I think there are few plausible reasons not mentioned.
We are still early in the tech. 2021 LLMs were not fit for any real purpose outside zany Madlibs. The tech has only really gotten fit for purpose since GPT4 and coherent models only got cheap and fast enough since llama 2 (July 2023). Even then smart models only got affordable around Deepseek v2 (May 2024) which was the first gpt4 level model to consistently serve subdollar per million tokens. You need at least a GPT4 level model to make a really interesting game. That's like a blink of time in game making time, indie games have 2-3 year cycles at best and AAA are like 5+. Even now fast, tool native interaction is only just coming online and there are no cheap models for that. Really fun game playing AI needs something like that to be magical.
There are AI games they just don't look like AI games neal.fun's Infinite Craft made a boat load of money and was very popular and it was powered by a Llama backend. Character ai and it's dozen copycats are like online storytelling/roleplay things. These are fairly popular. That could be a game or like a game platform or maybe it's more like fanfic. It has no fail state but people have made variants with fairly complex rules and states. Skryim and a few other games have pretty popular AI mods that let you add in NPC interactions that can talk, see, and even change their interactions with AI. The Skryim one Mantella has more than ~100k downloads.
We don't really have a northstar yet. Making a good game is really hard and usually takes someone doing something weird and clever. Minecraft is obvious in hindsight but many games on the way failed to make 'legos on the computer' fun. Incorporating AI breaks all the 'rules' of gaming. It operates slowly, it has high potential to break the rules and it has nearly unrecoverable failure states. Indie devs haven't figured out the best way to handle this and I suspect most big companies have just washed their hands of the whole thing until the AI gets faster, cheaper, and more stable.
The B2C case here seems off to me. The market of people who are going to pay the high price tag, have enough storage, and tolerate the extreme limitations and slowness of these machines seems to be: the rich, the elderly, and people with physical disabilities. The latter two categories come with a huge number of liability and regulatory costs that I think most of these companies are not willing to handle. That does not feel like a huge market.
I'll have a stronger belief in these things for consumers when we start seeing B2B adoption. Hotels have routine, highly structured cleaning tasks; hospitals have a need for extra strength, have highly structured cleaning tasks, and need to stock items; grocery stores have highly structured cleaning tasks and need to stock. Hospitals at least could tolerate the slowness of these things.
Without any B2B adoption it's hard to not see this as Roombas all over again. Cool for people who like it but low impact and still a toy 20 years later. I think generative AI makes these things better, though still perhaps struggles with long task adoption, but if you look at their movement they are still slow, weak, cognitively inflexible, and unstable. Maybe this tech is accelerating in some way I don't see and I'd love to be proven wrong here.
Visual anagrams popped up last year using similar, though simpler methods to those in the posted. Flips, internal rotations, rearrangements, color negatives etc.
[0]
Diffusion illusions did something similar at about the same time but with puzzles and multi-image color layer mixing. Some of the double puzzles they made are a lot of fun. They have great explainer videos [1] on that site including with Steve Mould.
Also there are diffusion double/hidden images using qrmonster, illusion diffusion, control net, or img2img that have been making the rounds. [2] for a random example. These work by using a fine-tuned diffusion model[3] to take an image and use it as a structuring element at various levels of following to a generated image. To see these illusions, squint or move the screen away. These are quite a bit more popular and easy to make than the other methods so many more examples show up around the internet.
You are right, current CRISPR systems tends to accumulate in the liver. Most CRISPR companies have shifted their focus to the liver over time because it's easiest to deliver there. Most viruses people use to target other organs are not large enough to carry CRISPR and lipid nanoparticles with CRISPR seem to like ending up in the liver and are hard to deliver at dose to hit other organ systems. It has been one of the big struggles of CRISPR companies. That being said, this is a huge deal and very encouraging.
As to the FDA stance, it tends to be more willing to go ahead with compassionate uses like this when it's clearly life or death.[1]
That's a good point, I assumed the farmers in the article were correct on getting squeezed, but individual producers rarely have a good view on the market. And McDonald's et al has been getting flack lately for the price of fries increasing. But looking at the potato price charts it's a pretty frothy market with a huge spike in 2023[0]. Maybe they've decided the suppliers are playing fair enough, at least with them. Or perhaps they are waiting to see if the frozen prices track the commodity price down before they decide to try to do something, I didn't realize the drop the article talks about was so recent.
While this seems true and I hope these companies get a large fine and some regulatory action takes place to wipe out these third party apps, I'm kind of surprised at the corporate learned helplessness here.
Back in the day McDonald's or one of their fast food competitors would have built their own frozen potato pipeline, made a massive marketing gimmick about cheapest fries and shattered this cartel quickly. It sounds like the companies are taking a high margin and the farmers would love to sell to anyone else. But it feels like the current managers at these fast food companies have gotten so used to outsourcing every part of production they lack the knowledge/remit to even try to set up a competing supply line.
It feels like only it's only efficient to focus core competencies if the people in charge of those stay smaller. Given how big companies can squeeze suppliers I see how they would end up consolidated. But if everyone is doing one thing and consolidating horizontally to negotiate better it becomes kind of red queen race.
Maybe when you have a multi-billion dollar supply chain and complex contract structure you can't just learn to do something new to solve a problem anymore.
The small players can't do much. As mentioned in the article, either potatoes and DIYing are cheaper or they aren't, but medium to large firms presumably could/should do something.
Absolutely, and in picking collabs with people who are willing to work with the weirdness and make it funny. Vedal is definitely a fantastic creator to make it work so well and the amount of fine-tuning and tweaking he must do must be unreal. But I think it still shows there is some hunger for this type of content, though you are probably correct that it still needs to be curated, gardened, worked with, and sometimes faked.
This was probably an okay idea terribly implemented. GenAI creators on social media kind of sense.
Neurosama, an AI streamer, is massively popular.
Silllytavern which lets people make and chat with characters or tell stories with LLMs feeds Openrouter 20 million messages a day, which is a fraction of it's totally usage. Anecdotally I've have non tech friends learn how to install Git and work an API to get this one working.
There are unfortunately tons of secretly AI made influencers on Instagram.
When Meta started these profiles in 2023 it was less clear how the technologies were going to be used and most were just celeb licensed.
I think a few things went wrong. The biggest is GenAI has the highest value in narrowcast and the lowest value in broadcast. GenAI can do very specific and creative things for an individual but when spread to everyone or used with generic prompts it start averaging and becomes boring. It's like Google showing its top searches: it's always going to just be for webpages. Making an GenAI profile isn't fun because these AIs don't really do interesting things on their own. I chatted with these they had very little memory and almost no willingness to do interesting things.
Second, mega corps are, for better or worse, too risk averse to make these any fun. GenAI is most wild and interesting when it can run on its own or do unhinged things. There are several people on Twitter who have ongoing LLM chat rooms that get extremely weird and fascinating but in a way a tech company would never allow. Silllytavern is most interesting/human when the LLM takes things off the rails and challenges or threatens the user. One of the biggest news stories of 2023 was an LLM telling a journalist it loved him. But Meta was never going to make a GenAI that would do self-looping art or have interesting conversations. These LLMs probably are guardrailed into the ground and probably also have watcher models on them. You can almost feel that safeness and lack of risk taking in the boringness of the profiles if you look up the ones they set up in 2023. Football person, comedy person, fashion person, all geared to advice and stuff safe and boring.
I suspect these things had almost zero engagement and they had shuttered most of them. I wonder what Meta was planning with the new ones they were going to roll out.
Valve actually backed off the ban and clarified their stance [0]. Now they allow AI art by the same standards as non-AI art if it is pre-generated and expect certain guardrails with regards to content and copyright if live generated. While their initial banning was probably over the top they did get better.
Linkin Park used it in 4 videos recently and seemed to use a combination of img2img and leaning into the glitchy style. [0] (These are 9-11 mo old so using much older models and techniques)
Peter Gabriel used it extensively in some recent music videos. The artists behind it also leaned into the glitchy style but are probably using some pre trained stuff to keep style.[1]
A writer/creative named Austin McConnell used AI art to make a 50 minute anime short to help market a book he wrote using AI [2]. Not sure how he kept consistency but this video got some flak and he has another video addressing his techniques.
Corridor crew did a second video which is a lot better but still a lot of work. [3]
I think a lot of projects using it are still kind of in the spec stage and are usually using a combination of loras, clip, generating multiple angles, and aggressive use of img2img and controlnets. And a few companies (Scenario [4] etc) are working on consistency. I think you won't see a ton of big projects using it yet because the tech is still early days and the early versions were really a bear to work with.
A lot of people on YouTube are using it like theyd use stock art (YouTube thumbnails, backgrounds, ads, or story boards for writing focused/story YouTubes).
There is still a stigma so a lot of people using it aren't announcing their use broadly, so I've found I usually have to stumble on their projects. Also, a lot of major companies aren't using it for that reason. I know Wizards of the Coast has had some arguments about that recently. Also consistency is still a problem as you mention which limits it's use in bigger projects.
Video games I think will be the first place we'll see it widely accepted. As people have been using AI generation techniques for background tools for like 20 years, see speedtree.
Too many users, I don't know Hugging Face's rules but they seem to limit how much each demo can use to a ceiling. When I ran it originally it was like 12 people using it, looks like the queue is now around 300 and Hugging Face doesn't spin up more instances. That being said the model is relatively small and can be run locally with at least 5 GB of VRAM according to the Stable Diffusion subreddit.
Seems super fast, some are saying 600x faster [0], than than the version made off of Google's paper. But it is a little less accurate. Point clouds are less useful but some on Reddit and the authors have tools to try to convert to meshes [1][2]. It does feel like stable diffusion level generation of good 3d assets is right around the corner. It will be interesting to see which tech wins out, whether it's some variant of depth estimation like sd2 and non ai tools can do, object spinning/multi angle view like Google's tool does, or whatever this tool does.
This is amazing but Google/OpenAI haven't released their models and don't seem to plan to. There is stable diffusion which was released and is probably slightly less good but still good if anyone wants to mess with it.
There is a huggingface instance, Collab notebooks, and local running notebooks here. [1] on the stable diffusion subreddit.
Also someone has packaged an exe that runs it with no fuss on computers with Nvidia GPUs that they posted on the media synthesis subreddit[0]
In my limited testing this compares ok to Dalle2. Style shifting works slightly less well and it's hard to force it away from normal images but with a little work it tends to be more accurate to your prompt.
Also seems to ignore that social pressure keeps people from account hoping.
Personally, I'd be much more likely to a la carte services on a month by month basis if I didn't have the subtle social pressure that 'well I'm not using this but my friends X, Y, and Z all are'. Even lacking the pressure element it seems like a small 'gift' to my family and friends who are on it and that it's not 'going to waste'. There is even the element of watching tings your friends are watching that keeps you watching. If they eliminated that I'd probably never keep more than one active at a time and likely not even that. Also once I've dropped a service for nonuse I'm much less likely to go back.
I think the biology revolution will look less like the computing (1940s-2010s) or transportation (1890s-1970s) revolutions and more like the chemistry revolution (1800s-1950s). It will be less iterating on one process and more years of seemingly stagnant progress while people tinker with the last big thing until the next big thing roars out of the gates. Iteration will be brief and seemingly unimpressive. The only real progress will be, that which was expensive and hard will become cheap, but unlike compute that won't necessarily be the driver of progress.
For instance early genetic testing revealed a huge number of disorders that could be linked to single genes that changed the game for how a certain sub-population has kids and treatment for a handful of disorders. Then it turned out a lot of stuff was multigenic or only very partially genetic and progress has stalled. As genetic testing has gotten cheaper we have learned more, but actual usable insights are rare.
Or take lab grown insulin from 1980s, it was a huge deal, everyone worked on similar technologies and knock ins and for a handful of disorders it was a complete game changer that petered out medically speaking. It saw a weird second life in farming and GMOs for a time that helped improve processes but didn't really move the needle in a big way. Then as it got cheaper, and new more efficient tools emergedm it reexploded as the field of biologicals, which have been great drugs since the mid-2000s. And as the technology gets really cheap we are seeing products like lab meats, Impossible leading the charge there, emerge.
Now we are on CRISPR, RNA-sequencing, CAR-T, and tissue engineering all of which, when they hit will have massive impact, but then slow down for a bit until the next big thing revs up.
This parallels chemistry where the theories of thermodynamics and gas laws emerged, sped everything forward, then petered out. Then organic chemistry emerged using a lot of the ideas developed by the previous field, sped everything forward for a while then slowed down. Then inorganic and solid state chemistry etc. But there was never a 18 month doubling it was more like a 10 year lull followed by 1000x in 5 years, if you were counting for example, the number of compounds developed. You could smooth the curve and pretend it was like compute but it's not. We'll just look back in 20ish years, or even look back today at the 1950s, and say 'wow there is a lot of stuff in our life that was based on biological research where did that come from'. Just like someone in the 1940s would look back and be like 'wow where did all this plastic/metal fab come from'.
I don't know the specifics for microscope lenses, I know a little about camera lenses, but I think it's largely high failure rate. It's a lot of mechanical grinding and heating and cooling with some final human checks. Each high quality lens has something like 10-15 lenses in it each which fails out pretty often. Also, shockingly, these lenses might still be hand assembled and checked? (I can't find a modern reference on this).
As to why it won't get cheaper anytime soon a couple reasons (at least what I'd guess):
1. Small batches and minimal incentives. It's feasible the high end microscope lens market just isn't worth disrupting. New processes could possibly make better lenses cheaper but just setting up the fairly massive tooling probably would eat up the margins. These lenses last forever so the yearly sale of these lenses is actually pretty small. (Funny enough there has been recent stories where competing manufacturers have started buying up some type of lens and actually completely drained ALL the available lenses of that type on the market, and that was a fairly cheap lens). Also people buying these lenses- labs and biotech tend to not care very much about price, so there is little incentive.
2. Replacement materials don't exist yet. The easiest way to beat the lens market would be to replace the glass lenses with plastic. This is what happened with phones/disposable cameras and is why camera prices plummeted. But at present this hasn't happened for higher end lenses.
3. There hasn't been cheap increased precision in machine movements really. So failure rates are unlikely to go down. If someone could adapt lithographic techniques to lens making it's possible it will get cheaper?
4. Kind of a dumb reason, but the biggest reason they aren't likely to get cheaper is that they haven't gotten cheaper. The process is about as automated as it can be and the microscope world is on the cutting edge of a lot of technologies (photolithogrpaphy, novel microscopy techniques, etc) if applying some technique like ML to some part of the process could make it cheaper, someone probably would have pushed it. There are also a lot of DIYers in this area. It's not a technological stagnant market. (I know this kind of competes with reason one, but what I'm saying is there is not 'easy win' to apply here.)
5. The big research in microscope/optics design isn't really focused on making high end lenses cheaper (As far as I've seen). It's more focused on stuff like making even higher end lenses that let you cheat the rules of optics, or like this paper, using low end lenses to do experiments that were impossible with more expensive lenses, making tiny/embedded systems, using beefy automated systems to image a ton of stuff simultaneously (can image a whole rat brain in like an hour now), and automating capture speed with the systems that exist (if your 100k system can image 10x faster and doesn't need a user it's like having a bunch of cheaper systems).
In summary, it's a well explored, small market, that no one is throwing money at and no obvious tech advances touch. Though this is largely speculation from talking to microscope reps and wandering around vendor fairs/automation conventions. Also I'm a biologist with an interest in optics not a microscope expert.
That is somewhat true but there are some fixed costs that cannot be easily brought down.
1. Optics- especially when you start talking 63x (looking at internal cell structures) color corrected objectives are generally going to run around 10k, there is really no way around that. A microscope like the one shown here (and the GRIN lens miniscopes used in brains recently and the ball lens microscopes) don't require as good optics as they aren't imaging at super high resolution and color misalignment won't hurt them.
2. The camera chip/housing/frame grabber- Generally the actual chip is pretty low cost. But the system to cool and grab a lot of frames really fast is pretty expensive. The rates of commercial cameras-120 FPS at their highest, is completely insufficient for a research machine. Also the sensitivity of an 'off the shelf' camera is not going to be appropriate for a research machine. These costs hold these cameras around 5-10k. This project doesn't require these systems because they are looking at whole cells so they can afford to 'lose' a lot of photons.
3. Lasers- if a microscope has a laser (generally only 2 photons or certain types of confocals), it's going to cost a lot. I don't know what costs go into making a laser but the technology is pretty well established and the power of laser required for these machines seems to consistently run into the 20-50k range.
Other parts that are can be made cheap, but cost more (1-2k) when they need to be long lasting and highly precise. Stages (xyz stepper motors in particular), dichroic mirrors/filters, housing body, LED/fluorescent lighting source. This project got around that by specializing these components and using off the shelf stuff. I think probably many of these components could have their prices dropped a lot.
I think there is a future in bringing some of the costs down, but until we get a cheaper way to grab a lot of frames quickly or make good optics on the cheap, I don't know if research scopes (except the ones currently under patent like STORM) will drop much. Though alternative scopes (like the miniscope- 1k) can sometimes surprise people and replace expensive microscopes (120k two-photon), but that's more about being able to build into the niche and cutting out the need for more expensive components. Though from what I've read/seen microscope companies have a ton of overhead.
I'm actually surprised this whole suit is being pursued by a professional society. Closed publications seem to be out of line with the interest of their members. Unless the society makes substantial money from their publications, which is unlikely[0], it's hard to see what the motivation of this organization would be and why the members of the society haven't loudly complained against it. I'd really expect a suit from this to come from Elseiver or NPG.
[0] From talking to a few representives of sceintific societies it sounds like they usually produce thier journals at or around cost.
"A study using 14C birthdating on sorted NeuN+ nuclei suggested that hundreds of new neurons are generated per day in the adult human hippocampus, with little decline with age. The results obtained from this method differ from the data presented here and other histology studies that show a sharp decline in markers of newly formed neurons during early postnatal development. Birthdating with 14C relies on the isolation of neuronal nuclei using NeuN antibodies, but subpopulations of oligodendrocytes and microglia can also express NeuN. 14C could also possibly become incorporated into DNA through methylation or DNA repair independent of cell division, processes that have been shown to occur at higher rates in the hippocampus. The proposed addition of new neurons to the adult caudate nucleus using this method is not supported by other work in the human or BrdU labeling in adult macaques. The 14C method is an innovative approach to perform birthdating in postmortem human samples, but it has not been validated in animal studies."
While this comes to the right conclusion I think it mistraces a lot of the issues.
For example: a fair amount of the depression and neurogenesis studies (and indeed neurogenesis manipulation studies) suffered from the same issue: it's really hard to make a population of cells die without generally poisioning the remaining cells. Most of the initial studies relied on things like radiation, ie radiate a rat and it starts acting depressed. There is now increasing evidence that radiation kills off a bunch of synapses and causes some wicked disregualtion across the system so that turned out to be wrong. Conversely BDNF enhances neurogenesis but it also enhances synaptic learning so even positive manipulations had this issue but in reverse. Newer timed genetic manipulation studies were more precise but the depression-neurogenesis link has been slowly losing steam in light of these studies.
But the BIG issue boils down to this: there was never strong evidence for the importance of neurogenesis. The dentate gyrus is a tiny region of the brain with neurons that act in unusual ways. In humans it's an even tinier region. While it is probably required, at least as a pass through, to learn new memories, it's surprising how much weight people were placing on neurogenesis in this tiny tiny structure. Given, networks can make small zones have big effects, but the weight of evidence should have always been on people pushing the area to prove the dentate was this massively important structure. This has not been conclusively done (studies have linked it to some specific subtypes of learning but those have limited it as much as they have found an important role for it). This is a structural issue in science broadly, there is an incentive to push forward with the next big finding but no incentive to go back and confirm the gulf of assumptions that a literature is resting on.
The good news is, human brains could always change, synaptic plasticity is present throughout life and has been shown by many good studies to occur throughout adulthood. We don't need new cells in a tiny part of the brain to learn new things, networks rewire and while that rewiring isn't something as dramatic as new cells, small effects can lead to big outcomes in dynamic systems. People obviously make new habits, learn new skills, and make new memories throughout life. The onus should have always been on science to show why it occurs, not to lend credibility that it occurs.
We are still early in the tech. 2021 LLMs were not fit for any real purpose outside zany Madlibs. The tech has only really gotten fit for purpose since GPT4 and coherent models only got cheap and fast enough since llama 2 (July 2023). Even then smart models only got affordable around Deepseek v2 (May 2024) which was the first gpt4 level model to consistently serve subdollar per million tokens. You need at least a GPT4 level model to make a really interesting game. That's like a blink of time in game making time, indie games have 2-3 year cycles at best and AAA are like 5+. Even now fast, tool native interaction is only just coming online and there are no cheap models for that. Really fun game playing AI needs something like that to be magical.
There are AI games they just don't look like AI games neal.fun's Infinite Craft made a boat load of money and was very popular and it was powered by a Llama backend. Character ai and it's dozen copycats are like online storytelling/roleplay things. These are fairly popular. That could be a game or like a game platform or maybe it's more like fanfic. It has no fail state but people have made variants with fairly complex rules and states. Skryim and a few other games have pretty popular AI mods that let you add in NPC interactions that can talk, see, and even change their interactions with AI. The Skryim one Mantella has more than ~100k downloads.
We don't really have a northstar yet. Making a good game is really hard and usually takes someone doing something weird and clever. Minecraft is obvious in hindsight but many games on the way failed to make 'legos on the computer' fun. Incorporating AI breaks all the 'rules' of gaming. It operates slowly, it has high potential to break the rules and it has nearly unrecoverable failure states. Indie devs haven't figured out the best way to handle this and I suspect most big companies have just washed their hands of the whole thing until the AI gets faster, cheaper, and more stable.