Has anyone tried to calculate how much LLMs would cost per prompt without the VC subsidy? Corporations have FOMO and are pursuing AI to avoid being left behind, while lacking many robust use cases. Actually paying for the stuff is the most likely thing to reset priorities.
Its pretty easy to calculate. You just need to know how much each part of the vertical is being subsidised.
Energy utilities are giving big Tech 25% off
Big Tech is giving places like OpenAI 30% off infra costs
OpenAI is still losing $10B+ a year
If we ignore training compute, aka lets say they have reached a point where they no longer wish to make new models, they only want to serve the current model, aka opex costs.
OpenAI would need to double its price just to get close to breaking even. Another price doubling if their energy and Big Tech subsidisies evaporate. AKA, a 4x price increase. $80 / month sub.....
And thats just text. Images are an order of magnitude more expensive, video and order more expensive than that....
The cost-per-prompt issue isn't the only one; as Utkash Karnwat explains, the use of LLMs as agents incurs geometric calls because of iteration while also introducing tolerance stacking. https://utkarshkanwat.com/writing/betting-against-agents/
What is one robust use case? At least where robust is direct sale in mission critical sectors federal sector where there are mil standards from an NIST Track. Semantic AI rules there.
TL;DR: The Semantic AI Model (SAM) plays a pivotal role in preserving human knowledge, augmenting AI capabilities, and ensuring equitable access to information. With its ability to capture, store, and retrieve data, SAM can act as a decentralized safety net against data loss and misinformation while enabling intelligent applications in robotics, language processing, and decision-making. By promoting collaboration and transparency in AI research, SAM contributes to a more resilient and innovative future for AI technologies.
Calculating the cost is one part, maybe the easy part.
How do you calculate the value you're getting from the output of the LLMs? This seems too subjective. Even for the easy case of programming, how do you calculate all the future debugging, bug fixes, recoding, etc to fix the AI slop code?
For more complex cases, how do you evaluate the productivity gains or financial gains from AI integration?
Thus far, AI fails to reliably solve even basic math problems when small changes are made to the problem's wording; therefore, how can we trust them to solve complex, multi-factorial problems that have serious real-world consequences?
The topic of AI is rife with either-or thinking. Either AI is going to usher in a UBI post-scarcity utopia or it will bring about human extinction. Either AI is on the verge of achieving human-level intelligence or is not even as smart as your cat or dog. Either the Singularity Is Near or "AI is bullshit".
Instead of the either-or trap, the reality is likely more nuanced. Anyone who denies the recent advances in AI is not looking at the facts, but anyone who thinks we are close to AGI using the current paradigm is equally misguided. The truth is somewhere in between-- but what that will look like in practice is not too hard to ascertain. In the next 5-10 years, I envision that the pronouncements of those like Ray Kurzweil and other techno-optimists will be deemed as laughably inaccurate. LLMs, the dominant paradigm, are already reaching the point of diminishing returns. The amount of energy, time, and expense to train the frontier models is already increasing faster than the models' capabilities. According to Market Watch on 10/26/2024: "The cost to society of AI chips, and the talent, electricity, water and more needed to manufacture them, currently dwarfs the payoff."
LLMs are predictive machines trained on vast amounts of data. They excel at "crystallized intelligence"; on the other hand, they still struggle greatly with what psychologists refer to as "fluid intelligence". Even small changes to the exams and problems used to test LLMs result in drastic performance reductions. A system with fluid intelligence would not be stymied by small changes to the problem set up that are not reflected in the training data. In combination with the problem of hallucinations, the inflexibility of these systems and their dearth of fluid intelligence will make their mass adoption by industry less likely. Until these models can reason fluidly with the kind of flexibility that humans can, then they will not be widely adopted in sectors like healthcare, law, energy, transportation, and so on.
The lack of widespread adoption by multiple industries is economically problematic for the AI enterprise and the flagship tech companies pouring billions of dollars into training these LLMs. Eventually, investors will want to see returns. But if AI is not being purchased and used throughout the economy in the way that people like Altman and Kurzweil think it will, then those flagship tech companies will not be able to recoup their upfront costs. They will not be able to pay back investors. Investors will in turn lose confidence in the AI industry. When that happens, a bursting bubble will become apparent. Capital will flee AI and research into new architectures beyond the LLM paradigm will suffer-- there won't be the money or investor confidence to keep things afloat let alone expand them. Those involved in the field will become more risk-averse. The end result will be a new "AI Winter".
Even though a new AI Winter is the most likely outcome in the next 5-10 years, without a doubt, AI will still have a significant societal presence. Most likely, we will all have the ability to talk to an AI on our phones that will sound just like a real human being (and will make Siri look like child's play). Image generation and LLMs will be barely noticeably better than they are now and there will be folks who spend more time interacting with “AI friends” than with other people, but as far as nanobots, "longevity escape velocity", space colonies, (functional, fault-tolerant) quantum computers, room temperature superconductors, climate change mitigation, and nuclear fusion, these *truly futuristic, materially relevant* technologies will still be "20 years in the future". A cyberpunk dystopia is the most likely outcome with mass surveillance and more cool, consumer electronics gadgets to distract people from the unfolding ecological collapse and mass extinction event. Hopefully a pivot toward more promising architectures will occur sooner rather than later and I will be proved wrong about all of this!
Isn't this a problem begging for a neuro-symbolic or some other hybrid models where we use encoding from LLMs and then build graphs that deterministically give the same answer everytime? The encoders are very good at building embeddings. How we put the embeddings to use is the reasoning. This can handle the brittleness problem (also known as learning by rote) that you mention.
Yes, I think this is a promising avenue. This is why I love the YT channel “Discover AI” because he avoids hype but acknowledges that there are promising ideas being developed by the AI community.
The fun thing is that you called the shots on GPT-5 being likely the tallest point of the hype three or four times in the last two years - and it may finally unravel to be the case. Can't wait to see how this goes...
Some say that GPT-5 is a cost reduction play, rather than an increased power or closer-to-AGI play. If so, it would make sense. Perhaps it signals OpenAI's acknowledgement of the diminishing returns of scaling and that they'd better figure out how to make money with the current level of LLM technology.
Sam is just using the proven Elon hype model. Get smart people to build a good beta product, take a pile of shares, hype it up, keep it private as long as possible so you don't have to actually disclose anything significant. 'Next year is AGI' is the new 'next year is full self driving'. At any sign of a problem pivot to something different and pretend that was the plan all along. Oh yeah, get something gold plated for Trump.
Plus: OpenAI employees are trying to sell a round of secondaries to SoftBank at $500B. That news dropped over the weekend. Too bad OpenAI is private and employees can’t dump their shares on retail.
Altmann certainly helped blow up the bubble, with help of course. Unlike the dot-com bust, this time we have names, and they will stick to this particular crash for a very long time.
Which makes me wonder just what is going on here. If ChatGPT5 was set up as the ultimate LLM product, with transcendent properties, or close enough, then why put out this version that has obviously fallen short? One does not promise the Holy Grail, then show up with a ceramic mug, however attractive. Why didn't they continue the promise that "the real deal is just around the corner"? That game has been around and working for a very long time. I'd like to know what caused Altman to light the fuse at this particular moment?
Megan McArdle seems to be the most prominent journalist taking up the "maybe it's a bubble, so what" call with her column on the 17th (https://www.washingtonpost.com/opinions/2025/08/17/ai-bubble-productivity-workers/). She includes Krugman's fax machine platitude and Solow's productivity statistics, as one does, because the dot-com bubble paid off for a few survivors. It then leans weakly into the anecdata-of-one approach ("Today, I use ChatGPT to kick-start much of the research I do for columns") followed by peer pressure ("when a technology is truly transformational, it’s probably safer to bet on premature optimism than to prematurely dismiss it out of hand").
What McArdle's column does NOT do is examine other bubble parallels like the Panic of 1873, in which overinvestment in underutilized railroads brought down the global economy for five years, or the telecom bubble of 2001, in which all the fiber laid by Worldcom and Global Crossing couldn't find any buyers to last-mile it into utility. Which is probably just as well for her sake as railroads that have lain fallow for decades can still be rolled on and dark fiber pulled in 1999 was finally carrying data by 2015 but the datacenters being built today will be obsolete by 2027, even presuming some "killer app" for AI starts making the rounds tomorrow.
I totally agree about the Paul Krugman meme. It’s become a lazy way of deflecting criticism of new technologies. I first encountered it when it was used to dismiss critics of Bitcoin, and now it’s fulfilling a similar role for ChatGPT fans.
The implied argument seems to be this:
The Internet is a new technology;
A famous expert (Paul Krugman) dismissed the Internet;
The internet succeeded.
.............
Large language models (LLMs) are a new technology;
A famous expert dismissed LLMs;
Therefore, LLMs will succeed! QED.
As a counterpoint, I could cite all the overly optimistic predictions from decades past about nanobots, cybernetics, genetic engineering, space colonization, or cold fusion that never materialized. E.g., I remember reading Isaac Asimov essays from the 70s predicting that ordinary people would be living in interplanetary colonies by the early 2000s. (Not to pick on Asimov, because he did write a lot of great popular science books.) I’m beginning to think that large language models are less like the Internet and more like space colonies or cold fusion—pipe dreams that never became reality.
AGI or AGi-like isn’t important I suspect. GenAI can be disruptive even if it is far from AGI. I would go as far as to say AGI is a red herring. GenAI has been *sold* with the AGI vibe, but don’t forget how the internet was initially sold: perfect information for everyone, democracy everywhere, and a new economy/long boom. All as nonsensical as AGI.
I suspect people like Sam may know this (at least they should by now) and they want to be the Google/Amazon/Meta that was still standing after the dotcom bust. Microsoft (business) and especially Apple (consumer) are in a different position.
If this is going to happen is uncertain that way depends on GenAI being cheap (uncertain) and good enough (also uncertain).
The dot-com comparison is weird because the Internet / WWW grew organically at first. I mean, Tim Berners-Lee wrote the first browser while working at CERN. Compare that to the extreme consolidation and concentration of expertise and funding when it comes to genAI.
The dot-com bubble only arose considerably later, after it became clear that the Internet + WWW would be the future worldwide network, relegating all alternatives to the sidelines. So it had already proven its worth, very dramatically so, and many use cases had already been accepted for a couple of years.
For genAI, this is not the case. It is still mostly a solution searching for a problem.
If one uses this technology for a while in a professional setting with unfamiliar, real-world tasks and doesn't just ask it for yet another toy project or yet another well-known exercise, its extreme restrictions become readily apparent.
Aside from the obvious hallucinations, there is also a subtle problem of just constantly missing the point. It's like genAI cannot crisply isolate concepts; concepts bleed into each other to an intellectual mush, and the more complex the task is, the worse it gets, rapidly so—to the point when you just get silly generalities. This isn't AGI nor AGI-like but just a waste of time and energy.
It's sobering how unintelligent and brittle it is, and not like a child, but in a fundamentally inhuman and bizarre way.
Aside from that, it cannot continually learn, but that's nearly a luxury concern as long as the other stuff isn't solved.
I guess, production of images is still the only clearly proven application. Because 1.) we can accept a higher rate of error here (people do not attentively look at low-end utility art for long), 2.) it is effortless for everyone to weed out bizarre and nonsensical images, and 3.) it makes sense to mass-produce variations of the same subject (if you, OTOH, have a program that works, it makes no sense to produce a variation just for the sake of it).
"the Internet + WWW would be the future worldwide network, relegating all alternatives to the sidelines."
The Internet was/is the interconnection of disparate networks. Everyone switched over to TCP/IP for numerous good reasons.
As for the rest ... yes, Gerben Wierda's comment is ignorant rubbish. Aside from the completely failed analogy to AGI:
"the Google/Amazon/Meta that was still standing after the dotcom bust."
Many companies are still standing after the dotcom bust ... e.g., before and during the bubble I worked for a Content Delivery Network company ... the technology we developed was eventually sold and is still in use. The bubble resulted from numerous startups that had no real product or business plan but went public with crazy stock prices ... the bubble burst and a lot of moms and pops who invested in these companies because they thought what goes up must keep going up lost their shirts or life savings. We might see similar things happen with AI but this is a market similarity, not a technical similarity.
"The Internet was/is the interconnection of disparate networks. Everyone switched over to TCP/IP for numerous good reasons."
Well, I was talking about the early online services such as CompuServe, The Source, and Prodigy. And those were primarily based on proprietary networking technology. There were also strange pre-Internet national networks in Europe, like Minitel in France.
Those providers later integrated TCP/IP (late 80s, early 90s) and then converted to true ISPs (mid 90s). And even later they switched off what had remained of their legacy networks (but that was very foreseeable development from the mid 90s).
IMHO, when certain technologies prevail over their competitors it's not fully because of technological superiority and everybody "seeing the light" but it also depends on many contingent sociological and political factors.
Let me clarify: the internet was indeed already there and it took off around 1994. Then came a well-documented period of big hype (idealistic in many ways). GenAI has been around since the 1990s (RNNs) and took a huge scaling step thanks to transformers in 2016-2017. GPT-3 is from 2019. It and the hype really took off with ChatGPT in late 2022.
My use of the word ‘initially’ for the 1994-2000 period was sloppy. It was the initial period of hype and selling, not the initial period of the internet.
Interesting behaviour by one of the commentators on my comment here. First post a reply, then immediately block me so I can neither see it (or any of this person’s comments), nor reply to it (not even saying “you’re right”).
Right, how much of the cost goes away if they stop trying to improve the base models with scaling? What are the ongoing costs to keep the LLMs working with new information, events, etc? Altman seems like a bad actor/excellent self promoter, which doesn’t discount the possibility that he was also fooling himself.
Has anyone tried to calculate how much LLMs would cost per prompt without the VC subsidy? Corporations have FOMO and are pursuing AI to avoid being left behind, while lacking many robust use cases. Actually paying for the stuff is the most likely thing to reset priorities.
Its pretty easy to calculate. You just need to know how much each part of the vertical is being subsidised.
Energy utilities are giving big Tech 25% off
Big Tech is giving places like OpenAI 30% off infra costs
OpenAI is still losing $10B+ a year
If we ignore training compute, aka lets say they have reached a point where they no longer wish to make new models, they only want to serve the current model, aka opex costs.
OpenAI would need to double its price just to get close to breaking even. Another price doubling if their energy and Big Tech subsidisies evaporate. AKA, a 4x price increase. $80 / month sub.....
And thats just text. Images are an order of magnitude more expensive, video and order more expensive than that....
The cost-per-prompt issue isn't the only one; as Utkash Karnwat explains, the use of LLMs as agents incurs geometric calls because of iteration while also introducing tolerance stacking. https://utkarshkanwat.com/writing/betting-against-agents/
I'd like to see that. I bet the economics would be bleak
What is one robust use case? At least where robust is direct sale in mission critical sectors federal sector where there are mil standards from an NIST Track. Semantic AI rules there.
Semantic AI is not genAI tho?
Right. This post explains.
http://aicyc.org/2025/06/10/the-crucial-role-of-sam-semantic-ai-model-in-preserving-knowledge-a-decentralized-safety-net-for-the-future/
TL;DR: The Semantic AI Model (SAM) plays a pivotal role in preserving human knowledge, augmenting AI capabilities, and ensuring equitable access to information. With its ability to capture, store, and retrieve data, SAM can act as a decentralized safety net against data loss and misinformation while enabling intelligent applications in robotics, language processing, and decision-making. By promoting collaboration and transparency in AI research, SAM contributes to a more resilient and innovative future for AI technologies.
Calculating the cost is one part, maybe the easy part.
How do you calculate the value you're getting from the output of the LLMs? This seems too subjective. Even for the easy case of programming, how do you calculate all the future debugging, bug fixes, recoding, etc to fix the AI slop code?
For more complex cases, how do you evaluate the productivity gains or financial gains from AI integration?
There is a bubble. The arithmetic makes that clear which is not to say there won’t be very significant developments after it’s burst, but much more slowly. There’s a nice analysis of the general form of bubbles on https://open.substack.com/pub/graceblakeley/p/the-ai-bubble-continues?r=b0jl4&utm_medium=ios.
It's often no fun seeing the future until proven correct, because only then do most others see.
Wait a second... are you saying that Sam just lied because he wants to make more money?
You are doing that, but at least you are doing it in a nicer way than I would
i never said any such thing, did I?
Well not in EXACTLY those words.
Exactly. You never said exactly that.
That is just the necessary consequence of what you said which also aligns perfectly with the relevant facts ;)
Many people have done exactly that, and Altman isn't exactly a paragon of virtue.
Thus far, AI fails to reliably solve even basic math problems when small changes are made to the problem's wording; therefore, how can we trust them to solve complex, multi-factorial problems that have serious real-world consequences?
The topic of AI is rife with either-or thinking. Either AI is going to usher in a UBI post-scarcity utopia or it will bring about human extinction. Either AI is on the verge of achieving human-level intelligence or is not even as smart as your cat or dog. Either the Singularity Is Near or "AI is bullshit".
Instead of the either-or trap, the reality is likely more nuanced. Anyone who denies the recent advances in AI is not looking at the facts, but anyone who thinks we are close to AGI using the current paradigm is equally misguided. The truth is somewhere in between-- but what that will look like in practice is not too hard to ascertain. In the next 5-10 years, I envision that the pronouncements of those like Ray Kurzweil and other techno-optimists will be deemed as laughably inaccurate. LLMs, the dominant paradigm, are already reaching the point of diminishing returns. The amount of energy, time, and expense to train the frontier models is already increasing faster than the models' capabilities. According to Market Watch on 10/26/2024: "The cost to society of AI chips, and the talent, electricity, water and more needed to manufacture them, currently dwarfs the payoff."
LLMs are predictive machines trained on vast amounts of data. They excel at "crystallized intelligence"; on the other hand, they still struggle greatly with what psychologists refer to as "fluid intelligence". Even small changes to the exams and problems used to test LLMs result in drastic performance reductions. A system with fluid intelligence would not be stymied by small changes to the problem set up that are not reflected in the training data. In combination with the problem of hallucinations, the inflexibility of these systems and their dearth of fluid intelligence will make their mass adoption by industry less likely. Until these models can reason fluidly with the kind of flexibility that humans can, then they will not be widely adopted in sectors like healthcare, law, energy, transportation, and so on.
The lack of widespread adoption by multiple industries is economically problematic for the AI enterprise and the flagship tech companies pouring billions of dollars into training these LLMs. Eventually, investors will want to see returns. But if AI is not being purchased and used throughout the economy in the way that people like Altman and Kurzweil think it will, then those flagship tech companies will not be able to recoup their upfront costs. They will not be able to pay back investors. Investors will in turn lose confidence in the AI industry. When that happens, a bursting bubble will become apparent. Capital will flee AI and research into new architectures beyond the LLM paradigm will suffer-- there won't be the money or investor confidence to keep things afloat let alone expand them. Those involved in the field will become more risk-averse. The end result will be a new "AI Winter".
Even though a new AI Winter is the most likely outcome in the next 5-10 years, without a doubt, AI will still have a significant societal presence. Most likely, we will all have the ability to talk to an AI on our phones that will sound just like a real human being (and will make Siri look like child's play). Image generation and LLMs will be barely noticeably better than they are now and there will be folks who spend more time interacting with “AI friends” than with other people, but as far as nanobots, "longevity escape velocity", space colonies, (functional, fault-tolerant) quantum computers, room temperature superconductors, climate change mitigation, and nuclear fusion, these *truly futuristic, materially relevant* technologies will still be "20 years in the future". A cyberpunk dystopia is the most likely outcome with mass surveillance and more cool, consumer electronics gadgets to distract people from the unfolding ecological collapse and mass extinction event. Hopefully a pivot toward more promising architectures will occur sooner rather than later and I will be proved wrong about all of this!
Isn't this a problem begging for a neuro-symbolic or some other hybrid models where we use encoding from LLMs and then build graphs that deterministically give the same answer everytime? The encoders are very good at building embeddings. How we put the embeddings to use is the reasoning. This can handle the brittleness problem (also known as learning by rote) that you mention.
Yes, I think this is a promising avenue. This is why I love the YT channel “Discover AI” because he avoids hype but acknowledges that there are promising ideas being developed by the AI community.
The fun thing is that you called the shots on GPT-5 being likely the tallest point of the hype three or four times in the last two years - and it may finally unravel to be the case. Can't wait to see how this goes...
I would love to See Altman brought up on charges of defrauding investors.
Fat chance. He'll buy some Trump crypto if he has to.
Some say that GPT-5 is a cost reduction play, rather than an increased power or closer-to-AGI play. If so, it would make sense. Perhaps it signals OpenAI's acknowledgement of the diminishing returns of scaling and that they'd better figure out how to make money with the current level of LLM technology.
yep
but they should have called it GPT-4-cheaper
GPT-Free-4-All
😂
Temu 4
Sam is just using the proven Elon hype model. Get smart people to build a good beta product, take a pile of shares, hype it up, keep it private as long as possible so you don't have to actually disclose anything significant. 'Next year is AGI' is the new 'next year is full self driving'. At any sign of a problem pivot to something different and pretend that was the plan all along. Oh yeah, get something gold plated for Trump.
Plus: OpenAI employees are trying to sell a round of secondaries to SoftBank at $500B. That news dropped over the weekend. Too bad OpenAI is private and employees can’t dump their shares on retail.
Currently, the real exponential growth in AI is in Google searches that include the term "AI bubble". Google Trends shows that the frequency of those searches has been blowing up since the start of August. https://trends.google.com/trends/explore?date=today%203-m&q=ai%20bubble
Altmann certainly helped blow up the bubble, with help of course. Unlike the dot-com bust, this time we have names, and they will stick to this particular crash for a very long time.
Which makes me wonder just what is going on here. If ChatGPT5 was set up as the ultimate LLM product, with transcendent properties, or close enough, then why put out this version that has obviously fallen short? One does not promise the Holy Grail, then show up with a ceramic mug, however attractive. Why didn't they continue the promise that "the real deal is just around the corner"? That game has been around and working for a very long time. I'd like to know what caused Altman to light the fuse at this particular moment?
Fabulous line: "One does not promise the Holy Grail, then show up with a ceramic mug, however attractive."
😂
Megan McArdle seems to be the most prominent journalist taking up the "maybe it's a bubble, so what" call with her column on the 17th (https://www.washingtonpost.com/opinions/2025/08/17/ai-bubble-productivity-workers/). She includes Krugman's fax machine platitude and Solow's productivity statistics, as one does, because the dot-com bubble paid off for a few survivors. It then leans weakly into the anecdata-of-one approach ("Today, I use ChatGPT to kick-start much of the research I do for columns") followed by peer pressure ("when a technology is truly transformational, it’s probably safer to bet on premature optimism than to prematurely dismiss it out of hand").
What McArdle's column does NOT do is examine other bubble parallels like the Panic of 1873, in which overinvestment in underutilized railroads brought down the global economy for five years, or the telecom bubble of 2001, in which all the fiber laid by Worldcom and Global Crossing couldn't find any buyers to last-mile it into utility. Which is probably just as well for her sake as railroads that have lain fallow for decades can still be rolled on and dark fiber pulled in 1999 was finally carrying data by 2015 but the datacenters being built today will be obsolete by 2027, even presuming some "killer app" for AI starts making the rounds tomorrow.
I totally agree about the Paul Krugman meme. It’s become a lazy way of deflecting criticism of new technologies. I first encountered it when it was used to dismiss critics of Bitcoin, and now it’s fulfilling a similar role for ChatGPT fans.
The implied argument seems to be this:
The Internet is a new technology;
A famous expert (Paul Krugman) dismissed the Internet;
The internet succeeded.
.............
Large language models (LLMs) are a new technology;
A famous expert dismissed LLMs;
Therefore, LLMs will succeed! QED.
As a counterpoint, I could cite all the overly optimistic predictions from decades past about nanobots, cybernetics, genetic engineering, space colonization, or cold fusion that never materialized. E.g., I remember reading Isaac Asimov essays from the 70s predicting that ordinary people would be living in interplanetary colonies by the early 2000s. (Not to pick on Asimov, because he did write a lot of great popular science books.) I’m beginning to think that large language models are less like the Internet and more like space colonies or cold fusion—pipe dreams that never became reality.
Whether AGI or an AG (attorney general) comes sooner remains to be seen.
Well formulated, succinct summary! Brilliantly consice.
This is what happens if the Woke EO law is being politically correct to access federal contracts not technically correct.
AGI or AGi-like isn’t important I suspect. GenAI can be disruptive even if it is far from AGI. I would go as far as to say AGI is a red herring. GenAI has been *sold* with the AGI vibe, but don’t forget how the internet was initially sold: perfect information for everyone, democracy everywhere, and a new economy/long boom. All as nonsensical as AGI.
I suspect people like Sam may know this (at least they should by now) and they want to be the Google/Amazon/Meta that was still standing after the dotcom bust. Microsoft (business) and especially Apple (consumer) are in a different position.
If this is going to happen is uncertain that way depends on GenAI being cheap (uncertain) and good enough (also uncertain).
The dot-com comparison is weird because the Internet / WWW grew organically at first. I mean, Tim Berners-Lee wrote the first browser while working at CERN. Compare that to the extreme consolidation and concentration of expertise and funding when it comes to genAI.
The dot-com bubble only arose considerably later, after it became clear that the Internet + WWW would be the future worldwide network, relegating all alternatives to the sidelines. So it had already proven its worth, very dramatically so, and many use cases had already been accepted for a couple of years.
For genAI, this is not the case. It is still mostly a solution searching for a problem.
If one uses this technology for a while in a professional setting with unfamiliar, real-world tasks and doesn't just ask it for yet another toy project or yet another well-known exercise, its extreme restrictions become readily apparent.
Aside from the obvious hallucinations, there is also a subtle problem of just constantly missing the point. It's like genAI cannot crisply isolate concepts; concepts bleed into each other to an intellectual mush, and the more complex the task is, the worse it gets, rapidly so—to the point when you just get silly generalities. This isn't AGI nor AGI-like but just a waste of time and energy.
It's sobering how unintelligent and brittle it is, and not like a child, but in a fundamentally inhuman and bizarre way.
Aside from that, it cannot continually learn, but that's nearly a luxury concern as long as the other stuff isn't solved.
I guess, production of images is still the only clearly proven application. Because 1.) we can accept a higher rate of error here (people do not attentively look at low-end utility art for long), 2.) it is effortless for everyone to weed out bizarre and nonsensical images, and 3.) it makes sense to mass-produce variations of the same subject (if you, OTOH, have a program that works, it makes no sense to produce a variation just for the sake of it).
"the Internet + WWW would be the future worldwide network, relegating all alternatives to the sidelines."
The Internet was/is the interconnection of disparate networks. Everyone switched over to TCP/IP for numerous good reasons.
As for the rest ... yes, Gerben Wierda's comment is ignorant rubbish. Aside from the completely failed analogy to AGI:
"the Google/Amazon/Meta that was still standing after the dotcom bust."
Many companies are still standing after the dotcom bust ... e.g., before and during the bubble I worked for a Content Delivery Network company ... the technology we developed was eventually sold and is still in use. The bubble resulted from numerous startups that had no real product or business plan but went public with crazy stock prices ... the bubble burst and a lot of moms and pops who invested in these companies because they thought what goes up must keep going up lost their shirts or life savings. We might see similar things happen with AI but this is a market similarity, not a technical similarity.
"The Internet was/is the interconnection of disparate networks. Everyone switched over to TCP/IP for numerous good reasons."
Well, I was talking about the early online services such as CompuServe, The Source, and Prodigy. And those were primarily based on proprietary networking technology. There were also strange pre-Internet national networks in Europe, like Minitel in France.
Those providers later integrated TCP/IP (late 80s, early 90s) and then converted to true ISPs (mid 90s). And even later they switched off what had remained of their legacy networks (but that was very foreseeable development from the mid 90s).
IMHO, when certain technologies prevail over their competitors it's not fully because of technological superiority and everybody "seeing the light" but it also depends on many contingent sociological and political factors.
"...how the internet was initially sold..."
The internet was not initially sold.
A useful idea that was not 'nonsensical.'
It was also not initially shilled far beyond any previous confidence scam.
It was devised, demonstrated, and sparked idealistic thinking.
Let me clarify: the internet was indeed already there and it took off around 1994. Then came a well-documented period of big hype (idealistic in many ways). GenAI has been around since the 1990s (RNNs) and took a huge scaling step thanks to transformers in 2016-2017. GPT-3 is from 2019. It and the hype really took off with ChatGPT in late 2022.
My use of the word ‘initially’ for the 1994-2000 period was sloppy. It was the initial period of hype and selling, not the initial period of the internet.
Interesting behaviour by one of the commentators on my comment here. First post a reply, then immediately block me so I can neither see it (or any of this person’s comments), nor reply to it (not even saying “you’re right”).
Among other errors, you confuse the internet with the web.
Sure, I was a bit sloppy. Doesn’t change the analogy between 1994-2000 and now regarding what was promised/expected and what happened.
The web is vastly bigger than what was "sold".
"All as nonsensical as AGI."
Ignorant rubbish.
I was an ARPANET developer ... we did good. Without us you would have no voice.
Right, how much of the cost goes away if they stop trying to improve the base models with scaling? What are the ongoing costs to keep the LLMs working with new information, events, etc? Altman seems like a bad actor/excellent self promoter, which doesn’t discount the possibility that he was also fooling himself.