Baldur Bjarnason"LLMs are here to stay," is like saying that the modem-based internet of the 90s was forever. It was a largely unusable piece-of-shit tech that was manifestly incapable of delivering beyond a fraction of what the dot-com bubble promised, but it prompted investment in fixing the one thing that prevented that promise: bandwidth.
I can't emphasise strongly enough how this ISN'T happening with LLMs. They are throwing all their money into building ON the broken tech and next to nothing in developing a replacement that has an actual chance at fulfilling the promise LLMs hint at, because there genuinely aren't any plausible pathways to get there from here
The modern AI Bubble is 100% a "modem speeds were good enough and the dot-com companies only failed because they didn't have enough funding to get them over the hump created by the government's regulatory fight with Microsoft" delusion.
Modem speeds are very much not enough and if broadband had proven to be infeasible for some unexpected technical reason, then the web would definitely have faded away into a hobbyist-oriented subculture in the style of ham radio.
And if you think LLMs are "good enough" as they are, then the only thing I can say is I hope you're right because at the moment it looks like the tech industry is trying to bake a broken toy into every aspect of our society in the belief that the only thing that actually "breaks" tech is regulatory interference
Addendum: even if we were on the path towards developing an “actually works” replacement to current era LLMs—which all suffer from the same problems—that hypothetical future tech is not what’s being baked into everything. Even if they did manage to magically come up with an LLM that “works” nobody’d be using it because broken shit will be what’s built into everything.
Responding to “modem speeds are unworkable and if we don’t fix that the internet is a dead end fad” with “the internet is still extremely useful” isn’t a counterargument, it’s self-identifying as a member of the subculture that will remain after the fad peters out
The difference between thinking about how any given tech impacts you individually—harm versus utility—and thinking about how it impacts society if adopted at scale is effectively the difference between concluding that:
“Cars are fine, actually, the bigger the better. So helpful.”
Versus…
“We’re destroying ourselves and our planet because public transport is less profitable than the car industry.”
jakob.pxi@baldur I don't know if I phrased it succinctly enough, but I think there are different and more incentives at play in business adoption than the oft discussed "it need not be able to replace you, just convincing enough for your boss to fire you over the promise of capabilities" and overleveraged data center investments. There is a landgrabbing rush for interface control going on that very much aligns with business tenets of controlling distribution channels.
@jakob Ah, that’s a good point
Cassandrich
DianeI suspect there is an unspoken motive for sticking LLM into everything.
I think they're trying to get us to pay for another tool of oppression. LLMs are effective at manipulating people and automating surveillance so they need fewer secret police to prevent us from overthrowing them as the climate crisis get to be unlivable
Angela Scholder@baldur Well, at least Packet Radio was a lot of fun to play with.
My first web page was on Packet Radio.
My first web page was on Packet Radio.
fedops 💙💛@baldur modem speeds very much were enough until the enshittification of the web.
Parade du Grotesque 💀@ParadeGrotesque ahem. My first modem was also 2400 bps, and while it was way before the web there very much was an Internet. I remember configuring uucp for mail and inn for netnews in 1991, having blagged access to the modem bank at my student helper job at the time.
@baldur
@baldur
Jeff Grigg@fedops @ParadeGrotesque @baldur
I was reading ASCII text as it came across at 110 baud and was printed on a teletype. 300 baud is faster than you can read text.
I did a fair amount of programming work at 1200 baud. Like all of my college work. Mostly remote.
9600 baud is annoyingly slow for images. But I eventually paid for higher bandwidth in order to download large software packages.
@fedops @ParadeGrotesque @baldur
I curse at our local water company for subjecting me (and everyone else) to full screen full motion video when I just want to login and pay my water bill. Even their own public relations people criticize it for not being very accessible, and for being irrelevant and misleading.
There's a lot of waste — bandwidth, memory, CPU time, disk space.
…
But high resolution streaming video is really nice to have!
Falcon Darkstar@fedops @baldur they were emphatically not. I remember trying to use the web in those days and the experience was very much one of clicking something and then coming back in a few minutes to read it.
That doesn't mean what is happening now is ok though, but in most cases it's not transfer time but rather the time parsing and executing wild amounts of JavaScript that makes it slow.
Guy on the runsubtopic
Did you just return home or something? ;)
Masahiro Miura (@[email protected])
#Dayton #Hamvention 2025 #ハムベンション #アマチュア無線 https://youtube.com/watch?v=0KVvuK6lqvc&si=_gM1GjV5huVz52NA
tamas@baldur I was with you up until this point, but I disagree here. We have largely used up the bandwidth increase with much larger assets on the web unnecessarily. Sure, maybe webapps will render a spinner and progressbar instead of looking at the browser load everything, but it's still not loading that fast. Everything did not get as fast as the difference between the 56 kbps connections of modems, compared to 1-2 GBit fiber connections today (and their improvements in latency). The user expectation is still that you have to wait a bit for websites to load, even though we could do much better. Similarly, if connections remained as slow as they were then, we could still have achieved many of the advancements that make the web what it is today, except we would still have a lot of pressure to deliver small file sizes. Many of the advancements in that area are flat out ignored because we have a lot of bandwidth.
@tamas I think you've forgotten just how incredibly slow the modem internet was. It barely did images, period. Let alone streaming video or audio. Also screen resolutions needed to improve match the experience of print to remain competitive and that meant that images had to follow. Images sizes are not the biggest issue with the web today. The modem experience was much slower than you remember.
@baldur My point is that we have ignored many software advancements that could make better use of the low bandwidth. For example, real-time voice chat was not possible back then, and yet, speex was developed in the early 2000s that made this possible in as narrow as 8kbps bandwidth. Similarly, we were limited in our use of graphics, but image compression codecs also vastly improved since then. All I'm saying is, I don't think the web would have faded away to a hobbyist subculture even without the bandwidth developments.
And it's not as if high speed internet did not exist at the time: The backbone and well connected sites had high bandwidth. It's just that most homes did not.
And it's not as if no homes had high bandwidth feeds: Many people did have "cable TV," which is high speed distribution to homes. And some had satellite (with high latency issues).
It was practically always an issue of cost and building out infrastructure.
LisPi@baldur @becomethewaifu While something of a detail, I'd be relatively satisfied with the speeds DOCSIS 3.1 is supposed to be capable of giving me symmetrically or even VDSL2+ asymmetrically. Of course literally no one offers either of those properly, the only (barely) adequately maintained & provisioned network infra is fiber.
90s-tier modems sucked though.
90s-tier modems sucked though.
Sara Joy 
@baldur or "hey copper wires are fine let's just keep throwing smaller and smaller square waves down it with more and more pre-distortion to overcome the lossyness and it'll be grand"
But in the end you still end up with a snarled up result at the other end where the 0s and 1s are blurred into each other and the noise floor, with ever more broken packets and errors.
D Ingram@sarajw @baldur Copper wires? Luxury! When I started on the internet we had a 128kb/s satellite connection. For the entire country of New Zealand. When there was severe sunspot activity we'd lose our connection to the world. https://en.wikipedia.org/wiki/Internet_in_New_Zealand#cite_ref-64
@sarajw @baldur It was pre web! I had to combine my own browser (Lynx) for Vax in 1994. NCSA Mosaic was around 1994 I think. There's a good write up of NZ internet history here. https://www.nethistory.co.nz/Internet_in_New_Zealand_Timeline/
Satellite went from 14.4k to 64k in 1992, and then to 128k in 1993. The improvement when the link went to fibre was amazing. The lag on telnet was horrible on satellite.
Satellite went from 14.4k to 64k in 1992, and then to 128k in 1993. The improvement when the link went to fibre was amazing. The lag on telnet was horrible on satellite.
Internet in New Zealand Timeline
'Connecting the Clouds - The Internet in New Zealand' is a history of the people, activities and events that contributed to the creation, then growth, of the Internet in New Zealand. Written by author Keith Newman, the book was commissioned by InternetNZ (the Internet Society of New Zealand Inc) and published by Activity Press.
@ingram ha oh gosh I bet!
@sarajw It still is, if you're using a satellite in geo synchronous orbit. Had a flash back about 9y ago when using an Inmarsat modem for remote equipment monitoring. SSH was horrible to use, but streaming speed was "good" at 400k.
It still amazes me that my home internet connection is now 5x faster than the LAN when I was at uni.
@ingram lucky you!
Parts of Germany are in the dark ages re consumer bandwidth. We can't get more than 100Mbps at home.
@sarajw I'm getting 50Mb/s at home. LANs were very slow -- we'd only just started the move from coaxial ethernet to twisted pair. I helped install the first 100Mb/s switch (FDDI) in the country when I had a part time job with the computer centre. It was only for the University backbone but it made a big difference.
@ingram amazing.
Yeah I remember going to Uni and being like wowwwww broadband is amazing!
Then home in the holidays to 56K 💀
I don't really think that home (and work) internet speeds were the main cause of "the dot com bubble bursting."
But it does look to me that many of the *other* widely accepted causes of the ".com bust" really do apply to (Generative) AI companies and products. And that that is almost certainly accurately predictive.
Joe@baldur so much of what LLMs do currently could have been done 5, 10 years ago without LLMs if all the big tech companies hadn't tried to suck the advertising and constant growth wells dry.
Dave Neary@baldur What is the LLM equivalent of bandwidth for the 1990s Internet?
@dneary If we look past the fact that prompts and chatbots are a fundamentally unsuitable paradigm for serious productivity work (much like touch screens are an unsuitable paradigm for touch typing), you'd still need to figure out parameterisation (structured input that's actually parsed by the model) for security. It's the only way of truly preventing prompt injection.
They also need to be made more deterministic and reliably factual
That's just for starters. All effectively impossible AFAICT
Chris Pitts@baldur Thinking of LLMs as the modern version of the HTML BLINK tag gives me hope 😁
su_liam@thirstybear I miss the blink tag. It was an obvious and reliable indicator that what you were about to read was bat shit insane. By gating the weird typography behind JavaScript or CSS skills, they effectively sanewashed a lot of the rightwing web that way.
Willow "Wolveric" Catkin@baldur "LLMs are here to stay" is like saying "Gamers will love NFTs!" when in reality, they only ever featured in two proper games, that both completely flopped... 
Ben Aveling@baldur disagree. Modems were useful. Sure, better was to follow, but modems helped build that.
LLMs are not useful, except for edge cases like academic misconduct, and LLMs are not a pathway to anything better.
LLMs are not useful, except for edge cases like academic misconduct, and LLMs are not a pathway to anything better.
Barry Schwartz 🫖@baldur Yesterday I proved once again that "AI" isn't even an attempt at actual intelligence, by asking one in Swedish whether it spoke Swedish. It didn't even recognize I was using a different language.
But the real proof is that human intelligence is obviously not language-based. Otherwise other highly intelligent mammals—especially great apes—would all have recursively structured language!
LLMs will NEVER be able to do actual mathematics. They can pass high school "math" tests. Big deal.
@baldur Oh, as to what LLMs ACTUALLY are, that's EASY. They are pattern recognition-generation software.
Not artificial intelligence.
They took grammar research, which led to the ability to process language well, and which was indeed an important contribution to the study of artificial intelligence.
They also took neural networks, which are based on the fallacy that mimicking neurons would produce intelligence.
They put these things together with fast database searches.
Joe Alexander@chemoelectric Apes have human intelligence?
@chemoelectric Great apes don't have language?
@joe I was very careful to state RECURSIVELY STRUCTURED language. In other words, language easily capable of infinite complexity.
@chemoelectric Your assertion is just that, only an assertion.
@joe The intelligence of a chimpanzee and the intelligence of a human must be almost the same thing, IN TERMS OF MECHANISM. And the brains are structured almost the same, as well.
If the requirement of similarity is not obvious to you then you might be either a Creationist or an LLM advocate. :)
Here is a problem that cannot be solved by juggling language like an LLM, and which even most mathematically educated humans solve incorrectly: With what probability is six the last digit of pi?
@joe What a chimpanzee and a human have in common is they can reason in pictures, by directions in relation to their bodies, etc. These are the types of reasoning needed to solve that problem.
@chemoelectric Please preach more about chimp problem solving and exactly how it works.
Angua 
@baldur
Oh but it was fun, the connection dropping mid download of a few hundred ks.
That inimitable sound of the handshake, especially when it didn't work.
Phone bills in the thousands...
No you're right, it was crap..
Oh but it was fun, the connection dropping mid download of a few hundred ks.
That inimitable sound of the handshake, especially when it didn't work.
Phone bills in the thousands...
No you're right, it was crap..