This is getting spooky. Once again my thoughts are just about 100% aligned with yours. I have been saying I am an AI Optimist for a long while now. My optimism is based on real world usage of GenAI tools at work (in cybersecurity) and outside of work.
On the AGI vs Useful AI front, I almost feel like there's a little bit of click bait type headlines approach from at least some writers- while obsessively following and being a very active user in the GenAI space for the last 20-ish months, I have never once expected AGI to arrive imminently. Never once felt impatient to see its arrival; if anything we're often lead to believe that the arrival of AGI will lead to more job losses etc.
For whatever it's worth I'm at the other end of the bored with or disappointed by GenAI spectrum. I am embracing/living this wisdom from Ethan Mollick:
"“always invite AI to the table” is the principle in my book that people tell me had the biggest impact on them. You won’t know what AI can (and can’t) do for you until you try to use it for everything you do. "
My main point is that GenAI doesn't have to live up to sky-high expectations or reach AGI to be a useful tool if you're willing to try it.
At the same time, I acknowledge that not everyone can seamlessly integrate AI into their routine and it doesn't always come natural to people. Which is why I often try to post stuff that hopefully nudges people from "AI is a complicated black box" to "I know how AI can help me with this specific thing."
"Just because you’re getting tired of these repetitive analogies, it doesn’t mean they didn’t help me make my point." <-- probably my favorite part; don't stop being silly!
I think the dot com bubble offers really good parallels. The internet absolutely changed the world over the last 25 years, but it was far from the overnight transition we were promised back then. This revolution, too, will take a bit of time to propagate (albeit almost certainly less since we now have an internet to begin with).
Don't think I could stop being silly if I tried, to be honest.
And exactly: I actually had the dot com bubble parallel in my early draft of the post but I couldn't find a very elegant way to work it into the narrative (partially because AI is different in many ways, as you also point out.) But both are manifestations of Amara's law.
One additional thing I'll say about financial bubbles: it's often very easy to pick the right trend, but it's almost always difficult to pick the right company.
People are the long pole here. Habits, business processes, norms; we are slow to change. The models will get better, gen AI integration and products will get winnowed out - who knows what the killer genai product is but we haven’t seen it yet. On the Nvidia front, I’m up here in Portland and people are shook at how Intel unraveled. Changes to the industry are deep but the ripples all the way out to end users, especially consumers take time. I like the Amara quote here.
I think when ChatGPT first came out, people saw this genie that can just do anything for them with little effort from their side. Now that all the things like "hallucinations" and shortcomings have surfaced, the initial magic feeling has gone and many people haven't put in the effort to learn what it takes to work around those downsides.
But I'm sure deep integration and an upwards learning curve are coming, even if it's not on the crazy exponential schedule some have thought it'd be.
It's my understanding that there's already a growing trend towards faster, cheaper, more specialized models that aren't as intensive to train and run. Which probably isn't the best news for NVIDIA in the long run, but they've certainly ridden the current frenzy to the top of the stock market for a while.
AI is already highly useful but not dummy proof. It takes a bit of effort and certainly nuance to learn and tease out the good from the bad. The issue is the hype made it sound easy and then, coupled with the panic, some people dove in too fast and others refuse to look.
100%. If it was fully intuitive, I'd have little to write about.
Also, I think a lot of people tried the free version of ChatGPT when it first came out (GPT-3.5) and figured it wasn't that impressive, so many of them never ended up getting a sense of what frontier models are capable of these days.
Like Ethan Mollick says, you need to invest about 10 hours into playing with AI before you can get a good sense of what it can and can't help you with. Not that many people invest this much time.
Brilliant piece. I agree with Gary Marcus, that the problem with ANY machine learning algorithm is that they're very bad at managing outliers. Making sure the LLM follows a symbolic set of rules is outside the scope of current machine-learning algorithms -- we need humans for that.
But most businesses are not ready for that. Most workers use AI, but they keep it a secret, yet productivity metrics and stock prices are not following the assumed productivity jump coming with AI.
Business processes are not ready to embrace AI, or "invite AI to the table" as Mollick says.
1. Generative AI as it stands doesn't appear to be on the path to ever managing outliers, eliminating hallucinations, or forming a coherent picture of the world. That's why alternative architectures and approaches are being explored.
2. Businesses (and educational institutions, and government organizations, etc.) don't quite know what to do with AI at the moment. Which I believe is exactly what Ethan's talking about when he mentions that it'll take many years to integrate AI and reap the benefits.
As for the "invite AI to the table" phrase, Ethan (and I) use it as a call to action for the individual user rather than a business. It's much easier for e.g. "Suzy in accounting" or "Mark in marketing" to grab an off-the-shelf AI tool to see if it solves a subset of their daily tasks than it is to get an organization to embrace AI wholesale and incorporate it at a general level.
Thanks for the response Daniel. I loved Andrey Karpathy's take on hallucination.
He said on X that "I always struggle a bit with I'm asked about the "hallucination problem" in LLMs. Because, in some sense, hallucination is all LLMs do. They are dream machines."
From a software perspective we are missing a key element in software architecture: one that can maximize consistency of neural networks by remaining within the parameters of a symbolic set of rules across modalities.
I've spent the last 18 months teaching people like "Suzy or Mark" (both online and in academia) how to invite AI to the table, but I just quit recently doing that. Because it's more complicated than that.
We are still lacking both vocabulary and paradigm that would allow businesses to embrace the changes needed to reap the benefits of AI.
The biggest bottleneck seems to be business as usual.
Yeah, the consensus seems to very much be that hallucinations are "a feature, not a bug" when it comes to LLMs. They're "plausible token prediction" machines rather than reasoning engines, so lying convincingly is kind of a built-in element.
What kind of challenges have you encountered trying to teach people how to use AI? Sounds like you've spend a good chunk of time working hands-on with this!
I think what makes it especially difficult is that people's routine and what AI can help with is a very individual thing.
For instance, I mostly use LLMs to act as beta readers for my drafts or sometimes to brainstorm top-level ideas. I haven't found a reliable way in my own routine to use something like Perplexity for thorough research or to get any chatbot to suggest written text that was anywhere close to where I'd need it to be.
At the same time, I know of people who swear by Perplexity or Gemini and say it makes their research significantly faster, etc.
Which is why I recommend people to just spend some time with AI and figure out how it can slot seamlessly into their routine. The answer may well end up being "It can't at all," but you can only learn that after trying and failing.
The biggest challenge is the ability to reframe problems.
I've done hundreds of consulting hours and they were always similar: Student comes with a problem, I whip up a frontier model of their choice and just start using it some creative way.
Instead of asking me to write me an article, at first I ask it to ask me questions. Then generate some rules. Then a table of content, etc.
You can teach all sorts of prompting techniques to people, but if they are unable to reframe and compartmentalize their problems, they will hardly be able to derive value from LLMs.
In my school we educated over 10k people and most of them are now confident with their use of LLMs but cannot really tangibly explain the value they get from it.
So I agree with you, it's very individual. We're at the stage where we're experimenting. Kind of like a steampunk era of AI, where we're trying to do everything and just FAFO. This is why I decided to focus on helping people directly one by one actually go all the way to seizing the productivity increase. Some needs automation, some needs documented processes, some are more convoluted narrow agents. It's very unique.
As Mollick says, after 10 hours of playing around you'll get some sense of what it can do to you. But I think there's still a chasm of actually achieving what you sensed.
I always scoffed at the term "prompt engineering" and felt it was overrated as a discipline. I'd say things like "It can basically understand natural language, so just tell it what you need."
But I think that grossly overestimates people's general ability to
a) Understand precisely what it is they're after in the first place.
b) Verbalize it in an effective way.
c) Be ready to work iteratively with LLMs if they don't get it right on first take (which is typically the case).
So yeah, your experience tracks!
I still feel there's lots of untapped potential, but the path to unlocking it isn't nearly as clearcut as we'd been led to believe.
I’m very apprehensive with AGI given how terrible AI is whenever it’s pushed to more complicated limits. Not in a nefarious oh no the robots are coming for us way, but in a the emperor has no clothes way. It’s just seems to really louse things up while people seem to think it’s great. I’m willing to be convinced otherwise though and when I read the part of “LLMs are an off ramp”, it seemed instinctively right though I have no clue what the alternative would be. I’d be interested in learning more. Because the independent thinking stuff is coming no matter what. I’d rather think we’re trying to set things up for success than throw some shit at the wall and see what sticks. That’s fine for some things but maybe not right for handling something like… um… everything?
Nah, you're spot on. Which is very much why LLMs are increasingly seen as a dead-end if AGI is what we're after. If LLMs can't reliably outperform humans on my tasks, then they're by the very definition of the word not AGI.
Also, while I make the point that LLMs are still useful to individuals despite their shortcomings, what I don't address is whether taking this off-ramp is justified when you take into account the financial and energy costs, negative societal impacts, and the fact that they're potentially a "distraction" if our sights are set on useful AGI.
This is getting spooky. Once again my thoughts are just about 100% aligned with yours. I have been saying I am an AI Optimist for a long while now. My optimism is based on real world usage of GenAI tools at work (in cybersecurity) and outside of work.
On the AGI vs Useful AI front, I almost feel like there's a little bit of click bait type headlines approach from at least some writers- while obsessively following and being a very active user in the GenAI space for the last 20-ish months, I have never once expected AGI to arrive imminently. Never once felt impatient to see its arrival; if anything we're often lead to believe that the arrival of AGI will lead to more job losses etc.
For whatever it's worth I'm at the other end of the bored with or disappointed by GenAI spectrum. I am embracing/living this wisdom from Ethan Mollick:
"“always invite AI to the table” is the principle in my book that people tell me had the biggest impact on them. You won’t know what AI can (and can’t) do for you until you try to use it for everything you do. "
Spot on!
My main point is that GenAI doesn't have to live up to sky-high expectations or reach AGI to be a useful tool if you're willing to try it.
At the same time, I acknowledge that not everyone can seamlessly integrate AI into their routine and it doesn't always come natural to people. Which is why I often try to post stuff that hopefully nudges people from "AI is a complicated black box" to "I know how AI can help me with this specific thing."
Yep, I try to do some of the same - especially for people in cyber.
To answer the second question, someone who’s better at math than I am should run a regression discontinuity study.
Yes, I too know some of those words!
Good stuff, and much needed.
"Just because you’re getting tired of these repetitive analogies, it doesn’t mean they didn’t help me make my point." <-- probably my favorite part; don't stop being silly!
I think the dot com bubble offers really good parallels. The internet absolutely changed the world over the last 25 years, but it was far from the overnight transition we were promised back then. This revolution, too, will take a bit of time to propagate (albeit almost certainly less since we now have an internet to begin with).
Don't think I could stop being silly if I tried, to be honest.
And exactly: I actually had the dot com bubble parallel in my early draft of the post but I couldn't find a very elegant way to work it into the narrative (partially because AI is different in many ways, as you also point out.) But both are manifestations of Amara's law.
One additional thing I'll say about financial bubbles: it's often very easy to pick the right trend, but it's almost always difficult to pick the right company.
I'm still holding on to my Enron shares. I hear they're getting more valuable by the minute!
People are the long pole here. Habits, business processes, norms; we are slow to change. The models will get better, gen AI integration and products will get winnowed out - who knows what the killer genai product is but we haven’t seen it yet. On the Nvidia front, I’m up here in Portland and people are shook at how Intel unraveled. Changes to the industry are deep but the ripples all the way out to end users, especially consumers take time. I like the Amara quote here.
Yup, that's a very good observation.
I think when ChatGPT first came out, people saw this genie that can just do anything for them with little effort from their side. Now that all the things like "hallucinations" and shortcomings have surfaced, the initial magic feeling has gone and many people haven't put in the effort to learn what it takes to work around those downsides.
But I'm sure deep integration and an upwards learning curve are coming, even if it's not on the crazy exponential schedule some have thought it'd be.
Appreciate the breakdown here, what say u about NVDA and the super computer gold rush?
I don't have a strong opinion about that, to be honest.
Clearly training LLMs is extremely GPU intensive and we're stretching the supply chain to continue rolling out frontier models.
Is there something specific you're thinking of?
Was thinking about their ability to compete as the tech shifts; I was hearing NVDA was mostly supplying hardware for the initials stages of AI
It's my understanding that there's already a growing trend towards faster, cheaper, more specialized models that aren't as intensive to train and run. Which probably isn't the best news for NVIDIA in the long run, but they've certainly ridden the current frenzy to the top of the stock market for a while.
AI is already highly useful but not dummy proof. It takes a bit of effort and certainly nuance to learn and tease out the good from the bad. The issue is the hype made it sound easy and then, coupled with the panic, some people dove in too fast and others refuse to look.
100%. If it was fully intuitive, I'd have little to write about.
Also, I think a lot of people tried the free version of ChatGPT when it first came out (GPT-3.5) and figured it wasn't that impressive, so many of them never ended up getting a sense of what frontier models are capable of these days.
Like Ethan Mollick says, you need to invest about 10 hours into playing with AI before you can get a good sense of what it can and can't help you with. Not that many people invest this much time.
And you have to be willing to work with it like a brand new intern who doesn't know nearly as much as you.
Yup, “alien intern” is the exact term Ethan Mollick uses too.
Brilliant piece. I agree with Gary Marcus, that the problem with ANY machine learning algorithm is that they're very bad at managing outliers. Making sure the LLM follows a symbolic set of rules is outside the scope of current machine-learning algorithms -- we need humans for that.
But most businesses are not ready for that. Most workers use AI, but they keep it a secret, yet productivity metrics and stock prices are not following the assumed productivity jump coming with AI.
Business processes are not ready to embrace AI, or "invite AI to the table" as Mollick says.
It's like trying to implement Excel in 1985.
Thanks, David!
And you're right on both counts.
1. Generative AI as it stands doesn't appear to be on the path to ever managing outliers, eliminating hallucinations, or forming a coherent picture of the world. That's why alternative architectures and approaches are being explored.
2. Businesses (and educational institutions, and government organizations, etc.) don't quite know what to do with AI at the moment. Which I believe is exactly what Ethan's talking about when he mentions that it'll take many years to integrate AI and reap the benefits.
As for the "invite AI to the table" phrase, Ethan (and I) use it as a call to action for the individual user rather than a business. It's much easier for e.g. "Suzy in accounting" or "Mark in marketing" to grab an off-the-shelf AI tool to see if it solves a subset of their daily tasks than it is to get an organization to embrace AI wholesale and incorporate it at a general level.
Thanks for the response Daniel. I loved Andrey Karpathy's take on hallucination.
He said on X that "I always struggle a bit with I'm asked about the "hallucination problem" in LLMs. Because, in some sense, hallucination is all LLMs do. They are dream machines."
From a software perspective we are missing a key element in software architecture: one that can maximize consistency of neural networks by remaining within the parameters of a symbolic set of rules across modalities.
I've spent the last 18 months teaching people like "Suzy or Mark" (both online and in academia) how to invite AI to the table, but I just quit recently doing that. Because it's more complicated than that.
We are still lacking both vocabulary and paradigm that would allow businesses to embrace the changes needed to reap the benefits of AI.
The biggest bottleneck seems to be business as usual.
Yeah, the consensus seems to very much be that hallucinations are "a feature, not a bug" when it comes to LLMs. They're "plausible token prediction" machines rather than reasoning engines, so lying convincingly is kind of a built-in element.
What kind of challenges have you encountered trying to teach people how to use AI? Sounds like you've spend a good chunk of time working hands-on with this!
I think what makes it especially difficult is that people's routine and what AI can help with is a very individual thing.
For instance, I mostly use LLMs to act as beta readers for my drafts or sometimes to brainstorm top-level ideas. I haven't found a reliable way in my own routine to use something like Perplexity for thorough research or to get any chatbot to suggest written text that was anywhere close to where I'd need it to be.
At the same time, I know of people who swear by Perplexity or Gemini and say it makes their research significantly faster, etc.
Which is why I recommend people to just spend some time with AI and figure out how it can slot seamlessly into their routine. The answer may well end up being "It can't at all," but you can only learn that after trying and failing.
The biggest challenge is the ability to reframe problems.
I've done hundreds of consulting hours and they were always similar: Student comes with a problem, I whip up a frontier model of their choice and just start using it some creative way.
Instead of asking me to write me an article, at first I ask it to ask me questions. Then generate some rules. Then a table of content, etc.
You can teach all sorts of prompting techniques to people, but if they are unable to reframe and compartmentalize their problems, they will hardly be able to derive value from LLMs.
In my school we educated over 10k people and most of them are now confident with their use of LLMs but cannot really tangibly explain the value they get from it.
So I agree with you, it's very individual. We're at the stage where we're experimenting. Kind of like a steampunk era of AI, where we're trying to do everything and just FAFO. This is why I decided to focus on helping people directly one by one actually go all the way to seizing the productivity increase. Some needs automation, some needs documented processes, some are more convoluted narrow agents. It's very unique.
As Mollick says, after 10 hours of playing around you'll get some sense of what it can do to you. But I think there's still a chasm of actually achieving what you sensed.
Ah, that makes a lot of sense!
I always scoffed at the term "prompt engineering" and felt it was overrated as a discipline. I'd say things like "It can basically understand natural language, so just tell it what you need."
But I think that grossly overestimates people's general ability to
a) Understand precisely what it is they're after in the first place.
b) Verbalize it in an effective way.
c) Be ready to work iteratively with LLMs if they don't get it right on first take (which is typically the case).
So yeah, your experience tracks!
I still feel there's lots of untapped potential, but the path to unlocking it isn't nearly as clearcut as we'd been led to believe.
I'd wager most people fail at a)
Which is a millennia-old problem, LLMs just make it more striking
I’m very apprehensive with AGI given how terrible AI is whenever it’s pushed to more complicated limits. Not in a nefarious oh no the robots are coming for us way, but in a the emperor has no clothes way. It’s just seems to really louse things up while people seem to think it’s great. I’m willing to be convinced otherwise though and when I read the part of “LLMs are an off ramp”, it seemed instinctively right though I have no clue what the alternative would be. I’d be interested in learning more. Because the independent thinking stuff is coming no matter what. I’d rather think we’re trying to set things up for success than throw some shit at the wall and see what sticks. That’s fine for some things but maybe not right for handling something like… um… everything?
Nah, you're spot on. Which is very much why LLMs are increasingly seen as a dead-end if AGI is what we're after. If LLMs can't reliably outperform humans on my tasks, then they're by the very definition of the word not AGI.
Also, while I make the point that LLMs are still useful to individuals despite their shortcomings, what I don't address is whether taking this off-ramp is justified when you take into account the financial and energy costs, negative societal impacts, and the fact that they're potentially a "distraction" if our sights are set on useful AGI.