I'm not on Bluesky or anything so I don't have perspective on some of the blanket AI dismissals, but I am a tech worker and am very much on the Luddite side of things with respect to LLMs. The Luddite view is frequently mischaracterized as "AI is fake and it sucks" and that's not right at all. But this characterization serves a purpose insofar as it allows the Casey Newtons of the world to sidestep the actual objections entirely. These objections range from the obvious, that the technology only works reliably in very limited domains or where accuracy & reliability aren't that important, or the more subtle, that whatever productivity gains from AI might occur will be be fully captured by the capitalist class.
Klein's interview with Buchanan is telling. Klein pushed him pretty hard about why no one on Buchanan's watch really seemed to plan for the social disruption that would occur should some sort of AGI emerge soon. I think the answer to this is obvious: neither Buchanan nor the people he worked with really believe that a world-changing AI will actually happen. But it never hurts to make a bold prediction: fully self-driving cars in 2016. No, wait, 2019.... Now....who knows.
Instead of fully self-driving cars, we got Teslas covered in cameras constantly running in "sentry mode", surveilling everyone and everything. *This* is the Luddite objection. Hype, promises, massive over-investment, extra surveillance, cars unsafe for pedestrians, and massive asset price inflation for the benefit of the few.
There is also some conceptual slippage going on: when I, a Luddite, say AI is fake, I refer to the AI industry, whose upside-down economics are premised the magical thinking that LLMs will lead to AGI, loosely defined. Or according to another view, the AI industry is essentially a fossil fuels & data center political economy play where the actual models are sort of secondary.
It's self-evident that there is real technology there and it can be passing useful for people in some ways. I used it the other day to help me figure out what to wear for an event and to brainstorm about a project I am working on. Handy! I couldn't imagine paying actual money for these conveniences, and less still for the ability to "vibecode" a recipe generator based on an image of my fridge. The lack of a killer consumer app, the lack of corporate uptake of OpenAI integration apart from some big consulting shops, and the complete commodification of LLMs in general make the industry pretty fake even if people can get a bit of mileage out of LLMs in their everyday lives.
I'm not even overly frugal by nature; I pay for your content!
I also work in the tech industry and this largely matches my thoughts. The strongest evangelists of LLM coding I've encountered have virtually all been over generalizing from using AI for simple greenfield or prototype apps. They really do help a novice coder create a simple web app, and that's impressive, but it also has very little to do with professional software engineering work. I have consistently found that chatbots often fail with even moderately complicated coding tasks, sometimes in ways that can waste quite a bit of time.
That's not to say that capabilities won't dramatically improve. I honestly have no idea if we're heading into a long plateau or more exponential gains. If these *do* become world changing I am pretty confident that average workers will end up far worse off, which worries me quite a bit.
Exactly. LLM coding is also great when you are creating what you think is a novel app that already sits in a public repo somewhere. The lunch recipe from fridge image is a great example of a problem humans already solved. I guess the LLM saves you the trouble of finding the code yourself.
If you keep an occasional bottle of Elmer's glue in your fridge, then an LLM app is good enough if you have the judgment to ignore its output and the brainpower to know how to make a tuna sandwich without instructions instead. But for any kind of enterprise application where you do get the occasional glue upstream, dang, what a nightmare living with the consequences of an LLM implementation could be.
Do you not get any use out of the more popular code completion tools like Cursor?
The other day I used Claude to write me a script that would perform a somewhat tricky codemod in my company's huge repo, and it felt like I did save real time on something that wasn't just a play app.
It's also nice to stand up unit test files really quickly, with cases and all.
I've sort of adopted an AI mantra of sorts to smooth out the high and lows of, as a technically-minded person, being interested in and every once in a great while aided by LLM type things, and horrified/frustrated at the centrality of the discussion, the costs, the claims - 'it is interesting that computers can do that now!'.
Because, well, it is! There are things like making new text that is like old text in interesting ways but different in interesting ways, is interesting! It was also interesting when computers were bad at chess and then were good at chess, and when pictures made on computers in movies went from looking bad and niche to looking pretty good (and then often bad again). It was interesting the first time a computer accepted a voice command (in the '60s) and a robot car drove across the country (in the '80s) and the first time someone had weird itchy feelings about a chatbot (in the '70s). It was interesting when you could push buttons and math answers came out! Computers are interesting! Sometimes they are useful!
But also, I dunno, *they're just computers*. The world filling up with computers has not radically bent its economic trajectory from the Before Times because they mostly do dumb things, and like anything they can mostly do things that don't take a lot of new figuring and waiting for happy accidents, and maybe they do them a little better (or actually do them a little worse because they're being driven to market by enormous piles of capital that can Do Things), and there are profound incentives for the bored pile of money, with its chip foundries so expensive they need to be kept in operation like Cold War shipyards, and the piles of data the wise were saying they probably shouldn't be collecting even if the reasons seemed thin, and the simple fact of being excited and blinkered by your work, to say that it's All Over- whatever that means. The fact that 'it is interesting that computers can do that now' places them firmly in a pantheon of things that have often been not super useful, or bounded, or premature, or misunderstood, or in the end, not that interesting.
Like, the chatbots are making fewer mistakes and seeming more like a search *because they're doing search.* Summoning up a surface summary on Madison et al is interesting, but also, it can do that *because Wikipedia is sitting right there* with ten page surface summaries.
The killer app for LLMs is text transformation, full stop. That's really neat! It's neat that my mess of poorly formatted notes cut and pasted from ten places are now formatted. It's neat that something can pull the topic sentences of ten papers and put them in one paper instead of looking at them all in different places on the front page of a search. But also these things in some ways don't seem that surprising if you said 'I took all the public facing documents in the world (and a few we stole because those are so bad) and had a computer look at them a billion times.'
Your point about web-searching and Wikipedia as the sort of giants on whose shoulders lots of this stuff is built is great, and I'd add that it's a good reminder that Wikipedia and Google are themselves astonishing technologies that have surely changed the world in myriad ways—but not ones that we typically think of in terms of some threshold moment or a "before or after." So much software and hardware of the last 50 years has both changed everything!! and also not really changed anything; it's hard to imagine LLMs will truly be different.
The mystery cult around 'AI' has basically nothing to do with the actual technology behind it, which is variations or expansions of what we've already been doing with computers - compression, interpolation, means, vectors - for decades, and everything to do with the conviction that since such basics are now capable of producing sufficiently flashy results, those who were secretly convinced that they were Actual Magic all along can now exit the closet. I wrote a Bluesky thread a month ago that gets into it a little https://bsky.app/profile/electriceden92.bsky.social/post/3lhmi4tfd3s2x
I wholeheartedly agree with this assessment. I work in the educational and medical field and the best use case is it's abulity to listen to my jargon and content heavy observation notes and assessments and turn it into a good, bland, professional report to put in my documentation. Really neat! Saves a few hours a week! Will pay a few bucks for that. Don't get how that makes an all encompassing job killer.
I’m probably more in the AGI skeptic group - though I use LLMs heavily for coding work as a data scientist / analyst - and I think their value there is incredible and if anything, still understated.
However - the source of my negativity to AI and the promise of AGIs boils down to the destructive energy costs, and more so how untrusting I am of the billionaires and billion/trillion corps that have everything to gain from feeding the hype cycle. They are unreliable and corrupt narrators.
After listening to the Ezra episode and then Casey Newton on Panic World this week, I detected this AI mood swing but couldn’t put my finger on what was going on until I read this post. Newton is a pure tech boosterite and Ezra is credulous, it makes sense the whispers from the AI insiders would sway them.
Also, Ezra’s Biden-admin guest spent half the episode touting the merits of AI for espionage in and statecraft. Why would the general listener root for that?
Proof that elites will use AI to serve whatever agenda they wanted to push anyways. “We must develop domestic AI so we can spy on China better than they can spy on us!”
Yes I didn’t even get a chance to get into the other group of people who are profiting or benefiting from “AGI” nonsense talk—China hawks! In some ways I find the “will China achieve AGI before us?” stuff worse and less grounded than the paperclip-maximizer apocalypse stuff.
Anyone who recalls the original hype is going to be more skeptical going forward—and I believe we should recall it. It does tell us something—that we are not in a reliable information space when it comes to AI or AGI. It’s never going to tell you what IS true to know that, it’s just an outlook for the assessment of information.
This is a well-measured take, thank you. I also get as annoyed with the knee-jerk reaction as I do with the breathless hype. From my perspective on the ground, the thing that is changing the narrative is actually that the models are getting better.
The scaling laws are not slowing down. The people who made that claim all last year have been proven wrong again and again. scaling model parameter still works, as evidenced by gpt 4.5. scaling inference time compute works, as evidenced by reasoning models, and even scaling RL training on these reasoning models also works.
The reason people are "feeling the AGI" is that The people closest to the models are using them to do real work and seeing the effects. sure, they're not perfectly reliable. you can't just set them loose on a task indefinitely and count on it being done. But you can't do that with humans either. The agentic capabilities of reasoning models are good enough to accelerate your day-to-day work, like having an a smart and infinitely patient coworker who always responds to your messages within seconds. This now accelerates the production of more & better AI systems.
It's smart to be skeptical, but it is also smart to update on new facts. I think it is good that kind of normie journalist like Kevin and Ezra are breaking the ground for this conversation. The alternative is people claiming it's snake oil right up until it takes their jobs.(Yes, even content creators like you and me 😉 )
This is a great post but as some who is (probably inadvisably) on bsky there really are plenty of people saying precisely "AI is fake and it sucks" by which they mean it has no practical uses and anyone who says it does is an NFT-level grifter. I find myself negatively polarized against both those people and the AGI believes simultaneously hah.
Those on BlueSky are more right than wrong. And I would go evenr further to say those prototypical posters will be the ones leading the public's apathy and eventual disdain for AI.
Slop was the word of 2024 is no coincidence.
Big Tech and silicon valley technocrats are deathly scared at the idea of the public turning against AI. AI has not even started to make much money, and is already facing application cancelling backlash.
I think technology is good for solving specific problems. I distrust technology that claims to solve every problem (“general”).
The idea that people can write ourselves out of the picture is a deluded fantasy. We’re stuck with ourselves and each other.
The people who are most obsessed with AI tend to think work can be separated from people. But people actually like and need to work. And jobs that are low-paying and mundane can actually be very fulfilling when there is a sense of community involved.
I listened to Newton’s interview on Panic World recently, in which he dunks on AI skeptics and freaks out about the potential consequences of a supposedly imminent AGI. It left a bad taste in my mouth. Part of that is because he argued that AI-skepticism is somehow rooted in a millennial inferiority complex over not being the most tech savvy person at the office any longer, which is totally frivolous, but the other reason I couldn’t put words on until I read your last paragraph.
I'm glad I'm not the only one who was really put off by that episode. I didn't have high expectations of Newton, but it was disappointing coming from Ryan Broderick.
"it’s no longer necessarily the case that, say, asking an LLM a question of fact is strictly worse than Googling it"
Isn't the difference that people don't assume the top Google result is a definitive answer? I mean, I hope most don't. A Google or other web search is the *beginning* of an investigation.
I've certainly had specific questions for which an LLM query yields worse (objectively false) results than a DuckDuckGo search plus a human sensibility reading.
I'd analogize it as follows: typing a question into a chatbot and trusting the answer : clicking the first result in Google :: using a RAG like Perplexity and checking the links : scrolling through the Google results and clicking on them.
Yeah, this is about right. And the other reason I say that LLMs are not strictly worse than Google is that ... Google is a lot worse than it used to be!
This is one of the saddest parts of the whole saga: that Google actively sabotages one of the greatest technological achievements of the last 50 years---their crown jewel---to be a middling provider in the overcrowded GenAI field.
The self-sabotage is not to be a middling provider of AI, but to make truly ungodly profits from advertising. One of the reasons I’m more bearish on AI is that it’s only a matter of time before ads pollute that knowledge source (such as it is) as well.
Great article Max. LLMs are deception machines. Because they generate fluent and grammatically perfect text on any subject it’s easy to miss the errors/hallucinations. It makes them dangerous to use for most practical purposes, particularly as autonomous agents.
It’s an inherent problem with neural net/transformer architecture. The technology is fundamentally unreliable.
Massively underrated take. I would’ve thought doing massively complex calculations is a surer sign of intelligence than the ability to compose a limerick. But we are narcissists. We like being serenaded.
Really solid piece, and agree on AGI increasingly becoming a useless term.
An insight on the split between skeptics and believers: it’s less about your assessment of the current usefulness and power of the tech than about the future projections you buy into. The skeptics think that next year the tech won’t be much more useful than it is now. The believers foresee a decade (at least) of exponential improvement—there’s a chart in their heads with a line going up, Moore’s Law style.
And part of the projected improvement would necessarily means the tech evolves in unexpected ways way beyond whatever ChatGPT 5 brings. Recent developments in the overbroad “AI” tech tree include things like materials science discovery algorithms and better weather prediction tech—lending some credence to the believer view. So the believers also have a greater willingness to handwave away issues like LLM hallucinations or extra fingers in generative images. If these models are just stepping stones, these problems don’t seem as salient.
The Benedict Evans article is underrated for how brutal it is. The fact that the example he criticizes is from OpenAI's marketing is just perfect. "OpenAI is promising that this product can do something that it cannot do, at least, not quite, as shown by its own marketing." :chefskiss:
I'm not on Bluesky or anything so I don't have perspective on some of the blanket AI dismissals, but I am a tech worker and am very much on the Luddite side of things with respect to LLMs. The Luddite view is frequently mischaracterized as "AI is fake and it sucks" and that's not right at all. But this characterization serves a purpose insofar as it allows the Casey Newtons of the world to sidestep the actual objections entirely. These objections range from the obvious, that the technology only works reliably in very limited domains or where accuracy & reliability aren't that important, or the more subtle, that whatever productivity gains from AI might occur will be be fully captured by the capitalist class.
Klein's interview with Buchanan is telling. Klein pushed him pretty hard about why no one on Buchanan's watch really seemed to plan for the social disruption that would occur should some sort of AGI emerge soon. I think the answer to this is obvious: neither Buchanan nor the people he worked with really believe that a world-changing AI will actually happen. But it never hurts to make a bold prediction: fully self-driving cars in 2016. No, wait, 2019.... Now....who knows.
Instead of fully self-driving cars, we got Teslas covered in cameras constantly running in "sentry mode", surveilling everyone and everything. *This* is the Luddite objection. Hype, promises, massive over-investment, extra surveillance, cars unsafe for pedestrians, and massive asset price inflation for the benefit of the few.
There is also some conceptual slippage going on: when I, a Luddite, say AI is fake, I refer to the AI industry, whose upside-down economics are premised the magical thinking that LLMs will lead to AGI, loosely defined. Or according to another view, the AI industry is essentially a fossil fuels & data center political economy play where the actual models are sort of secondary.
It's self-evident that there is real technology there and it can be passing useful for people in some ways. I used it the other day to help me figure out what to wear for an event and to brainstorm about a project I am working on. Handy! I couldn't imagine paying actual money for these conveniences, and less still for the ability to "vibecode" a recipe generator based on an image of my fridge. The lack of a killer consumer app, the lack of corporate uptake of OpenAI integration apart from some big consulting shops, and the complete commodification of LLMs in general make the industry pretty fake even if people can get a bit of mileage out of LLMs in their everyday lives.
I'm not even overly frugal by nature; I pay for your content!
Very well-said, and I think you're absolutely right.
Aw shucks, you're not just saying that because I'm a subscriber? :)
I also work in the tech industry and this largely matches my thoughts. The strongest evangelists of LLM coding I've encountered have virtually all been over generalizing from using AI for simple greenfield or prototype apps. They really do help a novice coder create a simple web app, and that's impressive, but it also has very little to do with professional software engineering work. I have consistently found that chatbots often fail with even moderately complicated coding tasks, sometimes in ways that can waste quite a bit of time.
That's not to say that capabilities won't dramatically improve. I honestly have no idea if we're heading into a long plateau or more exponential gains. If these *do* become world changing I am pretty confident that average workers will end up far worse off, which worries me quite a bit.
Exactly. LLM coding is also great when you are creating what you think is a novel app that already sits in a public repo somewhere. The lunch recipe from fridge image is a great example of a problem humans already solved. I guess the LLM saves you the trouble of finding the code yourself.
If you keep an occasional bottle of Elmer's glue in your fridge, then an LLM app is good enough if you have the judgment to ignore its output and the brainpower to know how to make a tuna sandwich without instructions instead. But for any kind of enterprise application where you do get the occasional glue upstream, dang, what a nightmare living with the consequences of an LLM implementation could be.
Do you not get any use out of the more popular code completion tools like Cursor?
The other day I used Claude to write me a script that would perform a somewhat tricky codemod in my company's huge repo, and it felt like I did save real time on something that wasn't just a play app.
It's also nice to stand up unit test files really quickly, with cases and all.
I think there are some great, totally usable features. Hard to see a great product in there, though.
True. I would sooner subscribe to a couple good substacks than pay for AI.
I've sort of adopted an AI mantra of sorts to smooth out the high and lows of, as a technically-minded person, being interested in and every once in a great while aided by LLM type things, and horrified/frustrated at the centrality of the discussion, the costs, the claims - 'it is interesting that computers can do that now!'.
Because, well, it is! There are things like making new text that is like old text in interesting ways but different in interesting ways, is interesting! It was also interesting when computers were bad at chess and then were good at chess, and when pictures made on computers in movies went from looking bad and niche to looking pretty good (and then often bad again). It was interesting the first time a computer accepted a voice command (in the '60s) and a robot car drove across the country (in the '80s) and the first time someone had weird itchy feelings about a chatbot (in the '70s). It was interesting when you could push buttons and math answers came out! Computers are interesting! Sometimes they are useful!
But also, I dunno, *they're just computers*. The world filling up with computers has not radically bent its economic trajectory from the Before Times because they mostly do dumb things, and like anything they can mostly do things that don't take a lot of new figuring and waiting for happy accidents, and maybe they do them a little better (or actually do them a little worse because they're being driven to market by enormous piles of capital that can Do Things), and there are profound incentives for the bored pile of money, with its chip foundries so expensive they need to be kept in operation like Cold War shipyards, and the piles of data the wise were saying they probably shouldn't be collecting even if the reasons seemed thin, and the simple fact of being excited and blinkered by your work, to say that it's All Over- whatever that means. The fact that 'it is interesting that computers can do that now' places them firmly in a pantheon of things that have often been not super useful, or bounded, or premature, or misunderstood, or in the end, not that interesting.
Like, the chatbots are making fewer mistakes and seeming more like a search *because they're doing search.* Summoning up a surface summary on Madison et al is interesting, but also, it can do that *because Wikipedia is sitting right there* with ten page surface summaries.
The killer app for LLMs is text transformation, full stop. That's really neat! It's neat that my mess of poorly formatted notes cut and pasted from ten places are now formatted. It's neat that something can pull the topic sentences of ten papers and put them in one paper instead of looking at them all in different places on the front page of a search. But also these things in some ways don't seem that surprising if you said 'I took all the public facing documents in the world (and a few we stole because those are so bad) and had a computer look at them a billion times.'
Fantastic mantra!!! I'm going to steal it.
Your point about web-searching and Wikipedia as the sort of giants on whose shoulders lots of this stuff is built is great, and I'd add that it's a good reminder that Wikipedia and Google are themselves astonishing technologies that have surely changed the world in myriad ways—but not ones that we typically think of in terms of some threshold moment or a "before or after." So much software and hardware of the last 50 years has both changed everything!! and also not really changed anything; it's hard to imagine LLMs will truly be different.
The mystery cult around 'AI' has basically nothing to do with the actual technology behind it, which is variations or expansions of what we've already been doing with computers - compression, interpolation, means, vectors - for decades, and everything to do with the conviction that since such basics are now capable of producing sufficiently flashy results, those who were secretly convinced that they were Actual Magic all along can now exit the closet. I wrote a Bluesky thread a month ago that gets into it a little https://bsky.app/profile/electriceden92.bsky.social/post/3lhmi4tfd3s2x
I wholeheartedly agree with this assessment. I work in the educational and medical field and the best use case is it's abulity to listen to my jargon and content heavy observation notes and assessments and turn it into a good, bland, professional report to put in my documentation. Really neat! Saves a few hours a week! Will pay a few bucks for that. Don't get how that makes an all encompassing job killer.
I’m probably more in the AGI skeptic group - though I use LLMs heavily for coding work as a data scientist / analyst - and I think their value there is incredible and if anything, still understated.
However - the source of my negativity to AI and the promise of AGIs boils down to the destructive energy costs, and more so how untrusting I am of the billionaires and billion/trillion corps that have everything to gain from feeding the hype cycle. They are unreliable and corrupt narrators.
After listening to the Ezra episode and then Casey Newton on Panic World this week, I detected this AI mood swing but couldn’t put my finger on what was going on until I read this post. Newton is a pure tech boosterite and Ezra is credulous, it makes sense the whispers from the AI insiders would sway them.
Also, Ezra’s Biden-admin guest spent half the episode touting the merits of AI for espionage in and statecraft. Why would the general listener root for that?
Proof that elites will use AI to serve whatever agenda they wanted to push anyways. “We must develop domestic AI so we can spy on China better than they can spy on us!”
“🆗,” as Max put it.
Yes I didn’t even get a chance to get into the other group of people who are profiting or benefiting from “AGI” nonsense talk—China hawks! In some ways I find the “will China achieve AGI before us?” stuff worse and less grounded than the paperclip-maximizer apocalypse stuff.
Anyone who recalls the original hype is going to be more skeptical going forward—and I believe we should recall it. It does tell us something—that we are not in a reliable information space when it comes to AI or AGI. It’s never going to tell you what IS true to know that, it’s just an outlook for the assessment of information.
This is a well-measured take, thank you. I also get as annoyed with the knee-jerk reaction as I do with the breathless hype. From my perspective on the ground, the thing that is changing the narrative is actually that the models are getting better.
The scaling laws are not slowing down. The people who made that claim all last year have been proven wrong again and again. scaling model parameter still works, as evidenced by gpt 4.5. scaling inference time compute works, as evidenced by reasoning models, and even scaling RL training on these reasoning models also works.
The reason people are "feeling the AGI" is that The people closest to the models are using them to do real work and seeing the effects. sure, they're not perfectly reliable. you can't just set them loose on a task indefinitely and count on it being done. But you can't do that with humans either. The agentic capabilities of reasoning models are good enough to accelerate your day-to-day work, like having an a smart and infinitely patient coworker who always responds to your messages within seconds. This now accelerates the production of more & better AI systems.
It's smart to be skeptical, but it is also smart to update on new facts. I think it is good that kind of normie journalist like Kevin and Ezra are breaking the ground for this conversation. The alternative is people claiming it's snake oil right up until it takes their jobs.(Yes, even content creators like you and me 😉 )
This is a great post but as some who is (probably inadvisably) on bsky there really are plenty of people saying precisely "AI is fake and it sucks" by which they mean it has no practical uses and anyone who says it does is an NFT-level grifter. I find myself negatively polarized against both those people and the AGI believes simultaneously hah.
Some people in this very comment section, even!
Those on BlueSky are more right than wrong. And I would go evenr further to say those prototypical posters will be the ones leading the public's apathy and eventual disdain for AI.
Slop was the word of 2024 is no coincidence.
Big Tech and silicon valley technocrats are deathly scared at the idea of the public turning against AI. AI has not even started to make much money, and is already facing application cancelling backlash.
I think technology is good for solving specific problems. I distrust technology that claims to solve every problem (“general”).
The idea that people can write ourselves out of the picture is a deluded fantasy. We’re stuck with ourselves and each other.
The people who are most obsessed with AI tend to think work can be separated from people. But people actually like and need to work. And jobs that are low-paying and mundane can actually be very fulfilling when there is a sense of community involved.
I listened to Newton’s interview on Panic World recently, in which he dunks on AI skeptics and freaks out about the potential consequences of a supposedly imminent AGI. It left a bad taste in my mouth. Part of that is because he argued that AI-skepticism is somehow rooted in a millennial inferiority complex over not being the most tech savvy person at the office any longer, which is totally frivolous, but the other reason I couldn’t put words on until I read your last paragraph.
I'm glad I'm not the only one who was really put off by that episode. I didn't have high expectations of Newton, but it was disappointing coming from Ryan Broderick.
"it’s no longer necessarily the case that, say, asking an LLM a question of fact is strictly worse than Googling it"
Isn't the difference that people don't assume the top Google result is a definitive answer? I mean, I hope most don't. A Google or other web search is the *beginning* of an investigation.
I've certainly had specific questions for which an LLM query yields worse (objectively false) results than a DuckDuckGo search plus a human sensibility reading.
I'd analogize it as follows: typing a question into a chatbot and trusting the answer : clicking the first result in Google :: using a RAG like Perplexity and checking the links : scrolling through the Google results and clicking on them.
Yeah, this is about right. And the other reason I say that LLMs are not strictly worse than Google is that ... Google is a lot worse than it used to be!
This is one of the saddest parts of the whole saga: that Google actively sabotages one of the greatest technological achievements of the last 50 years---their crown jewel---to be a middling provider in the overcrowded GenAI field.
The self-sabotage is not to be a middling provider of AI, but to make truly ungodly profits from advertising. One of the reasons I’m more bearish on AI is that it’s only a matter of time before ads pollute that knowledge source (such as it is) as well.
I wondered why my subscriptions have been spiking over the weekend. Thanks for citing my piece on AI skepticism.
Great article Max. LLMs are deception machines. Because they generate fluent and grammatically perfect text on any subject it’s easy to miss the errors/hallucinations. It makes them dangerous to use for most practical purposes, particularly as autonomous agents.
It’s an inherent problem with neural net/transformer architecture. The technology is fundamentally unreliable.
Massively underrated take. I would’ve thought doing massively complex calculations is a surer sign of intelligence than the ability to compose a limerick. But we are narcissists. We like being serenaded.
Really solid piece, and agree on AGI increasingly becoming a useless term.
An insight on the split between skeptics and believers: it’s less about your assessment of the current usefulness and power of the tech than about the future projections you buy into. The skeptics think that next year the tech won’t be much more useful than it is now. The believers foresee a decade (at least) of exponential improvement—there’s a chart in their heads with a line going up, Moore’s Law style.
And part of the projected improvement would necessarily means the tech evolves in unexpected ways way beyond whatever ChatGPT 5 brings. Recent developments in the overbroad “AI” tech tree include things like materials science discovery algorithms and better weather prediction tech—lending some credence to the believer view. So the believers also have a greater willingness to handwave away issues like LLM hallucinations or extra fingers in generative images. If these models are just stepping stones, these problems don’t seem as salient.
But at least Grok can swear now
Comedy is legal again—and it's automated!
100% accurate comedy (powered by 5G)
The Benedict Evans article is underrated for how brutal it is. The fact that the example he criticizes is from OpenAI's marketing is just perfect. "OpenAI is promising that this product can do something that it cannot do, at least, not quite, as shown by its own marketing." :chefskiss:
It’s not surprising that Ezra Klein would be easily fooled into thinking an artificial intelligence had value, he did it before with Paul Ryan. https://www.vox.com/policy-and-politics/2018/12/10/17929460/paul-ryan-speaker-retiring-debt-deficits-trump