Two years ago it would have been insane to say that you got help from ChatGPT to fix your outdoor plumbing, people here would have been frothing at the mouth for merely suggesting it. Four years ago it wasn't even on the radar of future possibilities.
The quality it is today, is the worst it's ever going to be.
Maybe? It's a plausible theory. But these operations are all wildly unsustainable financially at the moment, and it's not clear where they'll get future data from, having destroyed a lot of the incentives that generated their current source content.
Bubbles are not a great time to form intuitions. WebVan [1] and Kozmo [2] also seemed to herald a new age. Decades later, brick-and-mortar grocery stores and convenience stores are still doing fine.
> The quality it is today, is the worst it's ever going to be.
Someone in 2004:
A few years ago it would have been insane to say that you got help from Google’s I’m Feeling Lucky button to fix your outdoor plumbing.
The quality it is today, is the worst it's ever going to be.
It's still insane to do that today. Using any LLM to get info for things is a terrible idea, because they will slip inaccurate info into the mix and you have no way of knowing where it is.
I would still call you nuts for trusting ChatGPT with learning any construction or plumbing tasks. Atleast PVC is dead simple though since you basically just glue it together.
They already have all the data, all the money and all the chips. It’s actually the best it’ll ever be, as newer models will have to start paying off all that capex, newer models will be trained mostly on slop, and the SEO and influence op leeches will have begun their arms race to insert their products and values into the training data. We’ve seen this pattern before. AirBnB, Uber, social media, streaming video, et cetera didn’t get better once the VC money ran out and they needed to start turning a profit, they got much much worse.
For a slippery slope there has to be something about the slope that makes you start at the top. If you can go to any point on the slope, it isn't even a slope.
Anthropic can start censoring criticism of Dario completely independently from whether or not they censor Deepseek.
Maybe it’s like that trick where if a thousand people guess the amount of beans in a jar almost all of them will be wrong but their average will be very close to, if not, correct.
I doubt it. I'm stupid and I use LLMs a lot but I can still meditate for 30 minutes.
But apparently some of the smartest people in the world have lost the skill? But the commenter haven't, because why, they're 15 years older and thus immune to the same LLM-effects?
Plus, the issue with people having trouble sitting still for 30 minutes precede LLMs with decades.
Why is it so hard to believe? The young adults now have grown up with short form media and instant gratification / dopamine hits from apps. It's vastly different than people of the same age just a few years ago.
Not saying everyone else is immune, but those a few years older have also had a period without it.
> I'm stupid and I use LLMs a lot but I can still meditate for 30 minutes.
> apparently some of the smartest people in the world have lost the skill?
> But the commenter haven't
> why?
Perhaps because a correlation you assumed was there (more smartness = more ability to sit still alone with one's thoughts), is not actually as strong as you thought? If one does not start with that assumption, there is no inherent conflict in the 3 pieces of evidence you cited.
Or perhaps because you are smarter than you give yourself credit for :)
I didn't say I'm immune to those effects, I'm including myself in this as well. (also, I'm not older than my colleagues).
Most people definitely can't meditate for 30 minutes, so if you can do this, it's very impressive. Regardless, being able to think about poorly-defined problems and build completely new mental models from nothing is genuinely a really hard and uncomfortable task. If you don't use the skill you'll lose it.
> Most people definitely can't meditate for 30 minutes, so if you can do this, it's very impressive.
Maybe not traditional meditation, but I have no problem taking a 30 minute plus walk with nothing but my thoughts. It’s actually when I do most of my thinking. The other is in the shower/sauna where devices don’t work anyway.
I feel there's an interesting juxtaposition between these two quotes:
"The economy has absorbed automation before; agricultural employment collapsed from ninety percent of the American workforce to two percent and civilization continued. David Autor at MIT has shown that roughly sixty percent of today’s jobs didn’t exist in 1940. New technologies create new categories of work. True. But there’s a difference between an observation about the past and a law of nature, and the optimists consistently confuse the two."
and
"The assumption is that if you send people checks, they’ll find meaning in hobbies and community. They’ll paint. They’ll garden. They’ll finally write that novel.
This is ahistorical bullshit."
In the former the author establish that you can't make a observation about the past and take it as a law of nature. In the latter they refute arguments as "ahistorical bullshit" for not doing exactly that.
Especially since they then proceed with historical example, all of which had poverty as a strong contributing factor.
>Also, when did we stop liking to learn? Why is it a bad thing to know all the ins and outs of a programming language?
I do not know the inns and out of the assembly layer my high level code end up as. It's not because I don't like to learn, it's because I genuinely don't need to. At a certain level of AI performance, how will this be any different?
However, curious programmers who develop in high level languages will dabble with assembly maybe for fun, and will be much better off for it than those who treat parts of the stack like a black box never to be opened.
Because you may not know the specifics of the assembly being generated, but you’ve likely learned a language built on top of assembly. And the compilers do some great tricks behind the scenes to generate efficient assembly, but those tricks are specifically coupled to semantics of the source language.
An LLM is not coupled to anything and can generate output that simply does not relate to the input. This doesn’t happen with compilers, and if it does, then it’s a specific bug to be addressed. An LLM can never guarantee certain output based on the input.
If I write x < 100, I know exactly how the compiler will treat that code every single time, and I know what < means and how it differs from <=
If I tell an LLM that “I want numbers up to 100.” Will that give me < or <= and will it be consistent every single time, even the ten thousandth program that I write?
The language is ambiguous where the code is specific
To me this is semantics as far as it's related to "why don't you want to learn?"
I have a co-worker in another team that write java endpoins we consume. I can tell him what I need and I trust the output. I don't need to know java to trust him, it doesn't mean I don't want to learn.
There are thousand examples like this across every stack and abstraction level. From ssh-handshakes to gps.
Sure my co-worker is fundamentally different from a compiler which is fundamentally different from an LLM.
My argument is that the chain-of-trust where you offload knowledge to an external source is identical. We do it all the time but somehow doing it with an LLM means we no longer want to learn?
One difference is: to use a top notch compiler/assembler you don’t need to pay. They are open source and have a lot of support. To use the latest and greatest models (bc no one around likes to use non sota ones) you need to pay a premium price.
Multibillion dollars companies are now the gateway for every line of code you need to write. That’s dystopian. It sucks
Yes, but that's a completely different argument (that I agree with). Essentially, yes they are conceptually similar but one is bad because you have to pay rent to use it.
It's not novel in the sense that nobody knew about img2img. It's novel in the sense that nobody thought of using img2img to solve this problem in this way.
It's novel if you never played with img2img, including especially several forms of (text+img)2img. Or, if you never tried editing images by text prompt in recent multimodal LLMs.
That said, I spent plenty of time doing both, and yet it would probably take me a while to arrive at this approach. For some reason, the "draw a sketch, have a model flesh it out" approach got bucketed with Stable Diffusion in my mind, and multimodal LLMs with "take detailed content, make targeted edits to it". So I'm glad the OP posted it.
They’re actually quite good at it. I’ve had a number of situations where I’ve wanted to re-render some of my older comics. You can basically tell any SOTA multimodal model (NB, GPT-Image-X) to treat them as storyboards and prompt for a specific style: newprint, crosshatching, monochromatic ink sketch, etc.
Another thing I’ve gotten very used to doing is avoiding the “one-shot” approach. If I generate something and don’t like the results, I bring it into Krita, move things around, redraw some elements, and then send it back in with instructions to just clean it up (remove any smudges or imperfections). The state-of-the-art models can do an astonishing job with that workflow.
Ok it might just be me then. I view Nvidia‘s DLSS as a similar thing. There was even this meme that video games will in the future only output basic geometry and the AI layer transforms it into stunning graphics.
The quality it is today, is the worst it's ever going to be.