Andreas Thiel
Silver Contributor
FASTLANE INSIDER
Read Rat-Race Escape!
Read Fastlane!
Read Unscripted!
Speedway Pass
"Midjourney, Inc" is an intereting company to take a look at. I think diffusion models for asset creation have more short term potential than LLMs do.
Midjourney might be in a critical phase right now.
They do focus on the product and listen to their users. They do micro-polls on their Discord server. The founder lurks in the Discord channels.
It seems like they believe in the concept of a productocracy.
For those who want to get some behind the scenes impressions:
The founder David Holz (who I think had 11+ years experience running Leap Motion before working on Midjourney) does a weekly 2.5 hour long and public "Office Hours" call on Discord each Wednesday, where people can ask questions "on stage" if they are lucky and in "lightning rounds" by typing in a chat room after they go though a short progress summary for the past week. Apparently, he does that after a 2.5 hour long company internal one in back-to-back sessions.
CENTS:
Control - might be a problem, quite literally in the image creation process. As the models get bigger and more sophisticated, the results get better. There are fewer messed up hands etc. But text prompts currently offer only so much control over the results. Soon there will be inpainting, where you will be able to tell Midjourney to repaint a part of an image that you don't like. They throw away many trained models, because the results don't meet their standards.
Stable Diffusion tries to give the creators more control, for example with ControlNet that takes scribbles or poses into account, but the results are not great when I try it. Others seem to be more lucky.
They went with a Discord based interface, because people liked it most in an initial split test. Now they run into painful limitations and are trying to gain more control again by developing a web frontend and a mobile one.
Laws and copyright concerns could become an issue. Activists could become an issue.
On the other hand, they are bootstrapping the business and have rejected money from investors. They were also successful in developing an AI Moderation tool that made sure that their showcase does not get flooded with problematic content.
Entry - is working in their favor. The diffusion model papers are pretty accessible and there are books that explain the basics and even show code examples. But they have built out a crazy hardware infrastructure. I think they said that each job that gets started is distributed to 10 GPUs. Training models takes a lot of time. They had to fix countless issues. There are many people that cross the 10k image generation mark. The competitors probably don't come close to that level of mileage and experience.
Need - Need might be a concern. What I would call a "bold move" is that they don't seem to go the OpenAI route by offering API access to create a 3rd party marketplace. The Holz vision seems to be becoming a community and a platform. He considers Instagram the competition, not OpenAI. They want to add story telling capabilities and they want their "Showcase" page to beat the "Best Of Instagram". That might become interesting. People want to take their images elsewhere and they can do that. I wonder if people actually "need" Midjourney as a community / platform that showcases their creations. If they force this artificially, this could become a huge backlash story.
To compete with Instagram, they need people who are not creators but join just to look at pretty pictures. Will that work? There seem to be enough people signing up for the $48 per month plan and the mini-polls showed that there is demand for $200+ dollar tiers, so creators see a need. But what features will they demand in the future?
The platform focus also means that they are moving in the direction of creating "opinionated" - flashy and attention grabbing - pictures by default. Not everybody appreciates that.
Time - Here they are probably golden. Things are automated. Quote: "Time honored business are usually involved in the act of creation.". And Midjourney does create.
Scale - They seem to be able to scale things up pretty well. The is a "status" channel on their Discord server and they fix the service and bot issues relatively fast. I am pretty happy with the uptime.
Additional thoughts:
From the "Office Hours", it is apparent that David Holz is working towards a sale of Midjourney. He talked about which Startups he admires. All were companies that produce physical things or companies that do weird things like trying to establish communication with Beluga whales. He also said explicitly that he wants to start a more crazy business, probably space related and that Midjourney is just something that happens to be doable right now.
Competitors and Open Source will catch up. Other companies will offer APIs. While they are looking into 3D and animation, these are in a very early stage of development. If his backlog is an indication, he seems to be mentally prepared for running the company for another 5+ years, though.
Are the numbers inflated? Are there growth concerns? A personal observation is that, for a while, it became more difficult to get into the top 2000 when ranking images in order to get an additional "fast hour" (I pay $8 per month for 3.5 fast hours and can get an additional one pretty much each day by spending 20-30 mins ranking images). They incentivise engagement and possibly the creation of training data that way. Recently, it has become a lot easier to get into the top 2000 again. They might have hit a plateau. Then again, they don't seem to see a need to offer trial access again, so maybe it is not that bad and people being too lazy for ranking sessions is not a concern.
Anybody else who has an opinion?
Midjourney might be in a critical phase right now.
They do focus on the product and listen to their users. They do micro-polls on their Discord server. The founder lurks in the Discord channels.
It seems like they believe in the concept of a productocracy.
For those who want to get some behind the scenes impressions:
The founder David Holz (who I think had 11+ years experience running Leap Motion before working on Midjourney) does a weekly 2.5 hour long and public "Office Hours" call on Discord each Wednesday, where people can ask questions "on stage" if they are lucky and in "lightning rounds" by typing in a chat room after they go though a short progress summary for the past week. Apparently, he does that after a 2.5 hour long company internal one in back-to-back sessions.
CENTS:
Control - might be a problem, quite literally in the image creation process. As the models get bigger and more sophisticated, the results get better. There are fewer messed up hands etc. But text prompts currently offer only so much control over the results. Soon there will be inpainting, where you will be able to tell Midjourney to repaint a part of an image that you don't like. They throw away many trained models, because the results don't meet their standards.
Stable Diffusion tries to give the creators more control, for example with ControlNet that takes scribbles or poses into account, but the results are not great when I try it. Others seem to be more lucky.
They went with a Discord based interface, because people liked it most in an initial split test. Now they run into painful limitations and are trying to gain more control again by developing a web frontend and a mobile one.
Laws and copyright concerns could become an issue. Activists could become an issue.
On the other hand, they are bootstrapping the business and have rejected money from investors. They were also successful in developing an AI Moderation tool that made sure that their showcase does not get flooded with problematic content.
Entry - is working in their favor. The diffusion model papers are pretty accessible and there are books that explain the basics and even show code examples. But they have built out a crazy hardware infrastructure. I think they said that each job that gets started is distributed to 10 GPUs. Training models takes a lot of time. They had to fix countless issues. There are many people that cross the 10k image generation mark. The competitors probably don't come close to that level of mileage and experience.
Need - Need might be a concern. What I would call a "bold move" is that they don't seem to go the OpenAI route by offering API access to create a 3rd party marketplace. The Holz vision seems to be becoming a community and a platform. He considers Instagram the competition, not OpenAI. They want to add story telling capabilities and they want their "Showcase" page to beat the "Best Of Instagram". That might become interesting. People want to take their images elsewhere and they can do that. I wonder if people actually "need" Midjourney as a community / platform that showcases their creations. If they force this artificially, this could become a huge backlash story.
To compete with Instagram, they need people who are not creators but join just to look at pretty pictures. Will that work? There seem to be enough people signing up for the $48 per month plan and the mini-polls showed that there is demand for $200+ dollar tiers, so creators see a need. But what features will they demand in the future?
The platform focus also means that they are moving in the direction of creating "opinionated" - flashy and attention grabbing - pictures by default. Not everybody appreciates that.
Time - Here they are probably golden. Things are automated. Quote: "Time honored business are usually involved in the act of creation.". And Midjourney does create.
Scale - They seem to be able to scale things up pretty well. The is a "status" channel on their Discord server and they fix the service and bot issues relatively fast. I am pretty happy with the uptime.
Additional thoughts:
From the "Office Hours", it is apparent that David Holz is working towards a sale of Midjourney. He talked about which Startups he admires. All were companies that produce physical things or companies that do weird things like trying to establish communication with Beluga whales. He also said explicitly that he wants to start a more crazy business, probably space related and that Midjourney is just something that happens to be doable right now.
Competitors and Open Source will catch up. Other companies will offer APIs. While they are looking into 3D and animation, these are in a very early stage of development. If his backlog is an indication, he seems to be mentally prepared for running the company for another 5+ years, though.
Are the numbers inflated? Are there growth concerns? A personal observation is that, for a while, it became more difficult to get into the top 2000 when ranking images in order to get an additional "fast hour" (I pay $8 per month for 3.5 fast hours and can get an additional one pretty much each day by spending 20-30 mins ranking images). They incentivise engagement and possibly the creation of training data that way. Recently, it has become a lot easier to get into the top 2000 again. They might have hit a plateau. Then again, they don't seem to see a need to offer trial access again, so maybe it is not that bad and people being too lazy for ranking sessions is not a concern.
Anybody else who has an opinion?
Dislike ads? Remove them and support the forum:
Subscribe to Fastlane Insiders.
