AI Redraws Boston

TheRatmeister said:
It's pretty good at giving Wikipedia-level summaries of things with Wikipedia levels of accuracy. (Which given its training data should not be surprising at all.) This is reasonably useful, and something LLMs are quite good at.

I mean I've not done a thorough investigation but it seems fairly okay at this, which again shouldn't be surprising given how many cooking websites it's no doubt scraped.

Personally I've had success with even pretty complex math and physics problems. It can do fairly complicated systems of equations and integrals for example. However I highly suspect it's actually just plugging into Wolfram-Alpha or similar here and using the text conversation more to convey the results.

This seems to me to be by far the biggest use. The more code you write the worse it gets, but for writing adequate code that can do a small task, which is then integrated into a larger program, it's quite good.

Of course we live in a Capitalist society though. How does ChatGPT convert these things into revenue, specifically $1-2tn of revenue?

It doesn't. Coding is the only place I really see it really making an impact, and that's also something that involves a lot of back+forth queries. At 1¢ per query, costs start to add up pretty quick. (That's 20-40x higher than a Google search, for reference.) Nothing about paid ChatGPT is especially compelling to an average joe. Ooh I can play around with higher resolution images, fancy. No that's not worth $10 a month unless you're rolling in it.

Unless they're somehow able to turn a text prediction model into an AGI within about 2 years or so the only place OpenAI & Co are going at this rate is bankrupt, that much seems pretty clear. The whole gravy-train is fueled by a mix of venture capital hype and a game of hot-bag-of-money between OpenAI, Google, Nvidia, Oracle, Microsoft, etc. If one of them can't pay, all of them can't pay and there goes ~25% of the Dow and your nan's pension. It's not like the telecom bubble either where sure, we spent all this money for not much returns and everything went to $#^@ but now we've got all these fiberoptic cables that are good for many many decades. Any GPUs bought from the current boom will be snails in 10 years time, if they're still working at all.
Click to expand...
^I think you're spot on. Cheerleaders and resistors are turning this in to a simple good/bad argument. The issue is instead one of value, taking into account the cost tradeoffs of the parallel harms/opportunity costs: is this a wise investment of $Trillions at the pace that the money is currently being poured into it, considering where else that money could be spent? Almost certainly not. This is thirsty speculation, just like how Boston built too way many labs before taking its foot off the gas.
 
Patrick Winn said:
more impactful
Click to expand...
I guess the Luddites should've stopped wearing clothes "for impact". But, we live in a society, bro.

They protested unfair labor practices including wage theft and improper compensation. Because of the history of the labor movement and those predecessors, like the Luddites, it was possible for SAG AFTRA to strike to protect the fruits of their labor from devalument by gen AI video models.
 
type001 said:
AI everywhere you look
Click to expand...
Much of it is not even "AI" in the LLM sense, but, "AI" in the marketing department sense. In my experience, generative AI is at best a microwaving of leftovers - can be done well if you know how to use a microwave, but, most folks won't know how to do better than an awesome plate of nachos.
 
Charlie_mta said:
That looks very real, like how Boston could have looked with some different twists and turns in its long history. No Charlestown Naval Shipyard, a smaller Boston Common, no Public Garden, no Fort Point Channel, the Charles River basin configured differently, etc.
Click to expand...
and no way to get in or out of downtown.
 
as02155 said:
Much of it is not even "AI" in the LLM sense, but, "AI" in the marketing department sense. In my experience, generative AI is at best a microwaving of leftovers - can be done well if you know how to use a microwave, but, most folks won't know how to do better than an awesome plate of nachos.
Click to expand...
So true! This has been annoying too. So many things are being cast as AI inappropriately.
 
type001 said:
So true! This has been annoying too. So many things are being cast as AI inappropriately.
Click to expand...
Obviously, the marketing departments see that billionaires are pumping billions into "AI" - so they lemming what their billionaire (and soon trillionaire) bosses are going for.
 
KCasiglio said:
This is just categorically false. Have you actually checked the veracity of its "summaries"? They're riddled with hallucinations to the point of it being faster to do it oneself. It's not even capable of giving main ideas a lot of the time. Usually it's just a guess of what the document is about based on the file name or title*

*Edit: Which is charmingly appropriate considering how deeply trained it is on reddit
Click to expand...
I use chatGPT and Claude all the time for litigation prep. I find them highly accurate and incredibly insightful at times. Just try loading a .doc or .pdf into it. I get the sense you're not using it right.
 
Patrick Winn said:
I use chatGPT and Claude all the time for litigation prep. I find them highly accurate and incredibly insightful at times. Just try loading a .doc or .pdf into it. I get the sense you're not using it right.
Click to expand...
You get the wrong sense. I will admit to having not tried Claude but ChatGPT cannot consistently read and summarize an uploaded pdf without hallucinating.
 
as02155 said:
I guess the Luddites should've stopped wearing clothes "for impact". But, we live in a society, bro.

They protested unfair labor practices including wage theft and improper compensation. Because of the history of the labor movement and those predecessors, like the Luddites, it was possible for SAG AFTRA to strike to protect the fruits of their labor from devalument by gen AI video models.
Click to expand...
Wow what a historical connection.
 
KCasiglio said:
You get the wrong sense. I will admit to having not tried Claude but ChatGPT cannot consistently read and summarize an uploaded pdf without hallucinating.
Click to expand...
You can't rely blindly on it. It's a tool. A tool that most smart and imaginative people adopted a long time ago. A lot of it is knowing the right questions to ask.
 
The first one looks awful except for the ocean and islands. The second one could probably fool people, "Pruenthal" and all.

1778211511634.png


1778211527830.png

 
The second one is interesting in that it still shows the limitations of AI image generation when it comes to things like this. It may "know" that certain buildings recur in Boston images, but placing them in a geographic sense is beyond it's scope. You can even see the original images pop out in it - a quick Google search of "boston aerial night" came up with this iStock photo of the Custom House that it looks like the AI pulled from to create it's image. You can also see it kind of, sort of trying to render the Garden right next to the Zakim, but failing.

That first one though...oof.
 
What model are you using and how much context are you giving it? Just "Make me a picture of Boston" isn't really enough and is open to a huge wide amount of artistic interpretation by the model.
 
393b40 said:
What model are you using and how much context are you giving it? Just "Make me a picture of Boston" isn't really enough and is open to a huge wide amount of artistic interpretation by the model.
Click to expand...

I'm not the one doing it. I found them at the links below the pictures. They came up as part of my broader weekly search for new skyline photos.
 
Not saying that AI will ever approach anywhere near perfect but y'all need to be patient.

Gen AI as a mass product is only in its infancy. The fact some of y'all think it should be coming out as near perfect need to really get a grip with reality. Every product that exists today comes with a ton of bugs and likely have been duct taped with a ton of hacky solutions. Some just do a better job hiding them.

The sales pitch and demos that tech companies usually give on their tools are entirely curated and cleaned so that it appears perfect, but only for the demonstration's specific use case (something akin to how we call "overfitting" in the stats world). The moment you try playing around it with your companies data, there will be a ton of bugs and edge cases that they missed or are unaware of.

Look at Tableau, a BI tool that has been widely used for over a decade but still doesn't have the capability to do conditional formatting like google sheet or excel without having to create a ton of dummy variables.

Honestly, it's people who think Gen AI should be near perfect that scares me because that tells me that they think this is all "magic" and they have little understanding of how it works and would either 1) blindly trust whatever it says or 2) fail to adopt it because "it's broken" (kind of like how people say a machine "is broken" because they don't know how it works or how to fix it when all you need to do is replace the batteries).
 
Last edited:
I don't think many of the people who are pointing out AI's imperfections are doing so because they expect it to be near perfect.
 
I thought the first one was supposed to be Southie until I saw the golden dome - the high rise by the water's edge and the church steeple made me think of 1410 Columbia and Gate of Heaven respectively. Again, wonder if it's pulling from an aerial of Southie as much as Beacon Hill. The wildly out-of-scale Mass Eye & Ear is hilarious.
 
You must log in or register to reply here.

Back
Top