“Every time the hype cycle repeats , capex swarms in , which becomes a self fullfing prophecy for what we build next . This is why I feel it is inevitable “
Someone read Nick land and internalized this thesis about techno capital
This was an insightful podcast. Very informative. My only "but" is how poorly sound-edited it was. Did an Ai edit this without crossfades? It only needed crossfades! I had to stop listening and came here instead to read it.
The routing discussion suggests a useful way to connect model evaluation with the gross-margin problem: measure cost per successfully completed workflow, not cost per token. A cheaper model can become the expensive option once retries, escalation and human review are included. It would be interesting to compare a routed system with a frontier-only baseline on the same customer tasks, including those downstream costs. Has Ant found categories where routing improves margins without increasing the customer’s checking burden, and others where the apparent saving disappears?
Richard's point about a customer trying the software before asking for a proof of concept is lovely. By the time the sales conversation starts, there's something concrete to discuss: what worked, where it struggled and what the customer needs next. That makes the commercial value of openness much easier to picture.
Salute for giving us a well-steered conversation. It was extremely insightful and held my attention to the last word. I was also confounded by the same questions as raised by @Steve Sewall on occupational category graph. Secondly, the cross currents point about guardrail calibration.
While there are many concerns one can raise, I fully appreciate the key point Richard has made. The game has begun and our ships have sailed. We have to find our course and fable lands. The current play out is mimicking the raise of internet days. It was simply a matter of arrival of products akin to Mac, Amazon, Uber, WhatsApp genere. Thank you Irene and my salutes to Richard for his generosity.
The baseball analogy separates a question that often gets bundled together: technical capability may have reset to home plate, while commercialization is still waiting at first. Richard’s claim that gross margins—not capex—are the real open-source bottleneck is testable. We’d watch serving cost per useful task, rather than benchmark rank, as the business signal.
That Hugging Face detail is hard to shake. Frontier models refused to help parse the logs of an attack carried out by frontier agents, so the cleanup ran on a Chinese open-weight model from Z.ai sitting on a local machine. Richard Bian frames open source as a product strategy, and incidents like this show how that strategy gains ground quietly, by being the tool that simply works when American guardrails trip. Ant releasing the 560B-parameter Ling 3.1 Flash as open weights two days ago points the same direction. Curious whether US labs and policymakers read this as a guardrail calibration problem or a market-share problem, because security teams in the West might keep reaching for whatever model answers, wherever it was trained.
Wonderful post. For me the last two thirds were a baptism because I'm an AI outsider peeking inside. But the first third really got to me. Clarified a lot! Reminded me of Rob Urie's Zen Economics and Peter D'Errico's latest Dao post. (He's on Substack.)
But one question. May I set it up. These days I look everywhere for ANY AI uses dedicated to strengthening communication between citizens and government. For my money, such uses are the ONLY antidote to renegade frontier AI. More than that, in the equation between human and machine, aren't such uses indeed the OBVIOUS way to reverse the downward "Spaghetti Monster" spiral behind Richard's pessimism? These uses would necessarily be trustworthy, whether managed by humans or by an algorithm capable of building an informed consensus on the issues of the day. The output of these AI uses would be to give all citizens of a democracy or autocracy (yes, both!) an INFORMED VOICE in reaching the government (and business, by extension) decisions that affect their lives as well well-being of their nation. (And maybe someday the world). Why not?
But that's not my real question. I look everywhere but find no sign of such uses of AL. This strikes mas as a huge conceptual gap. Take, for instance, Anthropic's Occupational Category graphic cited by Richard. It lists 22 Occupations. Including Grounds Maintenance. But not Politics and Government. And no only is P&G MISSING, it's arguably EXCLUDED from this ostensibly comprehensive list. So here's my real question. Has anyone, or is anyone, using AI to connect citizens and government - ideally at all three levels, local, state and national - where citizens are members of government?
It wasn’t really the “age of discovery” though was it? It was the age of conquering, theft and genocide. Of claiming what wasn’t yours as yours by divine right. I see parallels to AI development. Where is democracy, transparency, and accountability? Did regular citizens ask for AI? I see them asking for lower rent and healthcare. Meanwhile datacenters are opposed in most places. This is the great divergence. It’s not about AGI, it’s about class struggle and democracy. It’s about unaccountable billionaires investing collective resources for private gain, no matter the consequences.
the distribution side is what makes the santé bet interesting from here: afu also lives inside alipay, which already does hospital booking and holds the 医保 code people scan at the pharmacy. a health model there gets used at the counter, not in a chat window. did richard say how many people actually use afu?
“Every time the hype cycle repeats , capex swarms in , which becomes a self fullfing prophecy for what we build next . This is why I feel it is inevitable “
Someone read Nick land and internalized this thesis about techno capital
This was an insightful podcast. Very informative. My only "but" is how poorly sound-edited it was. Did an Ai edit this without crossfades? It only needed crossfades! I had to stop listening and came here instead to read it.
But the content was top-notch. Love the access!
The routing discussion suggests a useful way to connect model evaluation with the gross-margin problem: measure cost per successfully completed workflow, not cost per token. A cheaper model can become the expensive option once retries, escalation and human review are included. It would be interesting to compare a routed system with a frontier-only baseline on the same customer tasks, including those downstream costs. Has Ant found categories where routing improves margins without increasing the customer’s checking burden, and others where the apparent saving disappears?
Richard's point about a customer trying the software before asking for a proof of concept is lovely. By the time the sales conversation starts, there's something concrete to discuss: what worked, where it struggled and what the customer needs next. That makes the commercial value of openness much easier to picture.
Salute for giving us a well-steered conversation. It was extremely insightful and held my attention to the last word. I was also confounded by the same questions as raised by @Steve Sewall on occupational category graph. Secondly, the cross currents point about guardrail calibration.
While there are many concerns one can raise, I fully appreciate the key point Richard has made. The game has begun and our ships have sailed. We have to find our course and fable lands. The current play out is mimicking the raise of internet days. It was simply a matter of arrival of products akin to Mac, Amazon, Uber, WhatsApp genere. Thank you Irene and my salutes to Richard for his generosity.
The baseball analogy separates a question that often gets bundled together: technical capability may have reset to home plate, while commercialization is still waiting at first. Richard’s claim that gross margins—not capex—are the real open-source bottleneck is testable. We’d watch serving cost per useful task, rather than benchmark rank, as the business signal.
That Hugging Face detail is hard to shake. Frontier models refused to help parse the logs of an attack carried out by frontier agents, so the cleanup ran on a Chinese open-weight model from Z.ai sitting on a local machine. Richard Bian frames open source as a product strategy, and incidents like this show how that strategy gains ground quietly, by being the tool that simply works when American guardrails trip. Ant releasing the 560B-parameter Ling 3.1 Flash as open weights two days ago points the same direction. Curious whether US labs and policymakers read this as a guardrail calibration problem or a market-share problem, because security teams in the West might keep reaching for whatever model answers, wherever it was trained.
Wonderful post. For me the last two thirds were a baptism because I'm an AI outsider peeking inside. But the first third really got to me. Clarified a lot! Reminded me of Rob Urie's Zen Economics and Peter D'Errico's latest Dao post. (He's on Substack.)
But one question. May I set it up. These days I look everywhere for ANY AI uses dedicated to strengthening communication between citizens and government. For my money, such uses are the ONLY antidote to renegade frontier AI. More than that, in the equation between human and machine, aren't such uses indeed the OBVIOUS way to reverse the downward "Spaghetti Monster" spiral behind Richard's pessimism? These uses would necessarily be trustworthy, whether managed by humans or by an algorithm capable of building an informed consensus on the issues of the day. The output of these AI uses would be to give all citizens of a democracy or autocracy (yes, both!) an INFORMED VOICE in reaching the government (and business, by extension) decisions that affect their lives as well well-being of their nation. (And maybe someday the world). Why not?
But that's not my real question. I look everywhere but find no sign of such uses of AL. This strikes mas as a huge conceptual gap. Take, for instance, Anthropic's Occupational Category graphic cited by Richard. It lists 22 Occupations. Including Grounds Maintenance. But not Politics and Government. And no only is P&G MISSING, it's arguably EXCLUDED from this ostensibly comprehensive list. So here's my real question. Has anyone, or is anyone, using AI to connect citizens and government - ideally at all three levels, local, state and national - where citizens are members of government?
It wasn’t really the “age of discovery” though was it? It was the age of conquering, theft and genocide. Of claiming what wasn’t yours as yours by divine right. I see parallels to AI development. Where is democracy, transparency, and accountability? Did regular citizens ask for AI? I see them asking for lower rent and healthcare. Meanwhile datacenters are opposed in most places. This is the great divergence. It’s not about AGI, it’s about class struggle and democracy. It’s about unaccountable billionaires investing collective resources for private gain, no matter the consequences.
the distribution side is what makes the santé bet interesting from here: afu also lives inside alipay, which already does hospital booking and holds the 医保 code people scan at the pharmacy. a health model there gets used at the counter, not in a chat window. did richard say how many people actually use afu?