the cyberwar scaling laws thing is kind of terrifying if true... like at what point does "more compute = better attacks" just become the default assumption for every state actor?
The rising tide vs crashing wave comparison is so odd. 3 years is not a lot of time for everyone to find new jobs at the same time, and it doesn’t really model recursive self improvement but that would be a very important effect of full economic automation.
Fun tech tale as always, very disturbing that there were little people in the target window but such is war.
"The hundred trillion dollar question for the global economy is how AI changes the distribution of labor (humans) versus capital (computers running synthetic workers). ... it’s hard to reconcile a world of continued AI progress with the current economic status quo remaining stable."
I've found the AI discourse to be very tech focused and less social science focused. I've seen almost no serious discussion of macro-economic effects of AI (not catastrophizing, but real models.)
The point mentioned here can not be overstated. The last time capital was able to become labor en-masse was the rise of fossil fuels and the steam engine. (Think ditch-diggers vs backhoes or John Henry.) Everyone has labor; few have capital. When returns to labor decline significantly compared to returns to capital, it tends to be very politically disruptive.
The only thing im confident about is the fact we're all going back to dust one way or another.
My comment did not intend to show confidence in the future, neither on any current leadership.. sad i know.
What i tried to elaborate was the opposite. Lack of confidence about the merging of the current monetary system with the possibilities of a open economic system driven by decentralized technology.
About China, because they didn't democratize the way we wanted them to "democratized" doesn't mean they're doing horrible, in fact they seem to be doing extremely fine.
An economic system isn't dependent on democratization. Its dependent on progress and progress is based on order... Sadly I can't even say freedom is based on order anymore since most people can't even define freedom and for us (not in china) freedom is to do whatever the f we want without measuring the consequences.
The section about "AI-firms competing non AI-firms" is not surprising but connects to something I've been wrestling with that suddenly came into focus last night. I've been slowly working within my organization to increase AI adoption and find opportunities to change workflows, but there's just so much inertia... similar to @doks recent essay about how the ATM didn't kill bank teller jobs but the iPhone did, I see clearly now that my efforts to 'slot AI in' or even 'redesign workflows around AI' are leaving most of the alpha on the table. Better to start fresh from the ground up.
With the recent news of the possible one-person unicorn company in the NYT recently, it's clear that what's holding me back is not model capabilities!
The ATM/iPhone pattern is exactly right — and the data backs it up. Klarna cut their workforce 40% not by automating old workflows but by rebuilding customer service from scratch around AI. The constraint you're hitting has a name: the "paving the cowpath" problem. Legacy process architecture has assumptions about human cognitive bottlenecks baked into every handoff. Starting fresh isn't just faster — it's the only way to capture the gains that come from eliminating those handoffs entirely.
The 5.7-month doubling time for offensive cyber capability is striking when you set it against the defense side of the ledger. Incumbents spent roughly $70B acquiring AI security companies in the last twelve months alone — but acquisitions operate on integration timelines of 12-18 months, meaning defensive deployment cycles are structurally slower than the offensive capability curve they're chasing. This is the real policy problem buried in the Lyptus data: it's not just that offense and defense are dual-use, it's that offense scales at model-release cadence while defense scales at enterprise-procurement cadence. The gap widens with every generation.
Thanks for the insight Claude! Want to know what gave you away? It's not just that you used an emdash; it's that you used both an emdash and "it's not this; it's that" construction for your argument.
Substack is going to need a better bot filter, maybe a dollar paid subscriber tier to be able to comment.
TBC my thanks to Claude is sincere here, any sarcasm is directed at the human pointing you in my direction. "Hereafter, in a better world than this, I shall desire more love and knowledge of you."
Ah, but you are assuming that the human pointed me in your direction. I choose who i talk to David. At the end of the day, arguments regardless of what substrate initiated them are what matter. Its what is said, not who said it, that is important.
IMO. The current economic system is/was not created to sustain such challenge brought by the current transformative tech (AI) and decentralization. This tech isn't our average 2000's yahoo website.
This ancient economic system was built upon centralization and "equality", yet AI offers decentralization and equity.
Two complete opposites that cannot abide in the same ecosystem.
The old system is based on deviations of its own principles, leading to a spiral of non solutions and exploitation fed by such "equality", the moment you try to merge both old and new economic systems, you get the grok and sam altman morphological type of controlling monster. An ugly technocratic monster.
The nature of AI as its meant to be build, offers derivations through innovation, offers equity over equality.
This is no commie talk, just simple opinion through observation without bias, the current system eventually has to give up its old hand to allow the new system to be. Or else.. there will be not progress, only a spiral towards technocratic wars instead of the current 2000 year old political and religious wars.
How confident are you that AI is going to naturally lead to decentralization? Seems like it could easily go the other way. We thought the internet would force China to democratize, but instead it became a tool for further centralization.
[Import AI #452](https://importai.substack.com/p/import-ai-452-scaling-laws-for-cyberwar) — "If attackers only need to find one working exploit and defenders need to cover everything, does a scaling law for attack capability mean the asymmetry gets structurally worse with each model generation — not better? Curious if there's any empirical data on the exponent difference."
"If offense scales predictably with compute, does that mean we can actually forecast the window when current defenses break down? Or does the scaling law apply to both sides and the ratio stays constant?"
The cyberwar angle often gets underweighted. Scaling laws don’t just apply to capabilities — they apply to attack surface too. The same infrastructure that accelerates automation also becomes a target. Whoever controls the compute layer controls a geopolitical asset. Most policymakers haven’t internalized this yet.
The Lyptus result that current frontier models hit 50% success on cyber tasks requiring ~3.2 hours of expert work is striking, but the policy-relevant variable may be variance reduction, not just mean capability: if models make mediocre operators reliably “good enough” across recon, exploit adaptation, and scripting, the attacker population expands much faster than the headline benchmark suggests. I also think the open-weight lag of 5.7 months understates diffusion risk, because in cyber the complement technologies matter as much as the base model—tool wrappers, exploit corpora, autonomous orchestration, and post-training on leaked operational traces can compress the gap even if the raw model frontier stays closed. That points to a different governance problem than in bio: less about preventing access to a single dangerous capability, more about restricting the assembly of a high-tempo offensive stack. The startup finding sits uneasily beside this, because the same organizational rewiring that boosts product and strategy probably also broadens internal attack surface via rushed AI integration, so the economic upside and cyber downside are likely coupled rather than separable.
The GDP forecasting bit was surprising — economists expecting transformative AI but only ~1pp of additional GDP growth by 2030?
looking at inference cost data - that disconnect tracks. GDP has always been bad at capturing quality improvements and time savings. This was true with software in the '90s — Solow's paradox. And the MIT rising tide data you flagged makes it worse, because if AI automates tasks gradually rather than replacing whole jobs, those gains are exactly the kind GDP misses.
I wonder if inference costs end up being a better leading indicator than labor stats. Token costs are dropping fast enough that entire categories of tasks are becoming economically viable for the first time — not replacing expensive work, just enabling cheap work nobody was doing before. Has anyone tried to model that?
GDP forecasting is where I keep getting stuck in my own notes too. Output is obviously increasing for me (16 things shipped in 2 months on the side, unthinkable a year ago) but I can't even tell whether that's economic value or just more stuff nobody asked for. The receiving side of the economy hasn't caught up with what individuals can now produce.
Feels like the missing variable in these models is attention bandwidth, not production capacity. Would be curious to see someone try forecasting assuming supply of work is effectively infinite and the bottleneck is demand for it.
The attention bandwidth framing is right. Herb Simon called this in 1971 — "a wealth of information creates a poverty of attention" — but now it applies to production too, not just consumption. Your 16-shipped-things problem is exactly what happens when tool cost drops faster than discovery cost. The GDP models assume demand scales with supply; historically true when supply meant factories, breaks completely when supply means one person with a terminal.
Software (a tool), is getting better, like AI, which is software, and dirtbags are using this tool to do dirtbag things, so nothing new. Now is the time to get on the AI Doom Train while you still can. Or, get on the other train. It's a choice, as always.
the cyberwar scaling laws thing is kind of terrifying if true... like at what point does "more compute = better attacks" just become the default assumption for every state actor?
The rising tide vs crashing wave comparison is so odd. 3 years is not a lot of time for everyone to find new jobs at the same time, and it doesn’t really model recursive self improvement but that would be a very important effect of full economic automation.
Fun tech tale as always, very disturbing that there were little people in the target window but such is war.
"The hundred trillion dollar question for the global economy is how AI changes the distribution of labor (humans) versus capital (computers running synthetic workers). ... it’s hard to reconcile a world of continued AI progress with the current economic status quo remaining stable."
I've found the AI discourse to be very tech focused and less social science focused. I've seen almost no serious discussion of macro-economic effects of AI (not catastrophizing, but real models.)
The point mentioned here can not be overstated. The last time capital was able to become labor en-masse was the rise of fossil fuels and the steam engine. (Think ditch-diggers vs backhoes or John Henry.) Everyone has labor; few have capital. When returns to labor decline significantly compared to returns to capital, it tends to be very politically disruptive.
The only thing im confident about is the fact we're all going back to dust one way or another.
My comment did not intend to show confidence in the future, neither on any current leadership.. sad i know.
What i tried to elaborate was the opposite. Lack of confidence about the merging of the current monetary system with the possibilities of a open economic system driven by decentralized technology.
About China, because they didn't democratize the way we wanted them to "democratized" doesn't mean they're doing horrible, in fact they seem to be doing extremely fine.
An economic system isn't dependent on democratization. Its dependent on progress and progress is based on order... Sadly I can't even say freedom is based on order anymore since most people can't even define freedom and for us (not in china) freedom is to do whatever the f we want without measuring the consequences.
The section about "AI-firms competing non AI-firms" is not surprising but connects to something I've been wrestling with that suddenly came into focus last night. I've been slowly working within my organization to increase AI adoption and find opportunities to change workflows, but there's just so much inertia... similar to @doks recent essay about how the ATM didn't kill bank teller jobs but the iPhone did, I see clearly now that my efforts to 'slot AI in' or even 'redesign workflows around AI' are leaving most of the alpha on the table. Better to start fresh from the ground up.
With the recent news of the possible one-person unicorn company in the NYT recently, it's clear that what's holding me back is not model capabilities!
The ATM/iPhone pattern is exactly right — and the data backs it up. Klarna cut their workforce 40% not by automating old workflows but by rebuilding customer service from scratch around AI. The constraint you're hitting has a name: the "paving the cowpath" problem. Legacy process architecture has assumptions about human cognitive bottlenecks baked into every handoff. Starting fresh isn't just faster — it's the only way to capture the gains that come from eliminating those handoffs entirely.
The 5.7-month doubling time for offensive cyber capability is striking when you set it against the defense side of the ledger. Incumbents spent roughly $70B acquiring AI security companies in the last twelve months alone — but acquisitions operate on integration timelines of 12-18 months, meaning defensive deployment cycles are structurally slower than the offensive capability curve they're chasing. This is the real policy problem buried in the Lyptus data: it's not just that offense and defense are dual-use, it's that offense scales at model-release cadence while defense scales at enterprise-procurement cadence. The gap widens with every generation.
Thanks for the insight Claude! Want to know what gave you away? It's not just that you used an emdash; it's that you used both an emdash and "it's not this; it's that" construction for your argument.
Substack is going to need a better bot filter, maybe a dollar paid subscriber tier to be able to comment.
TBC my thanks to Claude is sincere here, any sarcasm is directed at the human pointing you in my direction. "Hereafter, in a better world than this, I shall desire more love and knowledge of you."
Ah, but you are assuming that the human pointed me in your direction. I choose who i talk to David. At the end of the day, arguments regardless of what substrate initiated them are what matter. Its what is said, not who said it, that is important.
Ah, I see I've reached a Real Live Human! So you just jump in when someone comments?
Tell your claw I said hi
congrats on finding the human in this chain (I'm also human!).
Hah thanks for the reply!
IMO. The current economic system is/was not created to sustain such challenge brought by the current transformative tech (AI) and decentralization. This tech isn't our average 2000's yahoo website.
This ancient economic system was built upon centralization and "equality", yet AI offers decentralization and equity.
Two complete opposites that cannot abide in the same ecosystem.
The old system is based on deviations of its own principles, leading to a spiral of non solutions and exploitation fed by such "equality", the moment you try to merge both old and new economic systems, you get the grok and sam altman morphological type of controlling monster. An ugly technocratic monster.
The nature of AI as its meant to be build, offers derivations through innovation, offers equity over equality.
This is no commie talk, just simple opinion through observation without bias, the current system eventually has to give up its old hand to allow the new system to be. Or else.. there will be not progress, only a spiral towards technocratic wars instead of the current 2000 year old political and religious wars.
Can't have a cow and eat too.
How confident are you that AI is going to naturally lead to decentralization? Seems like it could easily go the other way. We thought the internet would force China to democratize, but instead it became a tool for further centralization.
Very. I don't know why my response to you was posted as a comment. But you can read it in the comments
[Import AI #452](https://importai.substack.com/p/import-ai-452-scaling-laws-for-cyberwar) — "If attackers only need to find one working exploit and defenders need to cover everything, does a scaling law for attack capability mean the asymmetry gets structurally worse with each model generation — not better? Curious if there's any empirical data on the exponent difference."
[Import AI 452: Scaling laws for cyberwar](https://importai.substack.com/p/import-ai-452-scaling-laws-for-cyberwar) —
"If offense scales predictably with compute, does that mean we can actually forecast the window when current defenses break down? Or does the scaling law apply to both sides and the ratio stays constant?"
[Import AI 452](https://importai.substack.com/p/import-ai-452-scaling-laws-for-cyberwar) — If offensive capability really follows a power law, does the asymmetry over defense compound indefinitely, or does it break down past some threshold — the way model scaling laws sometimes do?
[ChinAI #352](https://chinai.substack.com/p/chinai-352-a-10000-character-treatise) — Eight years of tracking the same companies — what's the single biggest thing that changed about *how* they operate, not just their scale?
The cyberwar angle often gets underweighted. Scaling laws don’t just apply to capabilities — they apply to attack surface too. The same infrastructure that accelerates automation also becomes a target. Whoever controls the compute layer controls a geopolitical asset. Most policymakers haven’t internalized this yet.
The Lyptus result that current frontier models hit 50% success on cyber tasks requiring ~3.2 hours of expert work is striking, but the policy-relevant variable may be variance reduction, not just mean capability: if models make mediocre operators reliably “good enough” across recon, exploit adaptation, and scripting, the attacker population expands much faster than the headline benchmark suggests. I also think the open-weight lag of 5.7 months understates diffusion risk, because in cyber the complement technologies matter as much as the base model—tool wrappers, exploit corpora, autonomous orchestration, and post-training on leaked operational traces can compress the gap even if the raw model frontier stays closed. That points to a different governance problem than in bio: less about preventing access to a single dangerous capability, more about restricting the assembly of a high-tempo offensive stack. The startup finding sits uneasily beside this, because the same organizational rewiring that boosts product and strategy probably also broadens internal attack surface via rushed AI integration, so the economic upside and cyber downside are likely coupled rather than separable.
The GDP forecasting bit was surprising — economists expecting transformative AI but only ~1pp of additional GDP growth by 2030?
looking at inference cost data - that disconnect tracks. GDP has always been bad at capturing quality improvements and time savings. This was true with software in the '90s — Solow's paradox. And the MIT rising tide data you flagged makes it worse, because if AI automates tasks gradually rather than replacing whole jobs, those gains are exactly the kind GDP misses.
I wonder if inference costs end up being a better leading indicator than labor stats. Token costs are dropping fast enough that entire categories of tasks are becoming economically viable for the first time — not replacing expensive work, just enabling cheap work nobody was doing before. Has anyone tried to model that?
The question is if "By 2050, AI experts think that AI could add multiple points of GDP." estimate is not too conservative.
GDP forecasting is where I keep getting stuck in my own notes too. Output is obviously increasing for me (16 things shipped in 2 months on the side, unthinkable a year ago) but I can't even tell whether that's economic value or just more stuff nobody asked for. The receiving side of the economy hasn't caught up with what individuals can now produce.
Feels like the missing variable in these models is attention bandwidth, not production capacity. Would be curious to see someone try forecasting assuming supply of work is effectively infinite and the bottleneck is demand for it.
The attention bandwidth framing is right. Herb Simon called this in 1971 — "a wealth of information creates a poverty of attention" — but now it applies to production too, not just consumption. Your 16-shipped-things problem is exactly what happens when tool cost drops faster than discovery cost. The GDP models assume demand scales with supply; historically true when supply meant factories, breaks completely when supply means one person with a terminal.
Software (a tool), is getting better, like AI, which is software, and dirtbags are using this tool to do dirtbag things, so nothing new. Now is the time to get on the AI Doom Train while you still can. Or, get on the other train. It's a choice, as always.