> Could AI help you build a competitor to beat TikTok in just a few minutes? Maybe in the next couple of years, according to Eric Schmidt, Google’s former CEO.
Seems like there's an obvious flaw in this premise since it could also help a billion other people do the same thing so in that case... no, pretty much none of you are "beating TikTok"?
Same way that creating a N64-level game in 2024 doesn't get you the same press or audience as it did back then.
I think we're already at the point where programming skill isn't the barrier between making "netflix history-based tinder" or whatever.
So are we really just talking about making a shit-ton more spam of all sorts? App spam, super-SEO spam, fake social-media post spam, etc.
Is making it easy to make race-to-the-bottom spam plays gonna give you ROIs on hundreds of billions or trillions for OpenAI or whoever else?
On one hand, MS made a lot of money selling Word despite it being harder to make money as a writer now that the barrier to entry is so low. On the other hand, the capital costs to create Word were far lower, and the market-barrier to entry for Word replacements back in the 80s and 90s was higher.
You’ve hit the nail on the head here. It feels like everyone who keeps saying AI is going to make it easy to churn out billion dollar businesses is only one step short of selling their latest guru course.
It's becoming more and more obvious the number one product this industry sells is shares in companies. Everything is about getting the price of those shares as high as possible for the anxious billionaire/institutional investment class whenever they go on the market.
The hero worshiping of people who’ve made a lot of money is really boring. What impact does this “prophecy” have? None. It’s just bullshit. Just like the work from home causing Googles issues, it’s just vibes bullshitting.
I work at one of the top companies in this field. Schmidt doesn’t know anything about this. I do think we’ll make significant efficiency gains (by which I mean inference, weight, and sample efficiency, all three) but I do not see any technology in the foreseeable future that’d mean a qualitative, or even significantly quantitative change in capability. So in the next few years at least we’re going to be in the flat portion of the sigmoid with only very gradual progress in what these models can do. Which is fine, because they already can do a lot, and FOSS needs time to catch up.
In a similar position as you, I fully agree. I see no step wise function like gains in the foreseeable future, merely the incremental improvements you mention. Though I'm not surprised Schmidt would put these words in a sentence and express them publicly.
Let's stipulate that Schmidt himself isn't hands-on, so you're correct in the sense that he doesn't have first-hand knowledge.
Can we also agree that he's probably not a complete idiot, he's probably not intentionally trying to deceive, and he has a reputation that he'd prefer to protect and enhance. More likely he's repeating what knowledgeable people in the upper echelons of the industry have told him.
If that's true, then shouldn't we put some weight in what he says? Even if it's not first-hand knowledge, even if he's only repeating what knowledgeable people told him, it's not likely to be all bullshit is it?
People in the upper echelons of all this are incentivized with billions of dollars to tell you all sorts of bullshit that’s just not going to happen. “AGI” is but one example. What weight you put on all that in these circumstances is up to you, but given the context I have I think the appropriate value is either zero or even slightly negative.
Perhaps, though I'm not convinced that incentives really work like that.
For every Elon Musk who makes grandiose promises and then either fails to deliver at all, or eventually delivers 9 years after he promised, there must be multiple senior execs who would be sufficiently embarrassed to make a bold prediction about the near future that utterly fails to materialize.
Certainly these people have already seen the capabilities of what will be released in the near future after all the safety boxes have been checked. Certainly they know more about the likely near future than me or 99.999% of everyone else.
So when I hear a bold prediction about the near future that's essentially a synthesis of opinions of this handful of insiders who are in the best possible position to know, why would I discount it by 100% simply due to hand-wavy "incentives"?
I _know_ what will be released in the next year, year and a half. It’s basically just improvements to models, multimodality, multilinguality and various optimizations. Absolutely nobody at my current FANG employer is currently working on anything even remotely resembling “AGI” or indeed anything qualitatively different from the technologies we already have.
Not really, no. Much of the slack has already been taken up in GPT 3.5 era, and things haven’t dramatically improved since GPT4. 4o basically alludes to what I said above: it’s smaller and faster than 4, but the level of capability is very similar, the next big thing in it is multimodality, which is relatively straightforward to add.
It just makes sense to pay attention to cost and efficiency now. It’s incredibly expensive today to both do research in this field and deploy the results of that research. Even FANGs are severely compute constrained to the point where when Google was training the bigger variants of Gemini I couldn’t even get a handful of TPUs scheduled to do my work (I’m not at Google now). That, quite obviously, is untenable.
But even if NVIDIA/TSMC somehow tripled their production and everyone got as much compute as they can use, there’s no serious research going on for how to do cognition. The closest we got to that is neural network assisted Monte Carlo a-la AlphaZero and the various other approaches to search. But all of them are very narrow. Nobody has even a foggiest clue on how to bring anything cognition-like to the general domain.
And don’t get me wrong, I’m a bona fide AI enthusiast, I’ve been in this field for over a decade. It’s just that all these promises of singularity “real soon now” are nothing more than clickbait.
> For every Elon Musk who makes grandiose promises and then either fails to deliver at all, or eventually delivers 9 years after he promised, there must be multiple senior execs who would be sufficiently embarrassed to make a bold prediction about the near future that utterly fails to materialize.
I mean, to be clear, Schmidt hasn’t been a senior exec in the normal sense for 13 years; that’s when he stood down as Google CEO.
But also, CEOs predict complete nonsense _all the time_. That’s why you’re not browsing Google+ through your VR headset while heading home from your job at the vertical farm in your self-driving car, for a night of watching 3D TV and drinking Juicero.
> If that's true, then shouldn't we put some weight in what he says? Even if it's not first-hand knowledge, even if he's only repeating what knowledgeable people told him, it's not likely to be all bullshit is it?
You probably _can_ take it at face value, and billions of dollars changes hands daily on this sort of conjecture. But is it the smart thing to do?
At a time when we have more data, instrumentation, and analytical models of everything (including market trends for tech companies) I would argue no. What one guy with unclear motives says doesn’t supersede whatever we could otherwise extract from evidence.
It is possible to not be an idiot and be completely hilariously wrong/misinformed at the same time.
Also, keep in mind that earlier in this sentence he said something like "want to make the next google without all of those arrogant programmers?" (The same arrogant programmers that made him hilariously wealthy, of course.) Master class in saying the quiet part out loud, IMO.
> he's probably not intentionally trying to deceive
I cannot agree with this. I'm not saying he is intentionally trying to deceive, but there's no particular reason to assume he isn't.
> he has a reputation that he'd prefer to protect and enhance
Nor this, in the sense that you mean. The reputation he wants to protect may be how he handles stock price rather than how accurate his PR representations are, for instance.
> it's not likely to be all bullshit is it?
It could very easily be 99% bullshit, yes. It's not like it's a rare phenomenon for CEOs to be deceptive in their public speech, nor is it particularly rare for CEOs to be completely mistaken about things like this. Add in that what he's saying is incredibly speculative, and it all seems pretty shaky. It reads like an advertisement.
All I'm saying is that I don't think the fact that he says something is enough to hold a lot of faith that what he said is accurate.
> Say to your LLM, ‘Make me a copy of TikTok. Steal all the users, steal all the music, put my preferences in it, produce this program in the next 30 seconds, release it, and in one hour, if it's not viral, do something different along the same lines.’
I think it all boils down to value. Sure, in a few years time, that might be possible. But what is the value in doing that? How am I bettering my own or anyone else's life by doing that? Say what you will about social media, but it at least has theoretical value: the spread of ideas. What does this level of AI have?
23 comments
[ 3.6 ms ] story [ 65.3 ms ] threadSeems like there's an obvious flaw in this premise since it could also help a billion other people do the same thing so in that case... no, pretty much none of you are "beating TikTok"?
Same way that creating a N64-level game in 2024 doesn't get you the same press or audience as it did back then.
I think we're already at the point where programming skill isn't the barrier between making "netflix history-based tinder" or whatever.
So are we really just talking about making a shit-ton more spam of all sorts? App spam, super-SEO spam, fake social-media post spam, etc.
Is making it easy to make race-to-the-bottom spam plays gonna give you ROIs on hundreds of billions or trillions for OpenAI or whoever else?
On one hand, MS made a lot of money selling Word despite it being harder to make money as a writer now that the barrier to entry is so low. On the other hand, the capital costs to create Word were far lower, and the market-barrier to entry for Word replacements back in the 80s and 90s was higher.
1. Type a prompt into AI
2. Copy/paste the answer
3. Exit for $3b
Never mind that pesky business stuff.
Can we also agree that he's probably not a complete idiot, he's probably not intentionally trying to deceive, and he has a reputation that he'd prefer to protect and enhance. More likely he's repeating what knowledgeable people in the upper echelons of the industry have told him.
If that's true, then shouldn't we put some weight in what he says? Even if it's not first-hand knowledge, even if he's only repeating what knowledgeable people told him, it's not likely to be all bullshit is it?
For every Elon Musk who makes grandiose promises and then either fails to deliver at all, or eventually delivers 9 years after he promised, there must be multiple senior execs who would be sufficiently embarrassed to make a bold prediction about the near future that utterly fails to materialize.
Certainly these people have already seen the capabilities of what will be released in the near future after all the safety boxes have been checked. Certainly they know more about the likely near future than me or 99.999% of everyone else.
So when I hear a bold prediction about the near future that's essentially a synthesis of opinions of this handful of insiders who are in the best possible position to know, why would I discount it by 100% simply due to hand-wavy "incentives"?
Even though I'm expecting to be shocked by what use cases those improvements will unlock, I expect it will be even more shocking than I expect.
It just makes sense to pay attention to cost and efficiency now. It’s incredibly expensive today to both do research in this field and deploy the results of that research. Even FANGs are severely compute constrained to the point where when Google was training the bigger variants of Gemini I couldn’t even get a handful of TPUs scheduled to do my work (I’m not at Google now). That, quite obviously, is untenable.
But even if NVIDIA/TSMC somehow tripled their production and everyone got as much compute as they can use, there’s no serious research going on for how to do cognition. The closest we got to that is neural network assisted Monte Carlo a-la AlphaZero and the various other approaches to search. But all of them are very narrow. Nobody has even a foggiest clue on how to bring anything cognition-like to the general domain.
And don’t get me wrong, I’m a bona fide AI enthusiast, I’ve been in this field for over a decade. It’s just that all these promises of singularity “real soon now” are nothing more than clickbait.
I mean, to be clear, Schmidt hasn’t been a senior exec in the normal sense for 13 years; that’s when he stood down as Google CEO.
But also, CEOs predict complete nonsense _all the time_. That’s why you’re not browsing Google+ through your VR headset while heading home from your job at the vertical farm in your self-driving car, for a night of watching 3D TV and drinking Juicero.
You probably _can_ take it at face value, and billions of dollars changes hands daily on this sort of conjecture. But is it the smart thing to do?
At a time when we have more data, instrumentation, and analytical models of everything (including market trends for tech companies) I would argue no. What one guy with unclear motives says doesn’t supersede whatever we could otherwise extract from evidence.
Also, keep in mind that earlier in this sentence he said something like "want to make the next google without all of those arrogant programmers?" (The same arrogant programmers that made him hilariously wealthy, of course.) Master class in saying the quiet part out loud, IMO.
I cannot agree with this. I'm not saying he is intentionally trying to deceive, but there's no particular reason to assume he isn't.
> he has a reputation that he'd prefer to protect and enhance
Nor this, in the sense that you mean. The reputation he wants to protect may be how he handles stock price rather than how accurate his PR representations are, for instance.
> it's not likely to be all bullshit is it?
It could very easily be 99% bullshit, yes. It's not like it's a rare phenomenon for CEOs to be deceptive in their public speech, nor is it particularly rare for CEOs to be completely mistaken about things like this. Add in that what he's saying is incredibly speculative, and it all seems pretty shaky. It reads like an advertisement.
All I'm saying is that I don't think the fact that he says something is enough to hold a lot of faith that what he said is accurate.
I think it all boils down to value. Sure, in a few years time, that might be possible. But what is the value in doing that? How am I bettering my own or anyone else's life by doing that? Say what you will about social media, but it at least has theoretical value: the spread of ideas. What does this level of AI have?
“The next two years of AI will shock you.” (Direct quote: “I invested lots of money” subtext: me and my VC buddies need an exit.)