, outpaining, upscaling, ControlNet, many different models, LORAs, textual inversions). You can much better define how the result should look like.
- No restrictions on what you can generate.
- Free if you install it locally.
https://www.runpod.io/console/templates
Look for "RunPod Stable Diffusion". I spent a whole $0.35/hr playing around with my own SD instance that I had running in minutes.
The main upside of running your own SD is that you can completely automate it, but I'm not sure how useful that really is.
No, the main upside of running your own SD web UI is that you can select and deploy your own checkpoints (not just using the base SD models), LoRas, embeddings, upscaling models, and UI plugins supporting additional services/models/features like multidiffusion (bigger gens and controls of which areas within the image different prompts apply to), ControlNet and associated models, video synthesis, combinatorial prompts, prompt shifting during generation to do blending effects, and, well, a million other things.
Also, you can completely automate it.
It's a little inconvenient to use non-base models and plugins this way (you pay extra for more storage), but it's definitely an easy way to use the full power of SD.
https://github.com/AUTOMATIC1111/stable-diffusion-webui
Or are you looking for the cutting edge stuff like control net?
If you want to use colab instead, I used this a month or two ago.
https://colab.research.google.com/github/TheLastBen/fast-sta...
I hope other people can give you further reading.
Even trained their own models using a cloud GPU.
The SD ecosystem is wild.
If you can afford Colab (which is free if you don’t want to use it too much), you can just click one of the existing A1111 colabs and run that, you don’t need to figure out python, git, or A1111 installs.
I was under the impression that midjourney was just running a form of SD and it's real secret sauce are the peripheral prompts it injects on the backend along with your prompts.
I could be totally off the mark here.
I could be entirely wrong, though. Maybe the person making that decision is just an idiot.
I have an IRC bot that has triggers for DALL-E, GPT3, and ChatGPT. I really want to make one for MidJourney, and I would happily pay MJ for the privilege. But I can't. Not without breaking some rules.
They're good enough technically at what they do that their audience is okay with the interface that they have to use, I'd reckon (I've heard similar beefs about the UI stuff though, so it sorta makes sense).
Would you mind explaining what you mean by that? Thanks!
The only way to do it right now would be to create a separate dedicated discord account linked to midjourney, and then have your bot control that user account and that's a good way not only to get your MJ access revoked but also to have your discord account banned.
As far as kneecaping their growth, they have one of the largest discord channels in history and they frequently run up against compute limits, so if this was one of their ostensible goals they failed at it pretty spectacularly.
AFAIK, this isn't possible.
Nothing so… social
By what measure? Midjourney v5 is massively better with every prompt topic I've thrown at it than SD. It's not even close. SD, however, is much better if you want an actual customizable toolchain or to do things like train it on your own face/character.
https://paperswithcode.com/sota/text-to-image-generation-on-...
Parti and Imagen are still on top, followed by Dall-E 2.
If their model is so great, why are they afraid of benchmarks?
For a while I was using an FID variant for evaluation during training, but didn't find it very helpful vs just looking at output images.
A 1000 V-rolls won't get you there. For something like this control net combined with inpainting is indispensable. Not to mention the excessively heavy handed censorship in MJ.
Midjourney excels in overall quality, but it completely falls down if you have an actual complex vision.
Here have another prompt: "Rapunzel has let her hair all the way down a tower. The hero has been tied up by the witch, and annoyed at Rapunzel's continual attempts to escape, the witch throws the bottom of her hair into a paper shredder at the base of the tower."
You could v-roll until the heat death of the universe without even getting close.
Midjourney is great if all you're capable of conceiving is 90s Mad Magazine templatized mad Libs, banal crap like "Darth Vader as a French pantomime street artist".
Unfortunately that also describes the majority of midjourney users.
https://www.bing.com/images/create/a-monk-playing-chess-agai...
https://www.bing.com/images/create/a-blue-book-on-top-of-a-y...
Simple (prompt only) use of generative models is quite good at creating simple artistic pictures you might actually hang on a wall. Trying to create a complex scene with just a prompt seems still a few years off though.
I use SD to augment MJ, like fixing hands with specific LORAs for example, so I definitely appreciate that it exists. But for actually creating a full image in one shot, they're not even comparable.
I've twice spent a couple of hours unsuccessfully trying to generate a simple background image that would be blurred out when rendering 3D models. SD out-of-the-box was far worse, but Midjourney still was not up to the task. It's incredible how well they can generate images of nearly any subject/object and make some changes to the style and placement, but trying to precisely achieve critical broad-stroke things like like perspective, sizing, lighting direction/amount/temperature, etc. was far too cumbersome. Prompt refining is just like having a program with a bunch of nebulous undocumented menu entries that you just have to click on to see what they do rather than just giving you the tools to make what you need to make. Was that the right entry or the wrong entry? Who knows! Maybe just try it again to see if it works better!
There's a fundamental disconnect between professional-level and consumer-level tools. Consumer tools must be approachable, easy to use, quickly yield appealing results, affordable, and require little maintenance. Professional tools need to be precise, reliable, capable of repeated results with the most demanding tasks, and easily serviceable into perfect working order.
These are consumer-level tools. If you merely need a cool picture of a space ship done in such and such style with such and such guns blah blah blah (that for some reason always looks 10%-50% Thomas Kinkaid,) these tools are great, but they abstract away the controls that really matter in professional work. Novices who get overwhelmed by all of those factors love it because they don't understand, and probably don't care about what they're giving up. For serious work, aside from getting inspo images or maybe generating deliberately weird bits of whatever, they're hit-or-miss at best. Without exception, doing a rough mock-up in a modelling program took FAR less time than trying to wrangle exactly what I needed from one of those generators.
I'm sure they'll get there someday but right now they're miles away from being professional-quality image generation tools.
I think a lot of people have unrealistic expectations of the tech-- they think they can get exactly what they want if they are articulate enough in describing it with words.
Feed your rough mock-up to img2img (or use inpaint sketch) and you'll land much closer to where you're trying to go.
It's a power tool. It will do tedious manual work (producing art) very quickly. The difference between professionals and consumers in how they use it is that the professional asks the machine to "finish what I started," whereas the consumer tells the machine to "do all of the work for me."
Canva fits most people's needs but it only fits the needs of a narrow slice of professionals. Trying to use it for professional work takes vastly more time and effort than using a proper professional tool. LaTeX fits academic paper publisher's needs and can pump out formatted and laid-out text far quicker than someone using InDesign but you'd go crazy trying to assemble a modern magazine or high-end book. It doesn't need polish or sheen. It needs something fundamentally structurally different.
I'm holding out for an instruct-based model that will take explicit instructions, or at least layered prompts. Mutating the original prompt along with the picture (or changing the entire thing to only describe certain parts, a la inpainting) is frustrating to me.
I think what most developers don't realize is that creating media, even snazzy, captivating images, is a tiny portion of the work commercial artists and designers do beyond Fiverr or the people working at sign shops. We've got bazillions of stock photos, asset stores, etc. etc. etc. available at our fingertips at a pretty low cost... and we use them, just as we'll use AI generated images in our processes... Just not for anything that actually matters. Reducing art and design to what so many people have is nearly akin to reducing coding to how fast someone can generate code regardless of it's suitability for the purpose. I could go into Adobe Illustrator and make most of the vector art you see on the net in a few minutes. Knowing what needs to go on the screen, exactly how it needs to go on the screen, and the effects of putting it there, what it communicates and to whom, etc. etc. etc. are the hard parts. You could have a tool that would instantly translate millions of pretty images from someone's imagination in a matter of milliseconds but that's not going to make them useful for communication or anything else.
So very much of this hype is people simply assuming that they understand something that they don't. And that's why the UIs for most FOSS applications suck yet the project contributors will defend them like their own children.
Who's fault is this though? The hype is absolutely hysterical.
We already now see deepfakes (e.g. of Trump or the Pope, recently even videos) that a far beyond what we saw in the years before, indicating that the old professional tools weren't so powerful after all. Now if we extrapolate this a few years into the future...
As I mentioned in another comment, so much of this hype is based on developers assuming they understand something that they don't. I've indulged in this hubris as a developer but straddling both sides of this line has been illuminating.
https://www.adobe.com/sensei/generative-ai/firefly.html
This sounds all very similar to the Stable Diffusion tool chain, though probably cloud based and with a more intuitive UI on top.
Being able to quickly generate and iterate on assets is great for inspiration but pretty useless for professional output without fine-tuned, predictable, repeatable controls. These tools will simply integrate with existing professional tools until they can do it better.
Imagine the first person to make an electric saw made some automated thing that could cut the wood to make a cool looking flat pack house somewhere in the neighborhood of your specifications in 5 minutes. The caveats: while it would assemble perfectly, the actual angles of the cuts might be unpredictable... Like 40 and 50 degrees rather than 45 and 45, and the layout was never quite what you expect even if it was OK more often than not. Pros knew those were fundamentally deal breakers for professional work, and remained more professionally useful with their hand saws because they had the required precision, control, and predictability. While enthusiasts were gong crazy exploring all of the different kinds of oh-so-slightly wonky structures they could generate and predicting the end of carpentry, the old school saw companies started making circular saws, chop saws, drills, and the like. The market for handyman-built dog houses, sheds and playhouses would immediately be lost to the automated machine but I guarantee you that all consequential work would still be done by carpenters with power tools.
Midjourney definitely generates really high quality art based on simple prompts, but the inability to really customize the output basically kills its utility.
We heavily use Stable Diffusion with specific models and ControlNet to get customizable and consistent results. Our artists also need to extensively tweak and post-process the output, and re-run it again in Stable Diffusion.
This entire workflow is definitely beyond a Discord-based interface to say the least.
https://twitter.com/P_Galbraith/status/1649317290926825473?c...
This is using Stable Diffusion and the Control Net Lineart Model. The coloured version is pretty rough but it was a quick test.
In my opinion Stable Diffusion is vastly superior to Midjourney if you have the skill to provide input to img2img/ControlNet.
I have some other earlier workflow experiments on Youtube if you're interested in this kind of thing https://www.youtube.com/pjgalbraith
I'd be interested to know where the art ends up in the game? Do you mean 2D backgrounds and billboards in-game? Or are we talking cut-scenes and menu screen art?
Excited by the potential for artists to accelerate their creative process with advances like ControlNet, but makes sense there's a lot frustrating about the process today.
Exploring tools that might help. If you might be interested in chatting, email me at [email protected]. Thanks!
None of this stuff is copyrightable so I dont care that its not private
Depends. If you want a script that will summarise and read your news feeds in the morning, or your email, you might prefer the slightly less intelligent but private models.
Or if you are batch processing NLP jobs, then you're better off with a fine-tuned small model (<15B). Faster, private, no restrictions.
API access is on waitlist.
UI has limit of 25 messages in 3 hours.
If you think big, known companies can get ahead of the waitlist and use it - short answer is no, they can't because of their IP. Nobody is going to sign off leaking out all internal knowledge to play with something.
ClosedAI seems to have big problem with capacity.
Those poems about your colleague's upcoming birthday do burn a lot of GPU cycles.
- Some scalability breakthrough will appear, if that's the case their moat disappears pretty much instantly and the cost of LLMs will plunge close to zero being a commodity. That's the future I'm betting on from what's happening now.
- No scalability breakthrough will appear and then it means that they will have a hard time to expand further as seen as the gpt4 limited access.
Either way, they are in a tough spot.
"End user computing" these days means mobile, and mobile is dominated by Linux (in Apple's case BSD, but we're splitting hair) and Chrome/WebKit - which began as KHTML.
The only area where opensource failed is the desktop, and that's also because of Microsoft's skill in defending their moats.
In this way, everyone's skills are being leveraged to innovate at a rapid pace.
Best only works till second best is "close enough" and cheaper/free
It's the internet all over again. How do you win the race to the bottom?
Once there
How do you compete effectively with free? Microsoft and Amazon will have billions on billions coming in to float their free offerings for what is effectively eternity in business terms. Probably Google and Meta will as well. What happens to everyone else?
I think you have to be in some niche market where you can charge. Because for everyone else, free is unsustainable.
Porn maybe? But there will be way too many competitors there. So something more like medical. Or semiconductors. Or construction or something.
That what I was getting to. Paid only makes sense if you're willing to provide stuff that OSS lacks, which is either "super specialized things not many people want to OSS" or, well good looking UI... (there seem to be massive lack of any UI/UX people vs developers in near anything OSS).
AI is neither so it will be commoditized and mostly run in few OSS projects, and probably for the best, the only thing worse than anyone having access to "near free copywriter bot that will write about anything you tell it to" is only people with money having access and control over it.
Both take zero technical knowledge to use the base models via first-party online hosts, but Midjourney is superior there.
SD offers a lot more capacity beyond what is available from the first-party online host, though, while with Midjourney, that’s where it begins and ends, there is nothing more.
> and doesn’t require a high-end gaming PC to run it.
Neither does SD (I use a couple-year-old business laptop with a 4GB Nvidia card; no sane person would call it a “highend gaming PC” to run A1111 locally, and there are options besides running it locally.)
I wonder why nobody is talking about Bing Image Creator
https://www.bing.com/images/create
which uses some much more advanced version of Dall-E 2 in the background (so Dall-E 2.5? 3?), while being completely free to use. It can produce some pretty mind blowing results with quite simple prompts, although apparently not as impressive as Midjourney V5. A few examples:
hyperrealistic https://www.bing.com/images/create/hyperrealistic/644fa0c48f...
an allegory for femininity https://www.bing.com/images/create/an-allegory-for-femininit...
portrait of a strange woman, hyperrealistic https://www.bing.com/images/create/portrait-of-a-strange-wom...
allegory of logic, portrait https://www.bing.com/images/create/allegory-of-logic2c-portr...
her strange bedfellow https://www.bing.com/images/create/her-strange-bedfellow/644...
Mrs fox https://www.bing.com/images/create/mrs-fox/6446e85a32134e649...
inside view https://www.bing.com/images/create/inside-view/6446f1dc573f4...
in the midst of it all https://www.bing.com/images/create/in-the-midst-of-it-all/64...
strange gal https://www.bing.com/images/create/strange-gal/6446e2a2ea7a4...
sighting of a strange entity in an abandoned library https://www.bing.com/images/create/sighting-of-a-strange-ent...
sleeping marble woman next to a wall of strange pictures inside an abandoned museum, close-up https://www.bing.com/images/create/sleeping-marble-woman-nex...
sculpture of a woman posing next to a wall of strange pictures, close-up https://www.bing.com/images/create/sculpture-of-a-woman-posi...
Easter https://www.bing.com/images/create/easter/643ae4968aff432684...
Christmas on board a spaceship, DSLR photograph https://www.bing.com/images/create/christmas-on-board-a-spac...
an angel, dancing to heavy metal https://www.bing.com/images/create/an-angel2c-dancing-to-hea...
Saturday afternoon in the streets of a buzzing cyberpunk city, photo-realistic, DSLR https://www.bing.com/images/create/saturday-afternoon-in-the...
The Dogfather https://www.bing.com/images/create/the-dogfather/6441d18950b...
the unlikely guest https://www.bing.com/images/create/the-unlikely-guest/644446...
Strange pictures in an abandoned museum https://www.bing.com/images/create/strange-pictures-in-an-ab...
strange woman in an abandoned museum, close-up https://www.bing.com/images/create/strange-woman-in-an-aband...
strange woman in an abandoned museum, strange pictures in the background https://www.bing.com/images/create/strange-woman-in-an-aband...
a wall of strange pictures in an abandoned museum in Atlantis, close-up https://www.bing.com/images/create/a-wall-of-strange-picture...
female sculpture in an abandoned museum in Atlantis, close-up https://www.bing.com/images/create/female-sculpture-in-an-ab...
the unlikely guest https://www.bing.com/images/create/the-unlikely-guest/644490...
an unlikely guest of the secret society in the lost city in a country without name, close-up https://www.bing.com/images/create/an-unlikely-guest-of-the-...
I think the quality of most of these pictures is far beyond what is achievable with Dall-E 2. One issue that still exists (though to s lesser extent) is the fact that faces have to cover a fairly large area of the image. Smaller faces look strange, e.g. here:
photograph of the unlikely guests https://www.bing.com/images/create/photograph-of-the-unlikel...
It is as if the model creates a good draft in low resolution, and another model scales it up, but the latter model doesn't know what a face is? (I have no idea how diffusion models actually work.)
Probably not the only factor but could be one.
https://www.theregister.com/AMP/2023/03/16/ai_art_copyright_...
So the images can also be used commercially, and this holds as much for Microsoft as for Midjourney. (Though note that Microsoft isn't denying the usage for commercial purposes, though they made it sound as if they did.)
Pardon the side discussion, but I think this is because of a few things.
1. OS-exclusive "killer apps" (Office, anything that integrates with an iPhone)
2. Games
The killer apps have better alternatives now, and games are starting to work better on Linux. Microsoft's business model no longer requires everyone to use Windows. (Mac is another story.) So I think that, at least for non-Macolytes, Linux end user dominance is certainly on the horizon.
I kid. I've been primarily a Linux Desktop user for 20 years.
TBH this feels like cope from Google; Bard is embarrassingly bad and they expected to be able to compete with OpenAI. In my experience, despite their graph in the article that puts them ahead of Vicuna-13B, they're actually behind... And you can't even use Bard as a developer, there's no API!
But GPT-4 is so, so much better. It's not clear to me that individual people doing LoRa at home is going to meaningfully close the gap in terms of generalized capability — at least, not faster than OpenAI itself improves its models. Similarly, StableDiffusion's image quality progress has in my experience stalled out, whereas Midjourney continues to dramatically improve every couple months, and easily beats SD. Open source isn't a magic bullet for quality.
Edit: re: the complaints about Midjourney's UI being Discord — sure, that definitely constrains what you can do with it, but OpenAI's interface isn't Discord, it has an API. And you can fine-tune the GPT-3 models programmatically too, and although they haven't opened that up to GPT-4 yet, IME you can't fine-tune your way to GPT-4 quality anyway with anything.
"There's no moat" and "OpenAI is irrelevant" feel like the cries of the company that's losing to OpenAI and wants to save face on the way out. Getting repeated generational improvements without the dataset size and compute scale of a dedicated, well-capitalized company is going to be very tough. As a somewhat similar data+compute problem, I can't think of an open-source project that effectively dethroned Google Search, for example... At least, not by being better at search (you can argue that maybe LLMs are dethroning Google, but on the other hand, it's not the open source models that are the best at that, it's closed-source GPT-4).
I'll also point out that paid api access to 3.5 (davinci-03) is frequently better than ChatGPT's use of 3.5. You get many fewer restrictions, and none the "awe shucks, I'm just a little 'ol LLM and so I couldn't possibly answer that".
If you're frustrated by having to go to great lengths to prompt engineer and ask ChatGPT to "pretend" then it's worth it to pay for API access. I'm just frustrated that I can't use the GPT-4 API the same way yet (waitlist)
I pay the $20 for ChatGPT Plus (aka, GPT-4 web interface access); personally I find it useful enough to be worth paying for, even in its rate-limited state. It already replaces Google for anything complex for me. I wish I could pay for the API too, and use it in my projects.
Often times it can just recite things from memory that Google can't even properly link to, and they've got a proper index to work from for fucks sake.
Google has the technology and talent to relaunch themselves in a leadership position, but the current executive team doesn’t seem to have what it takes. They’re custodians / accountants, running the company a bit like Microsoft in the Ballmer era. What google needs to do now is leap ahead, and I don’t see it happening without a leadership change.
It's not exactly a shortcut, and maybe it was just luck, but I suspect the key is just to start building with what you have and show a trajectory. The best part is that coding with ChatGPT-4 as a "colleague" has made the whole thing super fun.
Maybe it’s a bit too paranoid, I don’t know, but I’ve also been open here about my workplace experiences, if someone who knew my irl connection to them and decided to comb through my comments, in a way that my HR dept among others might not quite appreciate. Maybe I should setup a separate HN account connected to me Professionally for that sort of thing.
Also my use case for GPT-4 is data analysis. Using the paid “plus” version shows a lot of promise for quickly bootstrapping data exploration and consumption as a jumping off point for more detailed digging. Via the chat interface it can ingest very small aggregate datasets and spit out observations that only myself and my boss have the domain name expertise to produce in my organization. but the Chat interface is highly limited and often truncates even small (faked but plausible) data, so I really want API access, because it involves sensitive info I couldn’t put into the chat site or responsibly share with someone outside my org.
But really, thanks for the offer. What are you working on with it?
The wait for GPT-4 is not as long as it used to be, and when you're using the API directly there's no censorship.
I’ve been on the GPT-4 waitlist for about 6 weeks, but I’m not sure what the typical wait is.
Wait seriously?
Testing things out, it will produce vile and hateful content on demand. However it won’t say anything. If I specifically tell it to use some words I get the same type of content warning but also an extra note about those words, and that I have to contact openai support if my use case truly requires their use.
There’s also the fact that using or distributing such content is against TOS, so I suppose could simply ban you
Though they could still use "observational" censorship there, i.e. analyze your prompt with a different model. OpenAI does that, not sure about Microsoft.
Could you please describe how one "just" do that? I have been on the GPT-4 API waitlist since it was announced and I still don't have access to the GPT-4 API.
Little correction - 3.5 is not davinci. davinci is 3.0, 3.5-turbo (chatgpt) is a davinci variant that has been tuned and adjusted for chatting and conversation, including all those restrictions. It is much faster than davinci, way cheaper but as you know, results are… ok
davinci (3.0) is more untuned, slower, more expensive to use, not conversational, but can yield much better quality
Little correction of the correction. The base models are:
davinci = GPT-3
code-davinci-002 = GPT-3.5
They do only text completion and do not natively answer to instructions. There are also instruction tuned versions of the latter, e.g. text-davinci-003 and gpt-3.5-turbo-0301 (used in ChatGPT). See
https://platform.openai.com/docs/model-index-for-researchers
Note that code-davinci-002 is no longer available via the OpenAI API, but it is still on Azure. The GPT-4 base model is generally unavailable. Too powerful perhaps.
Thanks you for the offer, but I’m extremely conscious of avoiding a direct link from my comments here to who I am. Maybe it’s a bit too paranoid, I don’t know, but I’ve also been open here about my workplace experiences, if someone who knew my irl connection to them and decided to comb through my comments, in a way that my HR dept among others might not quite appreciate. Maybe I should setup a separate HN account connected to me Professionally for that sort of thing. Also my use case for GPT-4 is data analysis. Using the paid “plus” version shows a lot of promise for quickly bootstrapping data exploration and consumption as a jumping off point for more detailed digging. Via the chat interface it can ingest very small aggregate datasets and spit out observations that only myself and my boss have the domain name expertise to produce in my organization. but the Chat interface is highly limited and often truncates even small (faked but plausible) data, so I really want API access, because it involves sensitive info I couldn’t put into the chat site or responsibly share with someone outside my org. But really, thanks for the offer. What are you working on with it?
GPT-3.5, I think it is mostly suitable for:
1. Quick documentation lookup for non-essential facts
2. Lightweight documents writing and rewriting
3. Translation
Other use cases should go straightly to GPT-4
"In the context of [field I'm ramping up in], what does X mean, and how is it different than Y" -- it's not as good as GPT4 but it emits so much quicker and it normally gets me where I needed to go.
OpenAI spent 600m to improve GPT and made 200m from it and if costs dramatically fall for model development, it might be OpenAI that is shooting itself in the foot.
How do you know?
There is an API for the underlying model, it's just in alpha/beta/whatever they call limited invite-only release and you have to ask your devrel team to get access. I'm guessing we'll see better models very soon.
Google is, as usual, playing catch-up, but I have no doubt once the machine gets cranking they'll be fully competitive, at least similar to how GCP is now a totally viable AWS alternative. They never lead the pack because they can't (lawyers, regulation, monopoly, etc), but they know how to commit and execute.
MJ only does one style, and you can emulate that just fine in SD if thats what you want.
It seems just about every computing appliance in my home runs Linux. Then you have Android, ChromeOS, etc. which are also quite popular with end users, the first one especially. It may not have won, but I think it is safe to say that it is dominating.
They are when the end user is using them. Think things like TVs or even thermostats.
> the OS is incidental and under full control of the manufacturer.
For all intents and purposes Linux has won where those conditions aren’t met.
They did manage to get large buy in from the company after quite a significant journey. But it seems so much like a kind of outside event, something begat & pushed for not because it was a smart top down move, but because a couple super driven engineers made it their cause.
of course google doesnt want to acknowledge it too much.
Not me, I refuse to use OpenAI products but I do sometimes use vicuna 13b when I'm coding C. It's pretty good and I'm happy to see the rapid advancement of open source LLMs. It gives me hope for the future.
> Linux won in servers and supercomputing, but not in end user computing.
I use linux on all of my computers and I love it, many of us do (obviously). I'm aware that I'm a small minority even among other developers but I think looking at just statistics misses the point. Even if the majority will just use the most approachable tool (and there is nothing wrong with that), it's important to have an alternative. For me this is the point of open software, not market domination or whatever.
Open AI had years of advanced that almost vanished in a few months.
And we will see the rise of specialized models, smaller but targeted, working in team, delegating (Hugging GPT)
I would use a small and fast model that only speak english, is expert at coding an science and not much more. Then you fire up an question to another model if yours is out of it’s area.
Look at the first uses of StableDiffusion. It was either "you know python or you use Dall-E". Now we have one-click installers setting everything up, and nice user interfaces on top of it.
Best isn't defined just by quality though. In some instances for some groups, things like whether the model is trained on licensed content (with permission) and / or is safe for commercial use is more important.
This is one reason why Adobe's Firefly has been received relatively well. (I work for Adobe).
No, what the article said was:
> At that pace, it doesn’t take long before the cumulative effect of all of these fine-tunings overcomes starting off at a size disadvantage.
>Indeed, in terms of engineer-hours, the pace of improvement from these models vastly outstrips what we can do with our largest variants, and
> the best are already largely indistinguishable from ChatGPT.
^ The author did not note chat gpt is better, the author claims that the 7B koala model is 'largely indistinguishable from ChatGPT'.
and:
> While ChatGPT still holds a slight edge, more than 50% of the time users either prefer Koala or have no preference.
Which is highly misleading.
The koala authors rated their model by passing it to 100 people using the mechanical turk, noting:
> To mitigate possible test-set leakage, we filtered out queries that have a BLEU score greater than 20% with any example from our training set. Additionally, we removed non-English and coding-related prompts, since responses to these queries cannot be reliably reviewed by our pool of raters (crowd workers).
So.
What you have is a model that performs pretty well for some trivial conversational prompting tasks.
What you DO NOT have, is something that is: "largely indistinguishable from ChatGPT".
Anyway, regardless of the creative interpretation of the authors writing, the point that I'm making is that your point:
> So, the winning strategy is whatever strategy allows your model to compound in quality faster and to continue to compound that growth in quality for longer.
Is founded on the assumption from the post that:
> While the individual fine tunings are low rank, their sum need not be, allowing full-rank updates to the model to accumulate over time.
ie. If you fine tune it enough, it'll get better and better in an unlimited fashion.
Which is provably false.
If I have a 10-parameter model, there is no possible way that the accumulation of low rank fine tunings will make it the equivalent of a 7B, 13B of 135B model.
It is simply not complex enough to do some tasks.
Similarly, smaller models like 3B or 7B model, appear to have an upper bound on what is possible to achieve with them regardless of the number of fine tunings applied to them, for the direct and obvious same reason.
There is an upper bound on what is possible, based on the model size.
The 'best' size for a model hasn't really been figured out, but... I'm getting pretty sick of people saying these 7B models are as good as 'ChatGPT'.
They. Are. Not.
People will go to the best models, with the best licenses, but... those models are, it seems, unlikely to be fine tuned smallish models.
The author talks about Koala but notes that ChatGPT is better. GPT-4 is then significantly better than GPT-3.5. If you've used all the models and can afford to spend money, you'd be insane to not use GPT-4 over all the other models.
Midjourney is more popular (from what I'm seeing) than Stable Diffusion at the moment because it's better at the moment. Midjourney is closed-source.
The point I'm wanting to make is that users will go to whoever has the best model. So, the winning strategy is whatever strategy allows your model to compound in quality faster and to continue to compound that growth in quality for longer.
Open source doesn't always win in producing better quality products.
Linux won in servers and supercomputing, but not in end user computing.
Open-source databases mostly won.
Chromium sorta won, but really Chrome.
Then in most other areas, closed-source has won.
So one takeaway might be that open-source will win in areas where the users are often software developers that can make improvements to the product they're using, and closed-source will win in other areas.