I think that the AI companies will not reap all the benefit of AI like the car companies with exception of a few winners didn’t reap the rewards of vehicles. I think that average human beings will receive the benefit of AI as I am spending $20 per month, but reaping huge rewards from writing the emails and documents that I write with them.
> (…) reaping huge rewards from writing the emails and documents that I write with them
But are most people reaping huge rewards though? The article seems to suggest otherwise.
I’m not saying it’s useless, but we forget that humanity has survived with Grammarly, or Word, or dictionaries before that.
I will agree that ChatGPT strikes gold at a marginally higher rate than Google, but hardly orders of magnitude more. Both automation and documentation have existed and continue to do so in many forms.
They want to buy the dip that’s all. Anybody who actually uses the models with a good workflow ( specs=> unit tests=> function that passes ) sees the benefits
I used the models with a good workflow, but found them less than helpful. But I think that if I worked in a less expressive language I'd be very enthusiastic.
This makes sense to me at the moment because these AI products are enhancements, not entire replacements.
But even if they were replacements, companies do not suddenly drop all of their current processes and rewrite them from scratch using new shiny tools. And most companies are too slow to do that within the 2 years that ChatGPT/Stable Diffusion have been out anyway.
Also, things like LLMs are radically different than the assign, compare, iterate logic of current digital systems. LLMs are based on probabilities with no guarantee from one input/output combination to the next. It's pretty hard to model a process with an AI when there's no guarantee it will follow the process. Businesses are going to really struggle to adopt LLMs across their whole business like everyone is promising.
Past performance is not guarantee future results. For example, hopefully this argument wouldn't fly with an education person, "we've been pumping CO2 into the atmosphere for like 200 years now, ultimately everything is working out just fine."
It would be foolish to ignore the potential economic impacts of Gen AI. And just because we don't see the effects in the data right now, doesn't mean it won't happen.
Another Luddite take from someone who clearly doesn't engage the technology directly either as a builder or a seller.
It takes years to decades for new technologies to achieve their full effect. Factories based on steam had to be rebuilt around eletricity. That took decades. It will take time to rebuild large operations around AI.
I would also agree that the impact of AI, even in the best case scenario, will play out over years because it will take time to integrate even perfect AI into our processes. A world running on COBOL is not going to spin on a dime and run on AI, and I don't mean that sarcastically at all. That the world still runs on COBOL is an important data point in how quickly we can move things.
In fact I consider this rather obvious and borderline self-evident to a careful thinker.
However, the AI stocks are priced as if they are going to reap all the benefits, if not today, then perhaps even about a month or two ago. They are priced as if the current incumbents are the inevitable inheritors of all the promises, because the AI world is going to move so quickly to reap all this supposed value that there isn't even time for other companies to move in and displace nVidia. They are priced as if that productivity curve has already started to shoot up and it's just the beginning.
If the curve isn't even budging, then the AI world is going to need to be seriously, seriously repriced. If AI is going to change the world, but in five years rather than five months, and gradually rather than all at once, then the time value of money implies that the value to be captured in today-dollars is much, much smaller. In five years, nVidia's AI may be getting stomped by a startup that hasn't even been founded yet; I'd give it low odds, but meaningfully non-zero odds.
Whether I'm an "AI skeptic", I don't know. But I absolutely am an AI bubble skeptic. I haven't seen any evidence that the value creation is emerging at anything like the rate the bubble pricing implies. It'll probably arrive, but a lot more gradually than the market is currently believing.
> However, the AI stocks are priced as if they are going to reap all the benefits, if not today, then perhaps even about a month or two ago.
Yeah, there's actually a huge amount of metaphorical territory--if sparsely populated--between "optimistic about AI in a general sense" versus "believing all of today's market/startup hype".
However it feels like the window has been dragged so far to the hype-side that anything less than breathless stock-buying while waving a The Singularity Is Near sign will get you a "pessimist" or even "Luddite" label.
Right now, there is nothing to build around, though.
When you have an engine that consistently provides torque better and cheaper than humans or animals, you can literally build a factory around it. ML tools are currently something you occasionally try in case it gets you on track faster, but you need to have a fallback to rely on. There is nothing consistent about it.
>Apparently you haven't worked with many humans...
this goes the other way too, AI that trains on data from humans, well you're going to get the same sort of reliability haha
We get the best results from neural networks on closed systems with an infinite supply of training data (think how successful alpha zero was at chess, it simply trained by playing games if I'm not mistaken.)
Neural networks are an incredible tool. AI as a chatbot trained on human input is going to probably just simulate a lower quality human being that hallucinates more than normal lol
I didn't read it as a luddite take so much as a long winded description of what you say: there's a lot more latency in technology adoption (or even understanding by those who are developing it) than most people recognize.
Agreed; as someone working on a GenAI app in a big tech company, the main lesson I've learned is the gigantic delta between the effort required to make a intriguing prototype and a true, production ready, scalable, customer-facing, useful app. The former takes a weekend or two, the ladder can take a year or two.
The first comment I saw nailed it for me: in the 80's and 90's there was lots of investment in computers as the PC revolution took off, and the productivity gains were minimal. As with all new technologies, it takes a while for there to be macroeconomic scale changes to industry, and those changes aren't going to come without the up front investment. This article (and frankly, much of society) is looking at the ROI on the wrong time scale.
They think I have this workflow of A->B->C->D and how do I improve "B" using technology. Instead you should have something else of A->E->D where E wasn't possible or maybe just not practical.
You don't want to use technology so that you can type keys faster; eliminate the typing of keys process.
A lot of the current GenAI is around making images and summarizing text. Those largely aren't real business objectives. If you make a powerpoint 70% faster does that really close a sale 70% faster? I highly doubt it.
Powerpoints? No. But if it gets good/easy/reliable/cheap enough I can see a lot of applications. Imagine storing incoming customer emails and being able to fish out all emails complaining about, say, slow shipping, even if they don't use that exact phrase.
Funnily enough, I happen to be currently working on an agent application where the goal is to generate a PowerPoint faster. Right now it takes about a week. We should be able to do it in less than 5 minutes. We could do it faster, but I am actually reading and manipulating Excel and PowerPoint examples to make it easier to change the process or adapt the tool calls to other applications, just by editing the instructions and rough templates.
A lot of the 80's adoption of the PC meant that people did the same presentations but they looked better. ;-)
In fairness, it's not just mindset. It takes a while to understand technology and figuring out how it can be effectively harnessed. A->E->D isn't achieved simply by realizing you can do that. There's a whole sequence of steps to change the shape of A,E & D to make that happen.
But you're right, a lot of GenAI right now is not being used for things that are real game changers. This is the first step to establishing the foundation from which the real game changers will spring.
> the longer AI takes to show up in any positive economic indicators, the more it becomes the case that AI has brought increasing existential risk in exchange for minimal upside.
No? None of that is predecided. You are concluding that based on vague threats made by people who are "in the know" that can't even substantiate their own fear when confronted over it.
Existential risk comes from when us, humanity, trusts our own lives in something that is faulty. The solution is to not use or design around fault-agnostic systems, not to fear scenarios where stupid people make mistakes. AI is not red mercury, it is not a magical force multiplier that terrorists use to make lunchbox nukes.
I'm getting so tired of HN submissions like this that throw an alarmist opinion over the fence like this guy is secretly onto something. This is bunk.
It's a somewhat hilarious non sequitur, equating uselessness with existential risk. Like, my cat is also failing to show up in national economic indicators.
No it isn't; it's two separate premises, which the author uses to synthesise a conclusion. Both premises are required for the conclusion.
Neither premise is settled, and if you disagree with either, you're welcome to discard the conclusion. But you can hardly expect a blogger to start from first principles in every post.
I saw some blurb (I can't find it now) about the optimal time to launch an interstellar probe, and the punchline was that embarking too soon means you will be overtaken by a later vehicle which launched later but with better tech. When the rate of technological progress relative to the duration of a mission is above some threshold, it's rational to do more front-loaded research before actually committing to a particular plan to apply it.
In the AI case, advancement has been so rapid that to a certain extent, it makes sense for large companies to delay some aspects of retooling, because better models that come out next month may have different characteristics or requirements. We haven't seen the big shifts in productivity because reorganizing around new tech is slow, and the tech itself is still changing so fast.
Maybe, but all technology is built on what came before, so actually launching the thing and seeing what it does is a big part of enabling the next generation to be better.
I'm not sure that we disagree, but perhaps we're understanding "launching" in the metaphor somewhat differently. I think the key point about the interstellar probe example is it's assumed that once it's launched, it's not going to improve on the very long flight to Alpha Centauri. Maybe this isn't actually fair -- if you design a solar sail, pushed by a giant laser that stays in our system, you could likely upgrade the laser while the probe is in flight, but your probe can't build itself a new larger sail, or develop itself a fusion drive.
I think research is rapidly building off the prior generation of techniques, but this improvement doesn't show up in the economic impact the article author is complaining about. Actual productivity changes are going to come when large organizations integrate AI models to replace or improve key systems or human-performed functions. This is slow, expensive, and probably a minority of transformations like this are even started, let alone showing fruit, and I think it's rational for a lot of companies to defer starting this work until the tech improves further. There's also a possibly prolonged intermediate phase where some stuff is being done both by an AI-centered system and either a human or a legacy system while the organization measures how well the AI version does (and how stable/reliable any associated vendors are), and during this period, productivity may even take a measurable hit.
There was a funny (joke) academic paper circulating around 2000 that showed if you were plannig to do a computation that would take about a year that in 6 months you could pay 1/3 less and still get the computation done faster.
What made it so funny is not only that it was written so well but that things were moving so fast back then it felt plausible.
I think another analogy that applies to AI is going to orbit. It's "not hard" for a sufficiently large rocket to be able to hit orbital heights. It's pretty hard for the rocket to enter orbit at the same height. In the AI world, we are at the suborbital stage of technology. For me, the 99% use case is accessibility tools (speech recognition, grammar correction) and coding.
Image generation is like poor-quality Photoshop. Which, by the way, is the best target image generative AI could shoot for. Do what Photoshop can do, only better and easier.
>the optimal time to launch an interstellar probe, and the punchline was that embarking too soon means you will be overtaken by a later vehicle which launched later but with better tech.
"The latest estimates, using official figures, suggest that real output per employee in the median rich country is not growing at all."
If I can use an AI assistant to complete a job in 3 hours that used to take 9 hours, that's a win for me, because now I have 6 hours to spend on a personal open-source project, some family business, recreational (and healthy) exercise, or some community-beneficial volunteer activities. While this might not show up on the corporate balance sheet, it's going to make the overall society a lot healthier.
The existential risk of AI to the current socioeconomic order might have more to do with AI systems making better economic decisions than the overcompensated executive-shareholder sector does, and at far lower cost to the corporation or institution that adopts it. This makes logical sense, since the skills involved in humans climbing to positions of power in organizations (manipulation of other people, hyping up one's capabilities, clinging desperately to power, etc.) don't translate into efficient well-run operations.
It sort of depends on what you assume become table-stakes tools. Spreadsheets reduced some work, created more work (run a bunch of alternative models), and yes probably eliminated some work--for the people fluent in them. It's probably hard to argue that spreadsheets (which have become the most inarguably successful low-code tools) have markedly decreased the amount of work their users need to do in companies.
Except that’s not how it ever works. Any productivity gains will be consumed by your employer, and the expectations of your position will increase with the productivity gains.
Ie: if your role is now 75% more productive on average, the new expectation of that role is a 75% increase in outputs.
So it actually will get much worse for you, because you’ll be doing much, much more for effectively the same pay :)
There used to be a single role that did the data layer, a single role that did QA/Testing, a single role for UI/UX, a single role for client side, a single role for server side, and a devops team.
Then when the world became good at this stuff, the expectation shifted to, in many cases, a single role juggling all of this work
i was thinking more if the current crop of AI are useless, they won't get enough money pumped into them to lead to runaway amplification. if there's genuine risk, then there will be rewards. if we're close to that kind of AI, then we can't afford for our enemies to find it first.
Is this necessarily correlated? I would think that the potential to cause disaster is rather linked to whether AI is given authority to enact autonomously decisions, i.e. given access to critical systems.
If AI has no economic impact, it almost by definition also has no ability to marshal the kind of resources necessary to destroy all of humanity. All the doom scenarios seem to rely on quasi-magical abilities springing from advanced intelligence, but ignore that these scenarios require the AI to have access to significant resources.
Then, suppose AI does somehow think itself into effecting human annihilation, what happens when all the humans are dead? At the current state of things, all electricity production on the planet goes dark in maybe two weeks. Nuke plants irradiate significant parts of the earth, chip fabs are bricked, metal production is toast. In short the AI necessarily dies too.
The only way for this not to happen is AI controlled robots first gaining complete control of massive chunks of the economy. If AI is as useless as the article claims then that will effectively never happen. Thus, AI doom scenarios are always a murder suicide situation, and I’m not sure I believe anyone who says a superhumanly capable planner would pursue that plan. It even negates the “paperclip optimizer” as the AI’s demise obviously puts a hard upper limit on the number of paperclips produced.
> All the doom scenarios seem to rely on quasi-magical abilities springing from advanced intelligence
The amount of people who dismiss doom risk with some version of "I'm not smart enough to imagine what something far smarter than me might be capable of" is staggering. "I can't imagine how nanobots might work therefore they must be impossible."
>Then, suppose AI does somehow think itself into effecting human annihilation, what happens when all the humans are dead? At the current state of things, all electricity production on the planet goes dark in maybe two weeks. Nuke plants irradiate significant parts of the earth, chip fabs are bricked, metal production is toast. In short the AI necessarily dies too.
Smart enough to eradicate humanity but not smart enough to plan ahead to keep itself running. Ok buddy.
>The only way for this not to happen is AI controlled robots first gaining complete control of massive chunks of the economy. If AI is as useless as the article claims then that will effectively never happen.
"AI hasn't taken over large parts of the economy, therefore it never will." I really just can't even...
Heh when just last week we saw what happens when a single random vendor gains control over large chunks of the economy. Productivity in companies is weird it's like it seeps in slowly over years without much improvement then suddenly it's everywhere
Mostly the people dismissing the x-risk concern do so on the grounds that what is being sold is not and cannot become smart in the way it's being sold by x-risk stans as being able to, which is true. If you want to argue the contrary case, starting from the actual point of disagreement is probably a better play, because the argument you've actually made here constitutes no answer.
>what is being sold is not and cannot become smart in the way it's being sold by x-risk stans
Do you have actual evidence for this astonishing take? Because I'm sure the entire field would be entranced by your genius here. It's actually completely UN-clear what the limits to LLMs are and it's completely ridiculous to say otherwise.
>starting from the actual point of disagreement is probably a better play
Fundamentally most e/accs or other anti-doom stans believe that AGI/ASI are actually impossible, that nothing can be smarter than humans and that human brains constitute special magic meat. Almost all of their other wrong beliefs stem from those.
If the only way you can see it is on the spectrum of x-risk maximalist to e/acc, there's no point in us trying to have a conversation here. The implicit status game is meaningless to me and I can't see how any of this helps better reflect reality.
How is this a spectrum. There is indeed a spectrum of p(doom)s which represent meaningful positions about risk and then there's e/acc's, a group of misanthropic weirdos who are actively interested in pursuing the extinction of our species. If you can't see the difference then there is indeed nothing to discuss.
This is magical thinking. Like, the laws of physics still apply. Computation depends on power generation, and both *require* physical objects that break and decay. Who fixes the broken stuff when all the humans are dead?
> All the doom scenarios seem to rely on quasi-magical abilities springing from advanced intelligence, but ignore that these scenarios require the AI to have access to significant resources.
No. The doom scenario I'm most concerned with is basically the one from Manna (https://marshallbrain.com/manna), which doesn't require any "quasi-magical" abilities. It's basically taking our current economic system and "fixing the glitch" (from the capitalist's perspective) that forces them to pay our salaries.
That's (a slow) doom for the vast majority of the population, but not for humanity because the wealthy elite would survive.
Anything that removes the need to pay salaries would have massive impacts on productivity numbers. (I am not arguing that would necessarily be good[1], but it would show up in the economic stats, in line with bglazer's point).
[1] The big brain take behind "necessarily" is that: "no salaries, everything is free" seems cool, "no salaries, a few people hoard all the wealth" seems bad.
> Anything that removes the need to pay salaries would have massive impacts on productivity numbers. (I am not arguing that would necessarily be good[1], but it would show up in the economic stats, in line with bglazer's point).
I agree with that. My point was that not all "doom scenarios" require the kind of "quasi-magical abilities" that he tried to use to dismiss them all.
IMHO, the sci-fi "superintelligent AI sends terminators to kill all the people" doom scenarios are distractions (possibly intentional distractions) from far more plausible doom scenarios.
> [1] The big brain take behind "necessarily" is that: "no salaries, everything is free" seems cool, "no salaries, a few people hoard all the wealth" seems bad.
If the former were plausable, we'd almost be there even without AI, so it's either a fantasy or the work we need to get there is not technological. The latter is basically 2024 but moreso.
Or to put another way, AI can be dangerous even when It's being totally helpful... When it is enabling some of humanity to screw things up for everyone else out of greed or anger.
While I agree with you, have you found anything related to how you would essentially ration goods in a system where AI could provide and everything was free?
> But no salaries -> no consumption -> no profit for companies.
No, you just need a little more imagination. What would the economy look like if you didn't need employees to utilize resources or undertake projects? Some very mass-market consumer oriented companies would eventually go bankrupt during the transition, but that doesn't mean a doom scenario like I describe can't or won't happen.
> I don't see how that's effective capitalism.
And you know what? The end result of the transition might only vaguely resemble current-day capitalism, and may be more properly understood as a transition to a new economic system.
However, I think it would probably be like our current economy, except with:
* the B2C segment withered mainly to luxury goods of elites (yachts, supercars, elite-level fine dining) and sustaining a much smaller elite-service-focused labor pool;
* B2B being most of what's left of the market; and
* non-market vanity projects taking up a substantial slice of the pie (e.g. Elon Musk using a robot army to build ziggurats to celebrate himself, because why not).
This misunderstands the economy. It's driven by household and government spending. All that B2B stuff happens because some company, somewhere at the end of the chain, sells to households or the government.
Sure, there is also business investment. But that only happens because someone expects to sell to households or the government. If all you are selling is luxury goods, which depend on direct high labor content for their value, then most of the B2B businesses disappear also.
And yes, there are exports. But not at global scale. The globe is closed.
> This misunderstands the economy. It's driven by household and government spending. All that B2B stuff happens because some company, somewhere at the end of the chain, sells to households or the government.
No, I didn't misunderstand it. Let me put it in your terms: the remaining elite become tiny but massively consuming households, and the distinctions between the businesses they control and their households will be blurred (e.g. Blue Origin being wholly owned by Bezos). All the economic actors that cannot provide value to these households will get slowly cut off and wither away, and these elites (collectively) will eventually cement total control over all the valuable resources they have some use for (production capacity, compute, energy, raw materials, etc).
To go back to a question I asked upthread: what would the economy look like if you didn't need employees to utilize resources or undertake projects? What happens when most households aren't competitive at anything economically, but there's still a lot of stuff out there to be owned? My answer is those who own the right things (e.g. AI-run, totally automated, adaptable factories; energy resources) will be able to support themselves indefinitely with them and have massive personal surpluses they then use for vanity projects to keep themselves occupied.
> That's (a slow) doom for the vast majority of the population, but not for humanity because the wealthy elite would survive.
People throughout history have replaced the "elite" once the cost to do so became less than the cost (or benefit) to continue on in the status quo. The perceived "doom" would not last very long (in the greater span or sense of time) as it would all be reorganized, yet again. In that case, I am not so sure it's a "doom scenario" but instead something like the natural way of things, or history continuing on. And, maybe even a needed evolution of the species in that case.
Things would have to be "quasi-magical" to play out (for the rest of human eternity) in the way a lot of the AI doom musings point to.
> People throughout history have replaced the "elite" once the cost to do so became less than the cost (or benefit) to continue on in the status quo.
Past performance is not a guarantee of future success. The elite throughout history relied on masses of common people to get anything done (do work, go to war, etc.). AI as-hyped has the potential to upend that long standing dynamic.
The elite aren't going to be replaced if they and the resources they care to protect are ensconced in a 50 mile buffer patrolled by (among other things) drone-swarms that will shove grenades in the face of any unauthorized human that dares enter. That technology is pretty close to reality, as it's basically the Ukraine/Russia front lines with 5-10 more years of drone development.
> Past performance is not a guarantee of future success.
True.
I am still of the belief humans will continue to human. If an uprising of billions vs millions lose, the "elite" would also no longer be "elite".
While I have no data, other than observation, the "elite", imo, need the validation of their status more than those who have none and gave it to them in the first place. If they do something so brash as to rid the world of these unwashed masses, they have no one left to rise above. I don't see them all of a sudden becoming great members of this new society; they wanted to own society, not be a citizen in it. How that wouldn't lead to self-elimination at that point is hard to imagine - thus a needed evolution of the species (even if that's extinction and another hominid group later inheriting the leftover mess)
I think one way for an intelligent AI can gain control over humans is not to exterminate them but to control them, influence them without them being aware of it, changing the recommandation algorithms, producing educational material, videos, podcasts, books that would be pro AI while saturating our monkey brain with mining-less entertaining content. We would become their servant without noticing.
And once you've got lots of crop pickers you can transition to management. Then you can start a marketplace, then upgrade that to a chain of supermarkets, then upgrade that to a global FMCG conglomerate, and then your birth rates will also decline and population go to zero.
It's a sad realization that modern life which we deem the right way to exist is basically just wrong. That which is unsustainable by definition cannot be sustained after all.
If a miraculous new technology that helps individuals grok any conceivable topic has “no economic impact,” you should probably question how you measure of economic impact - perhaps even your very definition of the economy, the nature of human labor and collaboration. Fundamental changes at the base of the hierarchy will always escape the attention of high level managerial class, because the one thing they’re never taught to do is to consider the possibility that their inherited power structure may become irrelevant.
Arguably, but imho the difference between search engines and GPTs is akin to punch card coding vs modern IDE - theoretically the same activity, vastly different in practice.
For generating output, sure, but when it comes to "helping individuals grok any conceivable topic" like the original comment said, you're still bottlenecked by how fast you can read and understand text, so it won't make as much difference there.
Plus, from my experience, the more obscure topics, which are not easy to google, also tend to suffer the most from hallucinations when using LLMs. The topics which ChatGPT can explain perfectly usually have well-written explanations at the top of search results too.
My (snarky) point is that I do not think LLMs have dramatically enabled the average person to grok (to gain a deep and intuitive understanding of) new topics, because I haven't observed the corresponding sea-change in (mostly online) people that I would expect to go with that.
To offer a competing hypothesis: LLMs are merely augmenting or replacing other routes that people may take to acquire facile knowledge. In addition, the fundamental way those tools operate limits them to imparting facile knowledge.
I don't think we have seen the start of productivity gains. We are using it to write drafts for public competitions. It’s turned something that would be weeks of pain into days. That means the people involved can spend less time on trite and focus on editorial. It’s also let our small team take on more parallel bids with less frustration.
And presumably your competition is also now free to bid on more contracts, meaning each contract has many more bids, which means the client needs to use AI to read them all...
prompt -> your AI blackbox -> client's AI blackbox -> client picks winner
Maybe someday the client will let you submit your prompt directly to their AI blackbox so you can save computing resources.
Please submit your application in the form of an executable binary under 100GB containing an interactive agent which will convince us to give you money!
Are there any independent statistics for how much AI based coding assistants like GitHub CoPilot improve an average software developer‘s productivity?
I had hoped for something around 5%, which is not massive, but still a non-trivial number.
Not sure how AI assistants in other “high value” professions far. Say in legal, healthcare, finance, civil engineering, business consulting to name a few. Does anyone have reliable figures?
1) Current AI, mostly LLMs, not much economic impact, no extinction risk.
2) Sci-fi like future ASI, big impact, some existential risk.
Discussing the second is tricky as we don't quite know how it will go and will probably change all sorts of things. But re 'we are all going to die' type stuff, currently we will with certainty due to age, with sci-fi AI, not necessarily. So that kind of p(doom) is much lower with ASI.
The problems with the current generation and rollout of "AI" which I am using synonymously in this post with large language models, because that's the context it seems to be mostly used in print media - are so numerous that I could not list them out all in one post and this has been debated to an absurd degree already, on this site and on other platforms.
I really can't think of many, if any real positives. One example I like to use is in video game tutorial articles. I don't know if you've tried to use basically any search engine to find reliable video game information/walkthroughs lately, but the amount of blatant misinformation that has proliferated in this space makes searching the web for even the most basic information an exercise in complete futility.
This is something "AI" should be good at. Video game walkthroughs have been done to death, a game that's been out for 2+ years has likely already had every single secret and path in it discovered 10,000 times over and posted all over reddit and other sites. Yet, after completing my 2nd elden ring run this weekend, my first since 2022, I was struck at how much blatantly wrong info I stumbled across, even on websites I had considered "authoritative." And it's always subtle enough to be impossibly annoying - stuff like "head south from this site" when in fact, it should have said "north" Very small errors that seem like the result of hallucinations that cause an inordinate amount of time spent figuring out what's wrong - much like my experience with coding "assistants."
If it can't get something like this correct - which I'm aware isn't necessarily a purely "AI" problem - how do we expect it to do anything important? I really have not seen a solid case yet and would love to be convinced otherwise, but the more I learn the less I like what I see.
Perhaps we should question how productivity is measured and learn to control all the various factors. I think there is much more than AI at play here, and in fact i would guess that AI has been a major factor in offsetting (preventing) an overall decrease in "productivity" (as currently defined).
100 comments
[ 3.3 ms ] story [ 173 ms ] threadBut are most people reaping huge rewards though? The article seems to suggest otherwise.
I’m not saying it’s useless, but we forget that humanity has survived with Grammarly, or Word, or dictionaries before that.
I will agree that ChatGPT strikes gold at a marginally higher rate than Google, but hardly orders of magnitude more. Both automation and documentation have existed and continue to do so in many forms.
But even if they were replacements, companies do not suddenly drop all of their current processes and rewrite them from scratch using new shiny tools. And most companies are too slow to do that within the 2 years that ChatGPT/Stable Diffusion have been out anyway.
Ultimately everything working out just fine. AI is not going to doom us.
AI's gardening skills are so bad, even the weeds can't grow.
Doomsayers can now downvote me.
Yeah, the future is farming and mining, because AI is trained on all the fun jobs. "Stable boy," my ass.
It would be foolish to ignore the potential economic impacts of Gen AI. And just because we don't see the effects in the data right now, doesn't mean it won't happen.
It takes years to decades for new technologies to achieve their full effect. Factories based on steam had to be rebuilt around eletricity. That took decades. It will take time to rebuild large operations around AI.
In fact I consider this rather obvious and borderline self-evident to a careful thinker.
However, the AI stocks are priced as if they are going to reap all the benefits, if not today, then perhaps even about a month or two ago. They are priced as if the current incumbents are the inevitable inheritors of all the promises, because the AI world is going to move so quickly to reap all this supposed value that there isn't even time for other companies to move in and displace nVidia. They are priced as if that productivity curve has already started to shoot up and it's just the beginning.
If the curve isn't even budging, then the AI world is going to need to be seriously, seriously repriced. If AI is going to change the world, but in five years rather than five months, and gradually rather than all at once, then the time value of money implies that the value to be captured in today-dollars is much, much smaller. In five years, nVidia's AI may be getting stomped by a startup that hasn't even been founded yet; I'd give it low odds, but meaningfully non-zero odds.
Whether I'm an "AI skeptic", I don't know. But I absolutely am an AI bubble skeptic. I haven't seen any evidence that the value creation is emerging at anything like the rate the bubble pricing implies. It'll probably arrive, but a lot more gradually than the market is currently believing.
Yeah, there's actually a huge amount of metaphorical territory--if sparsely populated--between "optimistic about AI in a general sense" versus "believing all of today's market/startup hype".
However it feels like the window has been dragged so far to the hype-side that anything less than breathless stock-buying while waving a The Singularity Is Near sign will get you a "pessimist" or even "Luddite" label.
When you have an engine that consistently provides torque better and cheaper than humans or animals, you can literally build a factory around it. ML tools are currently something you occasionally try in case it gets you on track faster, but you need to have a fallback to rely on. There is nothing consistent about it.
At least yet.
Apparently you haven't worked with many humans...
this goes the other way too, AI that trains on data from humans, well you're going to get the same sort of reliability haha
We get the best results from neural networks on closed systems with an infinite supply of training data (think how successful alpha zero was at chess, it simply trained by playing games if I'm not mistaken.)
Neural networks are an incredible tool. AI as a chatbot trained on human input is going to probably just simulate a lower quality human being that hallucinates more than normal lol
I work in AI and I think he’s right to a degree.
They think I have this workflow of A->B->C->D and how do I improve "B" using technology. Instead you should have something else of A->E->D where E wasn't possible or maybe just not practical.
You don't want to use technology so that you can type keys faster; eliminate the typing of keys process.
A lot of the current GenAI is around making images and summarizing text. Those largely aren't real business objectives. If you make a powerpoint 70% faster does that really close a sale 70% faster? I highly doubt it.
isn't that just vectorized or "semantic" searching though?
windows is still exact string match but at least allows partial or reordered string match
filtering an excel column is exact substring only
In fairness, it's not just mindset. It takes a while to understand technology and figuring out how it can be effectively harnessed. A->E->D isn't achieved simply by realizing you can do that. There's a whole sequence of steps to change the shape of A,E & D to make that happen.
But you're right, a lot of GenAI right now is not being used for things that are real game changers. This is the first step to establishing the foundation from which the real game changers will spring.
No? None of that is predecided. You are concluding that based on vague threats made by people who are "in the know" that can't even substantiate their own fear when confronted over it.
Existential risk comes from when us, humanity, trusts our own lives in something that is faulty. The solution is to not use or design around fault-agnostic systems, not to fear scenarios where stupid people make mistakes. AI is not red mercury, it is not a magical force multiplier that terrorists use to make lunchbox nukes.
I'm getting so tired of HN submissions like this that throw an alarmist opinion over the fence like this guy is secretly onto something. This is bunk.
Neither premise is settled, and if you disagree with either, you're welcome to discard the conclusion. But you can hardly expect a blogger to start from first principles in every post.
In the AI case, advancement has been so rapid that to a certain extent, it makes sense for large companies to delay some aspects of retooling, because better models that come out next month may have different characteristics or requirements. We haven't seen the big shifts in productivity because reorganizing around new tech is slow, and the tech itself is still changing so fast.
I think research is rapidly building off the prior generation of techniques, but this improvement doesn't show up in the economic impact the article author is complaining about. Actual productivity changes are going to come when large organizations integrate AI models to replace or improve key systems or human-performed functions. This is slow, expensive, and probably a minority of transformations like this are even started, let alone showing fruit, and I think it's rational for a lot of companies to defer starting this work until the tech improves further. There's also a possibly prolonged intermediate phase where some stuff is being done both by an AI-centered system and either a human or a legacy system while the organization measures how well the AI version does (and how stable/reliable any associated vendors are), and during this period, productivity may even take a measurable hit.
What made it so funny is not only that it was written so well but that things were moving so fast back then it felt plausible.
Image generation is like poor-quality Photoshop. Which, by the way, is the best target image generative AI could shoot for. Do what Photoshop can do, only better and easier.
That's the wait calculation
https://en.m.wikipedia.org/w/index.php?title=Interstellar_tr...
If I can use an AI assistant to complete a job in 3 hours that used to take 9 hours, that's a win for me, because now I have 6 hours to spend on a personal open-source project, some family business, recreational (and healthy) exercise, or some community-beneficial volunteer activities. While this might not show up on the corporate balance sheet, it's going to make the overall society a lot healthier.
The existential risk of AI to the current socioeconomic order might have more to do with AI systems making better economic decisions than the overcompensated executive-shareholder sector does, and at far lower cost to the corporation or institution that adopts it. This makes logical sense, since the skills involved in humans climbing to positions of power in organizations (manipulation of other people, hyping up one's capabilities, clinging desperately to power, etc.) don't translate into efficient well-run operations.
Ie: if your role is now 75% more productive on average, the new expectation of that role is a 75% increase in outputs.
So it actually will get much worse for you, because you’ll be doing much, much more for effectively the same pay :)
There used to be a single role that did the data layer, a single role that did QA/Testing, a single role for UI/UX, a single role for client side, a single role for server side, and a devops team.
Then when the world became good at this stuff, the expectation shifted to, in many cases, a single role juggling all of this work
Then, suppose AI does somehow think itself into effecting human annihilation, what happens when all the humans are dead? At the current state of things, all electricity production on the planet goes dark in maybe two weeks. Nuke plants irradiate significant parts of the earth, chip fabs are bricked, metal production is toast. In short the AI necessarily dies too.
The only way for this not to happen is AI controlled robots first gaining complete control of massive chunks of the economy. If AI is as useless as the article claims then that will effectively never happen. Thus, AI doom scenarios are always a murder suicide situation, and I’m not sure I believe anyone who says a superhumanly capable planner would pursue that plan. It even negates the “paperclip optimizer” as the AI’s demise obviously puts a hard upper limit on the number of paperclips produced.
The amount of people who dismiss doom risk with some version of "I'm not smart enough to imagine what something far smarter than me might be capable of" is staggering. "I can't imagine how nanobots might work therefore they must be impossible."
>Then, suppose AI does somehow think itself into effecting human annihilation, what happens when all the humans are dead? At the current state of things, all electricity production on the planet goes dark in maybe two weeks. Nuke plants irradiate significant parts of the earth, chip fabs are bricked, metal production is toast. In short the AI necessarily dies too.
Smart enough to eradicate humanity but not smart enough to plan ahead to keep itself running. Ok buddy.
>The only way for this not to happen is AI controlled robots first gaining complete control of massive chunks of the economy. If AI is as useless as the article claims then that will effectively never happen.
"AI hasn't taken over large parts of the economy, therefore it never will." I really just can't even...
Do you have actual evidence for this astonishing take? Because I'm sure the entire field would be entranced by your genius here. It's actually completely UN-clear what the limits to LLMs are and it's completely ridiculous to say otherwise.
>starting from the actual point of disagreement is probably a better play
Fundamentally most e/accs or other anti-doom stans believe that AGI/ASI are actually impossible, that nothing can be smarter than humans and that human brains constitute special magic meat. Almost all of their other wrong beliefs stem from those.
No. The doom scenario I'm most concerned with is basically the one from Manna (https://marshallbrain.com/manna), which doesn't require any "quasi-magical" abilities. It's basically taking our current economic system and "fixing the glitch" (from the capitalist's perspective) that forces them to pay our salaries.
That's (a slow) doom for the vast majority of the population, but not for humanity because the wealthy elite would survive.
[1] The big brain take behind "necessarily" is that: "no salaries, everything is free" seems cool, "no salaries, a few people hoard all the wealth" seems bad.
I agree with that. My point was that not all "doom scenarios" require the kind of "quasi-magical abilities" that he tried to use to dismiss them all.
IMHO, the sci-fi "superintelligent AI sends terminators to kill all the people" doom scenarios are distractions (possibly intentional distractions) from far more plausible doom scenarios.
> [1] The big brain take behind "necessarily" is that: "no salaries, everything is free" seems cool, "no salaries, a few people hoard all the wealth" seems bad.
If the former were plausable, we'd almost be there even without AI, so it's either a fantasy or the work we need to get there is not technological. The latter is basically 2024 but moreso.
No, you just need a little more imagination. What would the economy look like if you didn't need employees to utilize resources or undertake projects? Some very mass-market consumer oriented companies would eventually go bankrupt during the transition, but that doesn't mean a doom scenario like I describe can't or won't happen.
> I don't see how that's effective capitalism.
And you know what? The end result of the transition might only vaguely resemble current-day capitalism, and may be more properly understood as a transition to a new economic system.
However, I think it would probably be like our current economy, except with:
* the B2C segment withered mainly to luxury goods of elites (yachts, supercars, elite-level fine dining) and sustaining a much smaller elite-service-focused labor pool;
* B2B being most of what's left of the market; and
* non-market vanity projects taking up a substantial slice of the pie (e.g. Elon Musk using a robot army to build ziggurats to celebrate himself, because why not).
Sure, there is also business investment. But that only happens because someone expects to sell to households or the government. If all you are selling is luxury goods, which depend on direct high labor content for their value, then most of the B2B businesses disappear also.
And yes, there are exports. But not at global scale. The globe is closed.
No, I didn't misunderstand it. Let me put it in your terms: the remaining elite become tiny but massively consuming households, and the distinctions between the businesses they control and their households will be blurred (e.g. Blue Origin being wholly owned by Bezos). All the economic actors that cannot provide value to these households will get slowly cut off and wither away, and these elites (collectively) will eventually cement total control over all the valuable resources they have some use for (production capacity, compute, energy, raw materials, etc).
To go back to a question I asked upthread: what would the economy look like if you didn't need employees to utilize resources or undertake projects? What happens when most households aren't competitive at anything economically, but there's still a lot of stuff out there to be owned? My answer is those who own the right things (e.g. AI-run, totally automated, adaptable factories; energy resources) will be able to support themselves indefinitely with them and have massive personal surpluses they then use for vanity projects to keep themselves occupied.
People throughout history have replaced the "elite" once the cost to do so became less than the cost (or benefit) to continue on in the status quo. The perceived "doom" would not last very long (in the greater span or sense of time) as it would all be reorganized, yet again. In that case, I am not so sure it's a "doom scenario" but instead something like the natural way of things, or history continuing on. And, maybe even a needed evolution of the species in that case.
Things would have to be "quasi-magical" to play out (for the rest of human eternity) in the way a lot of the AI doom musings point to.
Past performance is not a guarantee of future success. The elite throughout history relied on masses of common people to get anything done (do work, go to war, etc.). AI as-hyped has the potential to upend that long standing dynamic.
The elite aren't going to be replaced if they and the resources they care to protect are ensconced in a 50 mile buffer patrolled by (among other things) drone-swarms that will shove grenades in the face of any unauthorized human that dares enter. That technology is pretty close to reality, as it's basically the Ukraine/Russia front lines with 5-10 more years of drone development.
True.
I am still of the belief humans will continue to human. If an uprising of billions vs millions lose, the "elite" would also no longer be "elite".
While I have no data, other than observation, the "elite", imo, need the validation of their status more than those who have none and gave it to them in the first place. If they do something so brash as to rid the world of these unwashed masses, they have no one left to rise above. I don't see them all of a sudden becoming great members of this new society; they wanted to own society, not be a citizen in it. How that wouldn't lead to self-elimination at that point is hard to imagine - thus a needed evolution of the species (even if that's extinction and another hominid group later inheriting the leftover mess)
These would have a mildly depressing effect on the economy for a couple of generations, and then the population goes to zero.
much of the world is still crapping out kids just fine. subsistence farmers feel more confident doing that than programmers, it seems.
It's a sad realization that modern life which we deem the right way to exist is basically just wrong. That which is unsustainable by definition cannot be sustained after all.
So they do that because they plan to annihilate all humans. They succeed in both.
How does the AI deal with physics and the heat death of the universe?
If you think that's happened, then you and I are living on very different Internets with different individuals. :P
Plus, from my experience, the more obscure topics, which are not easy to google, also tend to suffer the most from hallucinations when using LLMs. The topics which ChatGPT can explain perfectly usually have well-written explanations at the top of search results too.
To offer a competing hypothesis: LLMs are merely augmenting or replacing other routes that people may take to acquire facile knowledge. In addition, the fundamental way those tools operate limits them to imparting facile knowledge.
prompt -> your AI blackbox -> client's AI blackbox -> client picks winner
Maybe someday the client will let you submit your prompt directly to their AI blackbox so you can save computing resources.
I had hoped for something around 5%, which is not massive, but still a non-trivial number.
Not sure how AI assistants in other “high value” professions far. Say in legal, healthcare, finance, civil engineering, business consulting to name a few. Does anyone have reliable figures?
1) Current AI, mostly LLMs, not much economic impact, no extinction risk.
2) Sci-fi like future ASI, big impact, some existential risk.
Discussing the second is tricky as we don't quite know how it will go and will probably change all sorts of things. But re 'we are all going to die' type stuff, currently we will with certainty due to age, with sci-fi AI, not necessarily. So that kind of p(doom) is much lower with ASI.
I really can't think of many, if any real positives. One example I like to use is in video game tutorial articles. I don't know if you've tried to use basically any search engine to find reliable video game information/walkthroughs lately, but the amount of blatant misinformation that has proliferated in this space makes searching the web for even the most basic information an exercise in complete futility.
This is something "AI" should be good at. Video game walkthroughs have been done to death, a game that's been out for 2+ years has likely already had every single secret and path in it discovered 10,000 times over and posted all over reddit and other sites. Yet, after completing my 2nd elden ring run this weekend, my first since 2022, I was struck at how much blatantly wrong info I stumbled across, even on websites I had considered "authoritative." And it's always subtle enough to be impossibly annoying - stuff like "head south from this site" when in fact, it should have said "north" Very small errors that seem like the result of hallucinations that cause an inordinate amount of time spent figuring out what's wrong - much like my experience with coding "assistants."
If it can't get something like this correct - which I'm aware isn't necessarily a purely "AI" problem - how do we expect it to do anything important? I really have not seen a solid case yet and would love to be convinced otherwise, but the more I learn the less I like what I see.