Silicon Promises, estimated statistics, sunnier realities: the slow diffusion of “artificial Intelligence”. rom steam to silicon, history’s most transformative inventions rarely deliver on schedule...
"Second, somebody ought to make it their thing to push the Silicon Law of Attention Conservation. If an improvement in information technology means that you and others can write three times as fast, that also means that you have three times as much to read and think about. If, initially, you spent more than 3/4 of your time writing, you come out ahead. If, initially, you spent less than 3/4 of your time writing, you come out behind."
It seems to me that there is yet another aspect when one considers economic productivity.
Assume that there is some generally available information technology that "means that you and others can write three times as fast". If this technology is generally available, then "others" will include all (or most) of your competitors, as well. And this means that, though all of you may be more "productive" in some sense, there will be little or not competitive advantage and thus little room for anyone to profit from that greater productivity. Which in turn means that any increase in GDP may turn out to be minimal.
A simple test. Download the text from a recent fiction book that you have recently read and ask an LLM to summarize it. I've found that the summary will make basic plot errors while bs'ing its way through with vapid insights.
Yes: **extreme** verbal fluency, and—I think—to the extent that the LLM is tied down by a huge amount of paragraphs containing true and valid insights that are in its training data that it judges as "close" to the prompt, it will come out with valid and even striking insights. But it has no idea whatsoever about the world—not even the fictional story-world that is the plot.
Watch Andrej Karpathy trying to teach people how to set up situations so that the machine is nooodged so that each GPT LLM call lands it in a spot in its vector-space regularization function of next-word continuation that will do something useful:
> Andrej Karpathy: +1 for "context engineering" over "prompt engineering" <https://x.com/karpathy/status/1937902205765607626/>: 'People associate prompts with short task descriptions you'd give an LLM in your day-to-day use. When in every industrial-strength LLM app, context engineering is the delicate art and science of filling the context window with just the right information for the next step. Science because doing this right involves:
> * task descriptions and explanations,
> * few shot examples,
> * RAG,
> * related (possibly multimodal) data, tools, state and history,
> * compacting...
> Too little or of the wrong form and the LLM doesn't have the right context for optimal performance. Too much or too irrelevant and the LLM costs might go up and performance might come down. Doing this well is highly non-trivial. And art because of the guiding intuition around LLM psychology of people spirits.
> On top of context engineering itself, an LLM app has to:
> - break up problems just right into control flows
> - pack the context windows just right
> - dispatch calls to LLMs of the right kind and capability
> - handle generation-verification UIUX flows
> - a lot more - guardrails, security, evals, parallelism, prefetching, ...
> So context engineering is just one small piece of an emerging thick layer of non-trivial software that coordinates individual LLM calls (and a lot more) into full LLM apps. The term "ChatGPT wrapper" is tired and really, really wrong...
This is serious, serious "Clever Hans" territory: you are doing an awful lot of work to nooodge the verbally-fluent system into a state in which it (falsely) appears (for a while) to have knowledge about the world. Also relevant is Bjarnesen (2023) <https://softwarecrisis.dev/letters/llmentalist/>:
> Kent: A simple test. Download the text from a recent fiction book that you have recently read and ask an LLM to summarize it. I've found that the summary will make basic plot errors while bs'ing its way through with vapid insights...
I remember in Fleming's Barrow's Boys that two arctic explorers chose steam engine-powered ships. The admiralty thought steam engine ships were idiotic and always would be; never leave tried-and-true sails. The admirals were right at the time, the steam engine ships were pathetically slow and constantly broke down.
"This productivity decline is not real—a better, human welfare-based measure would include the value to the user of interacting with Amazon via the natural-language interface."
Easy on there, Brad. Yes, of course it is possible that AI has some unmeasured human welfare improvements. But taking this for granted? That way lies madness. Once you invoke the daemon of unmeasurable, cryptic benefits, you cannot refute its doppelgänger, the unmeasurable, cryptic cost.
This doesn't take enshittification into account. Amazon search is hard to use because it was deliberately made hard to use. It used to be keyword and criteria based and much better. Now, Amazon sells product placement and search intentionally includes irrelevant items and the order of the items presented no longer provides useful information. Selling search ads and search position is a solid revenue stream for Amazon which raises a simple question: why would Amazon provide an AI search technology that makes it easier for users to find goods when an AI search technology that offers a revenue stream for placement and provides poor answers so it can present more paid advertising would be more profitable?
I don't see AI creating user value in this context. It's going to be something out of bad Monty Python skit with users asking about merchandise and getting verbose answers mentioning multiple goods with some of them being relevant, some non-sequitor and others paid placements. Follow up queries attempting to refine or focus the search will yield further hilarity. Throw in the occasional hallucination and Amazon profits nicely.
Worse, AI has already led to an increasingly intense battle between fake review generation and fake review detection. Product descriptions will become decreasingly useful. Users will have to wade through a deeper and deeper swamp. I've already dramatically cut back on my Amazon purchases since it is often easier to just use Kagi or directly go to a product category site. Perhaps Amazon will add a new tier for a few hundred dollars a year called Amazon Classic with Amazon's old search engine and a less corrupted AI component, but the company does have its responsibility to its shareholders.
Exactly how this is going to increase the GDP is unclear.
Yes: in my view, ad-supported AI is truly going to be a New Fresh Hell of DysCognition:
> Kaleberg: This doesn't take enshittification into account. Amazon search is hard to use because it was deliberately made hard to use. It used to be keyword and criteria based and much better. Now, Amazon sells product placement and search intentionally includes irrelevant items and the order of the items presented no longer provides useful information. Selling search ads and search position is a solid revenue stream for Amazon which raises a simple question: why would Amazon provide an AI search technology that makes it easier for users to find goods when an AI search technology that offers a revenue stream for placement and provides poor answers so it can present more paid advertising would be more profitable?
> I don't see AI creating user value in this context. It's going to be something out of bad Monty Python skit with users asking about merchandise and getting verbose answers mentioning multiple goods with some of them being relevant, some non-sequitor and others paid placements. Follow up queries attempting to refine or focus the search will yield further hilarity. Throw in the occasional hallucination and Amazon profits nicely.
> Worse, AI has already led to an increasingly intense battle between fake review generation and fake review detection. Product descriptions will become decreasingly useful. Users will have to wade through a deeper and deeper swamp. I've already dramatically cut back on my Amazon purchases since it is often easier to just use Kagi or directly go to a product category site. Perhaps Amazon will add a new tier for a few hundred dollars a year called Amazon Classic with Amazon's old search engine and a less corrupted AI component, but the company does have its responsibility to its shareholders.
> Exactly how this is going to increase the GDP is unclear...
Similarly, how does Google make more money by immediately providing users the info they want rather than paid rankings and clicks? Marketing does not optimize for the user's efficiency; quite the opposite.
I use Kagi to find sellers, often Amazon's source supplier, and haven't bought anything from Amazon year to date.
Again, excellent. Don't forget that for every Amazon there aren't just dozens of other legacy business put out of business, there are also a thousand other start ups trying to be Amazon and (when they fail) flushing lots of money and effort down the toilet. Technological innovation is expensive.
Again, chiming in very late, but...
"Second, somebody ought to make it their thing to push the Silicon Law of Attention Conservation. If an improvement in information technology means that you and others can write three times as fast, that also means that you have three times as much to read and think about. If, initially, you spent more than 3/4 of your time writing, you come out ahead. If, initially, you spent less than 3/4 of your time writing, you come out behind."
It seems to me that there is yet another aspect when one considers economic productivity.
Assume that there is some generally available information technology that "means that you and others can write three times as fast". If this technology is generally available, then "others" will include all (or most) of your competitors, as well. And this means that, though all of you may be more "productive" in some sense, there will be little or not competitive advantage and thus little room for anyone to profit from that greater productivity. Which in turn means that any increase in GDP may turn out to be minimal.
A simple test. Download the text from a recent fiction book that you have recently read and ask an LLM to summarize it. I've found that the summary will make basic plot errors while bs'ing its way through with vapid insights.
Yes: **extreme** verbal fluency, and—I think—to the extent that the LLM is tied down by a huge amount of paragraphs containing true and valid insights that are in its training data that it judges as "close" to the prompt, it will come out with valid and even striking insights. But it has no idea whatsoever about the world—not even the fictional story-world that is the plot.
Watch Andrej Karpathy trying to teach people how to set up situations so that the machine is nooodged so that each GPT LLM call lands it in a spot in its vector-space regularization function of next-word continuation that will do something useful:
> Andrej Karpathy: +1 for "context engineering" over "prompt engineering" <https://x.com/karpathy/status/1937902205765607626/>: 'People associate prompts with short task descriptions you'd give an LLM in your day-to-day use. When in every industrial-strength LLM app, context engineering is the delicate art and science of filling the context window with just the right information for the next step. Science because doing this right involves:
> * task descriptions and explanations,
> * few shot examples,
> * RAG,
> * related (possibly multimodal) data, tools, state and history,
> * compacting...
> Too little or of the wrong form and the LLM doesn't have the right context for optimal performance. Too much or too irrelevant and the LLM costs might go up and performance might come down. Doing this well is highly non-trivial. And art because of the guiding intuition around LLM psychology of people spirits.
> On top of context engineering itself, an LLM app has to:
> - break up problems just right into control flows
> - pack the context windows just right
> - dispatch calls to LLMs of the right kind and capability
> - handle generation-verification UIUX flows
> - a lot more - guardrails, security, evals, parallelism, prefetching, ...
> So context engineering is just one small piece of an emerging thick layer of non-trivial software that coordinates individual LLM calls (and a lot more) into full LLM apps. The term "ChatGPT wrapper" is tired and really, really wrong...
This is serious, serious "Clever Hans" territory: you are doing an awful lot of work to nooodge the verbally-fluent system into a state in which it (falsely) appears (for a while) to have knowledge about the world. Also relevant is Bjarnesen (2023) <https://softwarecrisis.dev/letters/llmentalist/>:
> Kent: A simple test. Download the text from a recent fiction book that you have recently read and ask an LLM to summarize it. I've found that the summary will make basic plot errors while bs'ing its way through with vapid insights...
I remember in Fleming's Barrow's Boys that two arctic explorers chose steam engine-powered ships. The admiralty thought steam engine ships were idiotic and always would be; never leave tried-and-true sails. The admirals were right at the time, the steam engine ships were pathetically slow and constantly broke down.
"This productivity decline is not real—a better, human welfare-based measure would include the value to the user of interacting with Amazon via the natural-language interface."
Easy on there, Brad. Yes, of course it is possible that AI has some unmeasured human welfare improvements. But taking this for granted? That way lies madness. Once you invoke the daemon of unmeasurable, cryptic benefits, you cannot refute its doppelgänger, the unmeasurable, cryptic cost.
This doesn't take enshittification into account. Amazon search is hard to use because it was deliberately made hard to use. It used to be keyword and criteria based and much better. Now, Amazon sells product placement and search intentionally includes irrelevant items and the order of the items presented no longer provides useful information. Selling search ads and search position is a solid revenue stream for Amazon which raises a simple question: why would Amazon provide an AI search technology that makes it easier for users to find goods when an AI search technology that offers a revenue stream for placement and provides poor answers so it can present more paid advertising would be more profitable?
I don't see AI creating user value in this context. It's going to be something out of bad Monty Python skit with users asking about merchandise and getting verbose answers mentioning multiple goods with some of them being relevant, some non-sequitor and others paid placements. Follow up queries attempting to refine or focus the search will yield further hilarity. Throw in the occasional hallucination and Amazon profits nicely.
Worse, AI has already led to an increasingly intense battle between fake review generation and fake review detection. Product descriptions will become decreasingly useful. Users will have to wade through a deeper and deeper swamp. I've already dramatically cut back on my Amazon purchases since it is often easier to just use Kagi or directly go to a product category site. Perhaps Amazon will add a new tier for a few hundred dollars a year called Amazon Classic with Amazon's old search engine and a less corrupted AI component, but the company does have its responsibility to its shareholders.
Exactly how this is going to increase the GDP is unclear.
Yes: in my view, ad-supported AI is truly going to be a New Fresh Hell of DysCognition:
> Kaleberg: This doesn't take enshittification into account. Amazon search is hard to use because it was deliberately made hard to use. It used to be keyword and criteria based and much better. Now, Amazon sells product placement and search intentionally includes irrelevant items and the order of the items presented no longer provides useful information. Selling search ads and search position is a solid revenue stream for Amazon which raises a simple question: why would Amazon provide an AI search technology that makes it easier for users to find goods when an AI search technology that offers a revenue stream for placement and provides poor answers so it can present more paid advertising would be more profitable?
> I don't see AI creating user value in this context. It's going to be something out of bad Monty Python skit with users asking about merchandise and getting verbose answers mentioning multiple goods with some of them being relevant, some non-sequitor and others paid placements. Follow up queries attempting to refine or focus the search will yield further hilarity. Throw in the occasional hallucination and Amazon profits nicely.
> Worse, AI has already led to an increasingly intense battle between fake review generation and fake review detection. Product descriptions will become decreasingly useful. Users will have to wade through a deeper and deeper swamp. I've already dramatically cut back on my Amazon purchases since it is often easier to just use Kagi or directly go to a product category site. Perhaps Amazon will add a new tier for a few hundred dollars a year called Amazon Classic with Amazon's old search engine and a less corrupted AI component, but the company does have its responsibility to its shareholders.
> Exactly how this is going to increase the GDP is unclear...
Similarly, how does Google make more money by immediately providing users the info they want rather than paid rankings and clicks? Marketing does not optimize for the user's efficiency; quite the opposite.
I use Kagi to find sellers, often Amazon's source supplier, and haven't bought anything from Amazon year to date.
Again, excellent. Don't forget that for every Amazon there aren't just dozens of other legacy business put out of business, there are also a thousand other start ups trying to be Amazon and (when they fail) flushing lots of money and effort down the toilet. Technological innovation is expensive.