Alberto, your point about the token rich and token poor divide is exactly right. Having been in the tech and startup space for a few decades now, I have seen every flavor of cash burn imaginable. We used to lose sleep over runaway cloud hosting bills or bloated early headcount. Now I am actively warning the founders I mentor that a sloppy autonomous agent loop will vaporize their entire seed funding over a holiday weekend. It is completely wild how fast you can drain a bank account with test time compute if you are not paying very close attention.
Your advice to stay in the valleys with Fable or Opus is precisely what I tell my teams. The sheer obsession with having the smartest model for basic, everyday tasks is a trap. Let the mega labs burn their billions up in the peaks. The rest of us need to focus on actual business logic and basic fiscal survival. Really sharp analysis here.
As someone who is just learning and only thinks of themselves as a casual user of AI, reading about the power that these new models are achieving and I agree with you that I will never use anything like Mythos level models. I like many other people have a cellphone that I only use maybe 10% of the applications that are on it. Sometimes I question myself,”why the hell do I pay for all this stuff I don’t use” when I buy these phones. Along the same lines, “why would I buy / pay for AGI (as Sam Altman wants to offer) when I probably will use a small percentage of the power of these powerful AGI’s like Mythos. Or maybe they will have tiered services (of course at tiered prices) where it is less expensive and less powerful like agi vs AGI. I see their use for big questions, cure cancer, development of miracle drugs, etc. But for the general population I think super model AGI or likely even average every day agi will be just like my cellphone where it won’t get used to its intended potential but will be one of those things that society has dictated is necessary for everyone to have.
I am not a computer or tech person but I like to read articles about it to stay informed. I make this comment as an average Joe off the street and genuinely questioning what is the end game of all this race for AGI in regard to the general masses.
Alberto, your point five is the one I am living right now, and I want to offer myself as a counterexample to point two.
I am a public high school chemistry and physics teacher. Normal salary, the exact reader you wrote this for. And I spent the last two days using the free window to build something no edtech company would ever fund: an interactive chemistry textbook where the game world literally is the phase diagram and the only way forward is understanding the chemistry. Eight working builds in two days. A boss fight you cannot win while ignorant. Things I sketched for years and could never afford to make.
So I think the divide is not quite peaks and valleys by profession or salary. It is whether you have something big you always wanted to build and never could. Mollick made the software researchers needed for years that was never profitable to create. There are a lot of teachers, nurses, and tradespeople sitting on that same category of unbuilt things. The window does not care what your day job is. It cares whether you show up with a blueprint.
Your point eight I will co-sign with both hands. The machine builds, the human writes. That division of labor seems to be holding.
... writing isn’t and will never be the main goal for AI labs, but at this point, it’s pretty clear they just can’t do anything about it" I am always wondering if it is really something they can't do, or just not a priority. As far as I understand the default to words that "can be relevant in many contexts" (=meaningless clichés) is intentional, but I am really not an expert, thought just ask. (note: resisting the urge to say things like "this is the most important sentence in the whole post and it gets buried so deep under simple letters that you never know what hit you" et al.)
"Recursive self-improvement . . . works like getting closer to a black hole. The closer you are, the stronger the pull and thus the faster you go, and thus the closer you are and thus the stronger the pull, and so on until you reach the event horizon (the singularity)." - Exponential improvement is just one possibility. It's the one that makes a great scifi story. However, another possibility is logarithmic improvement. And another is exponential degradation (self-improvement gone wrong). Personally, I think exponential improvement is the least likely of the three -- unless we are talking about an exponent less than 1 ;-)
I agree with you regarding writing. It sucks. But what it’s good for is assessing written work , but only if you are specific about the assessment process. So, for example, the mechanicals, including punctuation, logic gate, word choice and so forth. And, more than anything else, prohibit it for modifying any of your prose or using personal context to influence an assessment. In other words, what it knows about you from all your interactions, which could be used to color an assessment. Finally, it’s important to have hard rules about holding firm on disagreements where either you or the model is convinced that it is correct, or conceding the point of moving on when it’s not. The way that I describe this within the protocol is “collision without carnage. “This came from an particularly contentious disagreement, which ended with me telling Claude to stop being a prick about it.
Enjoyed this! The human part is still the whole thing. We're token rich compared to most indie people but we're definitely valley dwellers...can't justify burning thousands on test-time compute when the work that actually moves the needle is still the thinking and the editing a person does.
It really does appear that this time no one in China can bypass the restrictions. People used to rely on transfer stations and other methods to get around, but now, even if it's technically possible, it just doesn't seem worth the effort.
Like you pointed out, Fable 5 isn't gonna be that much of an improvement compared to other mainstream advanced models for people who only use AI for everyday tasks. And as AI access becomes increasingly geopolitical and honestly just annoying, I wonder if the safeguards will ever become so overboard that everyday users get turned off by it and stop using.
Alberto, your point about the token rich and token poor divide is exactly right. Having been in the tech and startup space for a few decades now, I have seen every flavor of cash burn imaginable. We used to lose sleep over runaway cloud hosting bills or bloated early headcount. Now I am actively warning the founders I mentor that a sloppy autonomous agent loop will vaporize their entire seed funding over a holiday weekend. It is completely wild how fast you can drain a bank account with test time compute if you are not paying very close attention.
Your advice to stay in the valleys with Fable or Opus is precisely what I tell my teams. The sheer obsession with having the smartest model for basic, everyday tasks is a trap. Let the mega labs burn their billions up in the peaks. The rest of us need to focus on actual business logic and basic fiscal survival. Really sharp analysis here.
As someone who is just learning and only thinks of themselves as a casual user of AI, reading about the power that these new models are achieving and I agree with you that I will never use anything like Mythos level models. I like many other people have a cellphone that I only use maybe 10% of the applications that are on it. Sometimes I question myself,”why the hell do I pay for all this stuff I don’t use” when I buy these phones. Along the same lines, “why would I buy / pay for AGI (as Sam Altman wants to offer) when I probably will use a small percentage of the power of these powerful AGI’s like Mythos. Or maybe they will have tiered services (of course at tiered prices) where it is less expensive and less powerful like agi vs AGI. I see their use for big questions, cure cancer, development of miracle drugs, etc. But for the general population I think super model AGI or likely even average every day agi will be just like my cellphone where it won’t get used to its intended potential but will be one of those things that society has dictated is necessary for everyone to have.
I am not a computer or tech person but I like to read articles about it to stay informed. I make this comment as an average Joe off the street and genuinely questioning what is the end game of all this race for AGI in regard to the general masses.
I am a computer tech person since the 70s and I agreed whole heartedly with you. I wish more people, like you, spoke up.
Alberto, your point five is the one I am living right now, and I want to offer myself as a counterexample to point two.
I am a public high school chemistry and physics teacher. Normal salary, the exact reader you wrote this for. And I spent the last two days using the free window to build something no edtech company would ever fund: an interactive chemistry textbook where the game world literally is the phase diagram and the only way forward is understanding the chemistry. Eight working builds in two days. A boss fight you cannot win while ignorant. Things I sketched for years and could never afford to make.
So I think the divide is not quite peaks and valleys by profession or salary. It is whether you have something big you always wanted to build and never could. Mollick made the software researchers needed for years that was never profitable to create. There are a lot of teachers, nurses, and tradespeople sitting on that same category of unbuilt things. The window does not care what your day job is. It cares whether you show up with a blueprint.
Your point eight I will co-sign with both hands. The machine builds, the human writes. That division of labor seems to be holding.
... writing isn’t and will never be the main goal for AI labs, but at this point, it’s pretty clear they just can’t do anything about it" I am always wondering if it is really something they can't do, or just not a priority. As far as I understand the default to words that "can be relevant in many contexts" (=meaningless clichés) is intentional, but I am really not an expert, thought just ask. (note: resisting the urge to say things like "this is the most important sentence in the whole post and it gets buried so deep under simple letters that you never know what hit you" et al.)
"Recursive self-improvement . . . works like getting closer to a black hole. The closer you are, the stronger the pull and thus the faster you go, and thus the closer you are and thus the stronger the pull, and so on until you reach the event horizon (the singularity)." - Exponential improvement is just one possibility. It's the one that makes a great scifi story. However, another possibility is logarithmic improvement. And another is exponential degradation (self-improvement gone wrong). Personally, I think exponential improvement is the least likely of the three -- unless we are talking about an exponent less than 1 ;-)
I agree with you regarding writing. It sucks. But what it’s good for is assessing written work , but only if you are specific about the assessment process. So, for example, the mechanicals, including punctuation, logic gate, word choice and so forth. And, more than anything else, prohibit it for modifying any of your prose or using personal context to influence an assessment. In other words, what it knows about you from all your interactions, which could be used to color an assessment. Finally, it’s important to have hard rules about holding firm on disagreements where either you or the model is convinced that it is correct, or conceding the point of moving on when it’s not. The way that I describe this within the protocol is “collision without carnage. “This came from an particularly contentious disagreement, which ended with me telling Claude to stop being a prick about it.
Enjoyed this! The human part is still the whole thing. We're token rich compared to most indie people but we're definitely valley dwellers...can't justify burning thousands on test-time compute when the work that actually moves the needle is still the thinking and the editing a person does.
It really does appear that this time no one in China can bypass the restrictions. People used to rely on transfer stations and other methods to get around, but now, even if it's technically possible, it just doesn't seem worth the effort.
Like you pointed out, Fable 5 isn't gonna be that much of an improvement compared to other mainstream advanced models for people who only use AI for everyday tasks. And as AI access becomes increasingly geopolitical and honestly just annoying, I wonder if the safeguards will ever become so overboard that everyday users get turned off by it and stop using.
Great post. The jagged edge of intelligence got real visible.