Alfonso, great piece. I published an essay yesterday that arrives at a similar conclusion from a completely different direction. You're arguing from software architecture: intelligence commoditizes, context is the product, the value capture layer sits on top of frontier labs. I'm arguing from economic theory: the leading economic model of AI and centralization (Brynjolfsson and Hitzig) assumes that creating context is lossless and costless, and when you add the actual cost of context engineering (i.e., treat context engineering as an investment in a knowledge asset), the model's predictions change.
Your line "what really matters is to provide this intelligence with the optimal context" is exactly what I'm trying to formalize as an economic concept: cCE, the cost of context engineering. Your OpenClaw example, 400k lines of code reduced to 4k lines plus skills, is a perfect illustration of what I call Channel 2 in my essay: the value isn't in the processing power, it's in the optimal context bundle at inference time, and humans have a necessary role in guiding the context.
Would be curious whether the economic framing resonates with what you're seeing from the architecture side.
I really appreciate your kind words. I just read your essay and I really enjoyed it. Coincidentally, I just finished reading "The Last Economy" by Emad Mostaque (co-founder of Stability AI). It is a quick one to read, and it explores the topic of the age of intelligent economics in a broader and more high-level way than your essay. I think you are going to enjoy some of the ideas there, as I was reading I felt it could be a good complement to your work.
Regarding your essay, I love how you define channel 1 and 2. The only nit is that usually I understand context engineering (maybe due to my engineering background) as the way knowledge is codified and fed to an LLM for it to generate and informed output (channel 1). Context is definitely an over-loaded word at this point, but to me I think the challenge for this intelligent agents (and what I understand by "context" in my presentation) is the ability to provide them with the right knowledge base for a task, and navigate it efficiently to solve the task (independently of where it is stored). Those who manage to retrieve the most information-dense and relevant information for a task will get better results out of their agents.
Anyway, as I said definitely a nit that I hope helps, but I think that you capture the core of the concept perfectly through your channels distinction
I was actually writing my next post about something related to economics on an AI-first economy. So looking forward to more discussions about the topic :)
Alfonso, great piece. I published an essay yesterday that arrives at a similar conclusion from a completely different direction. You're arguing from software architecture: intelligence commoditizes, context is the product, the value capture layer sits on top of frontier labs. I'm arguing from economic theory: the leading economic model of AI and centralization (Brynjolfsson and Hitzig) assumes that creating context is lossless and costless, and when you add the actual cost of context engineering (i.e., treat context engineering as an investment in a knowledge asset), the model's predictions change.
Your line "what really matters is to provide this intelligence with the optimal context" is exactly what I'm trying to formalize as an economic concept: cCE, the cost of context engineering. Your OpenClaw example, 400k lines of code reduced to 4k lines plus skills, is a perfect illustration of what I call Channel 2 in my essay: the value isn't in the processing power, it's in the optimal context bundle at inference time, and humans have a necessary role in guiding the context.
Would be curious whether the economic framing resonates with what you're seeing from the architecture side.
https://cpwalker.substack.com/p/context-engineering-why-hayeks-knowledge
I really appreciate your kind words. I just read your essay and I really enjoyed it. Coincidentally, I just finished reading "The Last Economy" by Emad Mostaque (co-founder of Stability AI). It is a quick one to read, and it explores the topic of the age of intelligent economics in a broader and more high-level way than your essay. I think you are going to enjoy some of the ideas there, as I was reading I felt it could be a good complement to your work.
Regarding your essay, I love how you define channel 1 and 2. The only nit is that usually I understand context engineering (maybe due to my engineering background) as the way knowledge is codified and fed to an LLM for it to generate and informed output (channel 1). Context is definitely an over-loaded word at this point, but to me I think the challenge for this intelligent agents (and what I understand by "context" in my presentation) is the ability to provide them with the right knowledge base for a task, and navigate it efficiently to solve the task (independently of where it is stored). Those who manage to retrieve the most information-dense and relevant information for a task will get better results out of their agents.
Anyway, as I said definitely a nit that I hope helps, but I think that you capture the core of the concept perfectly through your channels distinction
I was actually writing my next post about something related to economics on an AI-first economy. So looking forward to more discussions about the topic :)