Emily M. Bender

Bender's linked May 12, 2026 post "Frequently Unasked Questions", https://medium.com/@emilymenonbender/stochastic-parrots-frequently-una… , was a better read.

> The real issue in my opinion is that we will hit practical limits with training data and computational resources well before AGI turns us all into paperclips [...] I think you are correct, but also I think that even if that were not the case, the Thai Library Problem[1] strongly suggests that AGI will have to be built on something other than LLMs (even if LLM-derived systems were to serve as an interface to such systems). [1]: https://medium.com/@emilymenonbender/thought-experiment-in-the-nationa…

Dunno if vlad can see this, but https://kagi.com/images?q=baby+peacock ...shows that infamous AI-generated peacock image multiple times on the first row of results. Merely filtering out websites that tend to have lots of AI images does not prevent this failure case, since (for example): https://birdfact.com/articles/baby-peacocks has the fake image in there, as an example of "What does a baby peacock not look like?" As Emily Bender has correctly pointed out, AI images are like an oil spill, and the cleanup (if such a thing is even feasible) will be challenging: https://medium.com/@emilymenonbender/cleaning-up-a-baby-peacock-sullie… (edited to add that I'm a paying Kagi customer, and this failure case isn't a ding against my overall impression of what Kagi is, and I'll continue using it)

This perspective is interesting, but I don't think it refutes some of the key elements of concerns that Bender and colleagues have put forward. I really like the "Thai Library" thought experiment [1], and I encourage anyone to think through what you'd actually try to do, how you'd learn, and what you'd know or understand at the end. I can totally imagine that a sufficiently determined human in that scenario could develop many skills needed to generate continuations based on input texts, and use a different mix of skills for different situations, in a way that mirrors the setup described here. But ... would that change whether you actually "understand" Thai, or "know" the contents of texts you had previously processed? [1] https://medium.com/@emilymenonbender/thought-experiment-in-the-nationa…

> This is like saying "We understand practically everything about airplanes, except how they stay in the air." Even that is a sufficient level of understanding to correctly determine that a motorcycle is not an airplane. While we might not have a complete picture of what consciousness entails, we can at least list some necessary conditions for it to arise. Any system that lacks those conditions can at least be proven to not be conscious. With LLMs specifically, I think there is a very strong argument that they are not and cannot be conscious at all, regardless of how big of a corpus you throw at it or how many parameters it has. Emily Bender explains it well here: https://medium.com/@emilymenonbender/thought-experiment-in-the-nationa…