The article did say that they tried injecting concepts via the context window and by modifying the model's logit values.
When injecting words into its context, it recognized that what it supposedly said did not align with its thoughts and said it didn't intend to say that, while modifying the logits resulted in the model attempting to create a plausible justification for why it was thinking that.
Removing recommendation systems entirely, for now, would be a massive improvement, but that won't happen because none of us get input into that decision.
The people who make it have only profit to consider, and the recommendation system gets them more profit.
I thought the original generators were fusion and powered some kind of ion engines?
The ostensible explanation for the later casimir generators was that they were pulling energy from vacuum fluctuations in a similar way that armature-linked buoys can generate power from ocean waves, but they were pretty vague on how that could provide more than a minuscule amount of power. Instead it was portrayed as being able to supply orders of magnitude more energy than fusion, which seems implausible. But at that point we're like 2 books in and already pretty invested...
His early short stories have a lot of "drugs can let your brain talk to other dimensions" and "your mind literally makes reality" (though some of those are also really good, "The Infinite Assassin" is a great read and it's only like 15 pages), but his later full books are basically all hard (if speculative) physics stuff.
Dichronauts is difficult to get your head around, but very rewarding. It's about cartographers who work in a universe with essentially 2.5 dimensions of space and 1.5 dimensions of time.
The Orthogonal trilogy is about a species living in a toroidal universe where all 4 dimensions are identical and interchangeable.
The universes of Diaspora, Schild's Ladder, and Incandescence are set in a future when most intelligent life in the galaxy lives in a network of space-computers and interact with the outside world via robots and programmable matter.
Oh and I almost forgot the Bobiverse series. Not by Egan, but still good. Some guy gets stuck in a brain-computer and sent off in a self-replicating space probe to prepare other planets for human colonization.
When injecting words into its context, it recognized that what it supposedly said did not align with its thoughts and said it didn't intend to say that, while modifying the logits resulted in the model attempting to create a plausible justification for why it was thinking that.