The tech has always been mature enough but simply too expensive. But since combustion engines are sold out for the next years, data centers are rushing to get anything they can. So the extra millions for power generation will be a rounding error
someone did, determined it would make them look bad, and pawned the assignment off to a junior executive to take the blame.
if it works then they delegated successfully and the project is winning and if not then they blame the neophyte VP and walk away with another quarterly bonus anyway.
The cost of electricity is (almost) negligible compared to the cost of having GPUs sitting idle. Electricity from fuel cells can be twice as expensive, but even then, it represents only about an extra ~15% of the amortization cost associated with keeping the GPU hardware idle.
Yes, but given the 3-5 year lifespan of the server infra capex, the separation becomes less meaningful. Better to count the entire TCO or discounted cashflow accounting for the rapid asset depreciation.
There is no way it's 3% if they're getting electricity from the grid, even if they get their cooling from somewhere else (eg lake/river water or outside air in winter). Median DCs are anywhere from 25-40% (depending on the jurisdiction's electricity costs and how much cooling they need). Hypercaler datacenters tend to be even more dense than the median traditional ones.
I saw a breakdown (don't have on hand sorry) where server depreciation was by far the majority. Power was the largest cash opex, probably matching your intuition.
The stuff inside the buildings (servers, switches, storage, cooling, etc) basically the stuff using the electricity and salaries/sub contractors (in addition to some techs handling the servers think security, cleaning, storage management, receiving shipments of new hardware, etc). Some taxes (depends a lot where)
This feels very useful for training agents but almost completely useless to humans? Do I really want to go back to every change I made? Whatuse would that be?
Jetbrains has had a "local history" feature for many years that had this, and I only use it occasionally but it feels pretty essential now. I once accidentally wiped uncomimted work in the terminal, and brought up Local History to restore it in a click. Or more casually, I'm struggling with a CSS design: what did it look like at 3pm? It's an effortless escape or experimentation valve between commits.
(It's also why you never need to manually save a file in jetbrains. It just assumes you want to save, because you can always safely go back if needed. Editors with manual saving feel cumbersome now)
To me it feels so weird that people are trying to push these model for online shopping like "here is how this dress/shirt/pants would look on you". But these models will always make the clothes fit your body and show you in flattering light and so on. How the actual garment fits is still as elusive as before these tools
The short-term goal of a tool like this is to sell products. The more ambitious long-term goal is to shift cultural norms, blurring the lines between advertising and reality until the question you're asking is no longer consciously asked. At least, not by average people, and not at the point of purchase.
I find it easy to envision a world, maybe 50 years from now, in which the very concept of "truth in advertising" is viewed as a lost, idyllic fantasy. Something people are nostalgic for, but feel powerless to regain.
> I find it easy to envision a world, maybe 50 years from now, in which the very concept of "truth in advertising" is viewed as a lost, idyllic fantasy.
From what I hear (fact-check, etc. needed) that's already the case right now with much more consequential transactions, like renting real estate in NYC. Square footages that are blatant lies, etc.
>I find it easy to envision a world, maybe 50 years from now, in which the very concept of "truth in advertising" is viewed as a lost, idyllic fantasy. Something people are nostalgic for, but feel powerless to regain.
Alternatively, perhaps we'll see models fine-tuned or steer-able towards accuracy that customers can themselves use to get a more honest view on what the product would look like in person.
The funny thing about these tools is that they can go either direction, but it sure seems like there's the potential for it empower individuals and shift the balance. Clothing sales shifting online has given sellers the advantage/ease to deceive without much customers can do other than hope the reviews aren't manipulated (they are) even before AI. Maybe this can turn things around as people start to shop with personal agents.
> Alternatively, perhaps we'll see models fine-tuned or steer-able towards accuracy that customers can themselves use to get a more honest view on what the product would look like in person.
we live in a world where the norm is to use filters on pictures of your own face to lie about how you look, imagine thinking than people actually care about truthness
>imagine thinking than people actually care about truthness
I think they actually care about "truthness" when it comes to how others will perceive them. You might filter your face for an online photo.
However, you're not going to want to look at yourself through your selfie camera with a filter that hides your imperfections when you're checking to see if something is stuck in your teeth on a date, don't you think?
There's nuance. People will fudge the truth to be perceived better, but don't want to be lied to when gathering data about how to be perceived better.
> However, you're not going to want to look at yourself through your selfie camera with a filter that hides your imperfections when you're checking to see if something is stuck in your teeth on a date, don't you think?
No, but you definitely use it in all the pictures you put on dating apps
Sure, but just because I may want to filter the photos I post, doesn't mean I want my mirror to show me a distorted view of reality. Isn't this use case more like the mirror, and less like the dating app?
That's a fun thought. A gradual division between "institutional models" trained to have capabilities that solve for the needs of corporations, vs. "personal models" trained to have capabilities that solve for the needs of individuals.
I wonder what kind of capabilities those might be?
> I find it easy to envision a world, maybe 50 years from now, in which the very concept of "truth in advertising" is viewed as a lost, idyllic fantasy.
There hasn't been "truth in advertising" for many, many years. The only thing that has changed recently is that you don't even have to hide most of the lies.
> I find it easy to envision a world, maybe 50 years from now, in which the very concept of "truth in advertising" is viewed as a lost, idyllic fantasy. Something people are nostalgic for, but feel powerless to regain.
It's infuriating the amount of effort people will expend to claim that what was achieved in the past is literally impossible to do now. It's pervasive, especially from allegedly-smart people like software engineers.
> I find it easy to envision a world, maybe 50 years from now, in which the very concept of "truth in advertising" is viewed as a lost, idyllic fantasy.
I've noticed something similar in Facebook marketplace ads for used furniture. Most of the images are AI generated to look like a Pottery Barn catalog, then the last image will be the actual item, full of scratches and other damage, sitting in a messy garage.
I've started to see this on Etsy and Wayfair too, where there will be a listing that is clearly just MDF flatpack being resold from China, but the AI-generated images wildly exaggerate the proportions of it.
Ironically, ChatGPT is decently good at ferreting these out. Like I sent it a screenshot of that listing and it not only helped me find where the original item was for sale, but also pointed out how the dimensioned diagram shows it as being just 49" tall, whereas the "in real life" image looks like it's at least six feet, based on it coming up over the top of the picture frame.
I ended up engaging a local woodworker to make me a piece like it instead. Obviously an order of magnitude difference in price, but it will actually be real solid walnut and finished to match my dining table.
Yes and even if all they do is ask it to stage an empty room picture with furniture and decorations it will make improvements like adding a doorway to a room that doesn’t exist.
The worst I've seen changed the view out a window from an alley to a private garden. Most also add natural light that isn't possible, and in some cases I'm convinced the generate image is depicting the space as larger than it actually is, with furniture and spaces that wouldn't fit in reality
I once heard about a company that specialized in making furniture about 10% smaller than normal for use in show rooms to make the spaces look larger. Not surprised they do it with AI if they can.
Wide angle lenses have been a thing in real estate photography since forever, but you would always have the reference of the furniture to ground your perception. Having furniture be slightly downsized is diabolical.
Many product/model shots actually use clips to make the clothing look like it's perfectly fitted to the body. [1] This image is actually over 15 years old at this point. I think there should be laws that prevent this because it veers into false advertising though others believe it's alright because you can theoretically tailor the clothes to fit like this.
Regardless, I'd rather see real clothing on a real person when it comes to my purchasing decisions. I buy a lot of vintage clothes online and I've noticed a dramatic uptick in AI images of models wearing the clothes. I've never once bought from those sellers because it feels disingenuous. Sometimes they have fake runways which is actual false advertising because it makes the item appear more expensive than it really is. I've also noticed that the AI models' body types are always thin even if the item is a L or XL. Needless to say, the AI isn't showing me what an XL looks like on a small model; it's showing what a small model would look like if the item fit perfectly.
Nope. I used to assist a photographer who did shoots for catalogs, and clips were used quite a bit. The same model has to wear dozens of pieces of clothing throughout the shoot, and not everything is going to fit their body well, but the clients obviously want everything to look well fitted.
I bet if you had a way to easily train a LORA on you trying on various clothing types the models would do pretty well but I agree without that I can't think of a way to get it to work. Reminds me of the flattering mirrors scam https://finance.yahoo.com/news/company-swears-controversial-...
One joy of online shopping, especially for people doing it in an impulsive and/or addictive way (it's not really rare) is the satisfaction they get from the imagination of having it.
An acquaintance of mine was buying many and not wearing most, as she did not attend that many social occasions. Still, she kept buying.
Eventually, she had to face the actual problem in her life that bothered her. She ended up dealing with it, terribly.
There's definitely a predatory "everything will look good", but you can also leverage exactly that part for your own purposes - I've done pre-shopping a couple of times by asking for something like "a grid of 9 versions of my photo, wearing different types of X that look good". Definitely helped with choosing a good style.
Definitely. I think the appeal to some isn't GenAI per se but image-to-image generation like "make this item look like a talented photographer shot it".
But I can't imagine that your conversion wouldn't benefit by showing actual imperfect photos of your products, especially when you run a platform with 1k new listings per month. I'm personally turned off by something that looks like 3D, especially when all the different colour variations look the exact same.
The cynic in me says this is just an advancement and logical continuation from the known problem of sycophantic behavior in text-to-text chatbot format LLM, to image generation models.
I use nanobanana 2 to test changes in paint and flooring/tiling to great success. And the real pro move is taking those images to a designer to tweak the remaining 20% or so.
You just described how people are using AI in general. And I think that last step is actually a huge gap. People will brainstorm or improve their understanding of the task with AI and create something. Ideally they would hand it off to a professional for finalization like you did.
I wish a service existed where I could make something in AI (text, images, whatever), then pass that AI output to an actual human who would use it as a guide to produce an actual product.
I used it recently to explore a few placements of a pool and glass building on my property. It's not authoritative, by any means, but it did answer some questions
And at this level it's barely any different than an architectural render, more for "how could this look" rather than "how will this look"
Good or bad, I’m pretty sure showing things in the best light has always been the point of “marketing”. There is a reason ads aren’t filled with ugly people with misshaped bodies, and it isn’t because the intention is to reflect reality, so what you allude to as problematic is what a lot of businesses call a feature.
I'm pretty sure I saw somewhere people can sell re-sell clothes. They can include a picture of themselves wearing the item and it replaces literally everything but the clothing with someone/someplace "prettier".
In the end the one thing that is completely honest is the portion of the picture that is the item you are re-selling. But somehow to me the entire thing feels disingenuous.
Marketing a product is often disingenuous. You didn’t actually need that shirt, or that bag of chips. That’s the definition of successful marketing in their mind.
If marketing had to be honest, an entirely different set of products would likely be the most popular. I don’t think that’s a good thing by any means, just that the problem hasn’t really changed much beyond more focused targeting, which has always been a goal of marketing anyway.
It’s hard to convince someone who isn’t interested to buy something which is why tech ads sell on tech channels and car ads sell on car channels. Now you get to star in your very own clothes ads…for yourself…targeting you…
I suspect companies drool for that prospect because you are selling it to yourself with only some nice gentle nudging. That is the grail of marketing; that you don’t even realize you are in the midst of being marketed to.
Well from the images I've seen they seem to be pushing text-heavy "infograph" capabilities pretty hard. Outside of that, I have serious doubts that it'll be better than other proprietary models like NB Pro and gpt-image-2.
User-targeted fashion advice is the polite goal. AI designed to generate images of people is actually racing to capture the entertainment markets. They want to be ready to replace models/actors in everything from fashion mags to porn studios. That is where the money is.
I've been struggling with this question myself. But isn't this a model training/use problem (a.k.a skill issue )? Isn't there a way to make these models be faithful to how people will actually look?
i'd like it to notice things i would miss, like "this is ring-spun shirt, so it will sit like this on your torso" or "these pleats will require you to iron them" etc
> But these models will always make the clothes fit your body and show you in flattering light and so on
This is something that can be fixed over time. And if this forces clothing manufacturers to stick more to their advertised "specs" (width/length), then it's a win for us.
The "climate" cares little about carbon intensity of labor or dollars. It cares about absolute tons of green house gases in the atmostphere and if you use more to produce more you still use more.
You should always assume these that benefit most from the technological forefront think they can outrun the output of climate change or any of the outputs.
It generally is called "effective altruism", eg, techno jesus will solve whatever problems creating techno jesus creates.
What prevents a bad actor from posting "easy" problems to the board, solving them and getting quick reputation. Then as a validator they can easily validate their own malicious changes to someones software?
I always wondered how these models would reason correctly. I suppose they are diffusing fixed blocks of text for every step and after the first block comes the next and so on (that is how it looks in the chat interface anyways). But what happens if at the end of the first block it would need information about reasoning at the beginning of the first block? Autoregressive Models can use these tokens to refine the reasoning but I guess that Diffusion Models can only adjust their path after every block? Is there a way maybe to have dynamic block length?
Does anyone know what kind of RL environments they are talking about? They mention they used 15k environments. I can think of a couple hundred maybe that make sense to me, but what is filling that large number?
Download every github repo
-> Classify if it could be used as an env, and what types
-> Issues and PRs are great for coding rl envs
-> If the software has a UI, awesome, UI env
-> If the software is a game, awesome, game env
-> If the software has xyz, awesome, ...
-> Do more detailed run checks,
-> Can it build
-> Is it complex and/or distinct enough
-> Can you verify if it reached some generated goal
-> Can generated goals even be achieved
-> Maybe some human review - maybe not
-> Generate goals
-> For a coding env you can imagine you may have a LLM introduce a new bug and can see that test cases now fail. Goal for model is now to fix it
... Do the rest of the normal RL env stuff
The real real fun begins when you consider that with every new generation of models + harnesses they become better at this. Where better can mean better at sorting good / bad repos, better at coming up with good scenarios, better at following instructions, better at navigating the repos, better at solving the actual bugs, better at proposing bugs, etc.
So then the next next version is even better, because it got more data / better data. And it becomes better...
This is mainly why we're seeing so many improvements, so fast (month to month, from every 3 months ~6 monts ago, from every 6 months ~1 year ago). It becomes a literal "throw money at the problem" type of improvement.
For anything that's "verifiable" this is going to continue. For anything that is not, things can also improve with concepts like "llm as a judge" and "council of llms". Slower, but it can still improve.
Judgement-based problems are still tough - LLM as a judge might just bake those earlier model’s biases even deeper. Imagine if ChatGPT judged photos: anything yellow would win.
Agreed. Still tough, but my point was that we're starting to see that combining methods works. The models are now good enough to create rubrics for judgement stuff. Once you have rubrics you have better judgements. The models are also better at taking pages / chapters from books and "judging" based on those (think logic books, etc). The key is that capabilities become additive, and once you unlock something, you can chain that with other stuff that was tried before. That's why test time + longer context -> IMO improvements on stuff like theorem proving. You get to explore more, combine ideas and verify at the end. Something that was very hard before (i.e. very sparse rewards) becomes tractable.
Every interactive system is a potential RL environment. Every CLI, every TUI, every GUI, every API. If you can programmatically take actions to get a result, and the actions are cheap, and the quality of the result can be measured automatically, you can set up an RL training loop and see whether the results get better over time.
With the rough numbers from the blog post at ~1k tokens a second in Cerebras it should put it right at the same size as GLM 4.7, which also is available at 1k tokens a second. And they say that it is a smaller model than the normal Codex model
reply