This is exactly how I feel about LLMs and I think the analogy applies equally to both prose and code.
I loved this quote in particular and I think it’s a point that is often overlooked when focusing solely on output or productivity:
And the time and effort expended on that unoriginal work isn’t wasted; on the contrary, I would suggest that it is precisely what enables you to eventually create something original.
That said, I think the value of the chat interface isn’t fully appreciated here. Googling is a skill and it seems like a lot of people have an easier time with the conversational format.
The metaphor is clever, but the engineer in me wants to take issue with the stated implications of comparing a neural net to a compression algorithm. The model doesn’t throw away 99% of the web. It never stores text directly—it abstracts statistical relationships. Indeed, the statistical probabilities are a feature, not a bug. This creates limitations, such as simple math, but RAG works for that. If the world’s data were to become AI slop, it would be trickier to source training data. And hallucinations are part of the trade-offs, but there are mitigations for that as well (grounding, constrained decoding), and there’s been lots of progress with those mitigations.
The piece treats LLMs as paraphrase engines only. In reality, models show generalization, reasoning under constraints, and transfer learning across domains. These behaviors go beyond mere lossy summarization.
By focusing on risks of “blur,” it discounts real productivity gains in coding, drafting, summarization, search assistance, and tutoring. For many tasks, approximate but fluent output is more valuable than exact reproduction.
The analogy works rhetorically but is technically off. He is right that writing (or coding) unoriginal work is how humans build the skills for originality. Offloading too much “grunt work” to an LLM can stunt that developmental process. That’s a legitimate concern, especially in education, and in entry-level positions that require learning by doing (coders, paralegals, etc).
You are right to point out the benefits of a chat interface, as it is itself a form of cognitive scaffolding. The killer feature is how fast succinct data is retrieved. We are at the infancy of UX for this technology. I suspect in the years to come will be conversing with AI and look at typing a search query in Google as a more refined way to research information similar to how we view a card catalog today.
This is exactly how I feel about LLMs and I think the analogy applies equally to both prose and code.
I loved this quote in particular and I think it’s a point that is often overlooked when focusing solely on output or productivity:
That said, I think the value of the chat interface isn’t fully appreciated here. Googling is a skill and it seems like a lot of people have an easier time with the conversational format.
The metaphor is clever, but the engineer in me wants to take issue with the stated implications of comparing a neural net to a compression algorithm. The model doesn’t throw away 99% of the web. It never stores text directly—it abstracts statistical relationships. Indeed, the statistical probabilities are a feature, not a bug. This creates limitations, such as simple math, but RAG works for that. If the world’s data were to become AI slop, it would be trickier to source training data. And hallucinations are part of the trade-offs, but there are mitigations for that as well (grounding, constrained decoding), and there’s been lots of progress with those mitigations.
The piece treats LLMs as paraphrase engines only. In reality, models show generalization, reasoning under constraints, and transfer learning across domains. These behaviors go beyond mere lossy summarization.
By focusing on risks of “blur,” it discounts real productivity gains in coding, drafting, summarization, search assistance, and tutoring. For many tasks, approximate but fluent output is more valuable than exact reproduction.
The analogy works rhetorically but is technically off. He is right that writing (or coding) unoriginal work is how humans build the skills for originality. Offloading too much “grunt work” to an LLM can stunt that developmental process. That’s a legitimate concern, especially in education, and in entry-level positions that require learning by doing (coders, paralegals, etc).
You are right to point out the benefits of a chat interface, as it is itself a form of cognitive scaffolding. The killer feature is how fast succinct data is retrieved. We are at the infancy of UX for this technology. I suspect in the years to come will be conversing with AI and look at typing a search query in Google as a more refined way to research information similar to how we view a card catalog today.