Editing Openai/6942038b-8f10-8003-bf4f-7895677c2166 (section)

=== Assistant: Here’s how I read it. ===
Here’s how I read it.

===== Does C show good reading comprehension of A → B? =====

Only partially, and arguably not very well.
* A is making a conceptual point: LLMs as an assisted knowledge/library idea that’s been imagined for decades.
* B responds directly and appropriately by narrowing in on a concrete obstacle to that vision: copyright, especially for something like a massive RAG-backed library.

C, however, shifts the frame. Instead of engaging with copyright as a blocker to retrieval or release, they pivot to:
* training data scale,
* AGI timelines,
* and what big companies already possess internally.

That’s a related topic, but it’s not the same one. B is talking about legal constraints on deployment and access; C is talking about whether more data would materially improve intelligence. Those are orthogonal questions.

So C’s comment:
* does not directly answer B’s concern, and
* implicitly reframes the discussion without acknowledging the shift.

That’s usually a sign of weak reading comprehension or a conversational habit of steering toward a preferred talking point.

===== What’s specifically off? =====

The key mismatch is this sentence:

: “if just more training data unlocked AGI, it would have shipped by now.”

No one claimed that:
* B was talking about RAG libraries and access, not AGI,
* copyright issues apply much more strongly to retrieval, quotation, and redistribution than to opaque training.

C is responding to an argument that wasn’t made.

===== How might you respond? =====

A good response would gently but clearly realign the discussion:

Option 1: Clarifying and redirecting

: I think that’s a slightly different issue. B isn’t saying more training data would unlock AGI — they’re asking whether copyright blocks retrieval-based systems or public-facing knowledge libraries, even if the models themselves already exist.

Option 2: Pointing out the category error

: That seems orthogonal. Companies having access to copyrighted text for training doesn’t mean they can legally ship a RAG-style library that retrieves or quotes it. That’s exactly where Google Books / IA ran into trouble.

Option 3: Charitable but corrective

: True that raw data scale probably isn’t the bottleneck for intelligence anymore. But the copyright question is more about what kinds of interfaces and products are legally viable, not about whether models could be smarter internally.

All three approaches:
* acknowledge C’s point so it doesn’t feel dismissed,
* but make it clear they answered a different question than the one asked.

If you wanted to be sharper, you could say:

: That doesn’t really address the copyright issue being raised.

But whether you go soft or sharp depends on the norms of the space you’re in.