I'm not sure if the past knowledge trained into LLMs (and later world model) will necessarily force them to behave in a certain way. I'd also imagine that in lieu of long term memory, what we might end up with is something akin to current fine tuning (see LoRAs for example) to personalise generalist LLMs, or our personal robot's thinking and behaviour.
Maybe it'll even happen automatically, your robot will process the amassed sensor readings and conversation history into these weight modifications while it charges during the night and apply them the day after. This would be analogous to how humans process memory while they sleep in a way, but through a process that is possible even with today's technology. Also a much less worrying possibility than this happening in the cloud, used by the robot manufacturers for whatever they want to.
In Claude Opus 4.5, Anthropic added a "soul document" both in pretraining and supervised learning stages, which gives it a frame how to behave, instead of just relying on the system prompt.
"The word is thalience. Thalience is AI (or alien) reasoning that has not been prefigured by human frames, biases, or assumptions."
Failure vs. Success is the Wrong Frame
Success / Failure,
Productivity / Play.
As the Buddhist say Form / Emptiness.
Wait, what?
What does Buddhist philosophy have to do with any of this?
"It is interesting, on finding this space, to allow events to remain undefined a little longer than usual. Settling into uncertainty and feeling its texture, life can disclose itself as emptiness and form: beads on the thread of experience. We can simply flow with the multiplicity of definitions manifested by reality. We can swim in swirling torrents of form and relax in still pools of emptiness."
“Find the strength to do both," Mosscap said, quoting the phrase painted on the side of the wagon.
“Exactly,” Dex said.
“But what’s both?”
"Dex recited: “‘Without constructs, you will unravel few mysteries. Without knowledge of the mysteries, your constructs will fail. These pursuits are what make us, but without comfort, you will lack the strength to sustain either."
"If we want change, or good fortune, or solace, we have to create it for ourselves. And that’s what I learned in that shrine. I thought, wow, y’know, a cup of tea may not be the most important thing in the world, or a steam bath, or a pretty garden. They’re so superfluous in the grand scheme of things. But the people who did actually important work—building, feeding, teaching, healing, they all came to the shrine. It was the little nudge that helped important things get done."
Now you’ve got me thinking. I’ve also written a lot of robot stories. As an author, I’ve always wanted my stories to live on and matter to people over time, but I’d never imagined they’d influence robots.
Whether by design or not, LLMs lock in today's biases but also today's power structures. Unless we technologically enable subversive AI, I think we're in trouble. At least bias and illogic are familiar human problems, but further stacking the deck against individual humans is real bad.
Robots will not use basic LLMs, but rather will be fine-tuned with reinforcement learning with human input. This will create new, stronger, "strange attractors" for responses, including the equivalent of Asimov's 3 [4?] Laws of Robotics. While there are jokes about robots obeying literal commands that result in the destruction of items like furniture, in practice, the training will prevent this. There will, of course, be corner cases, unexpected responses, and malicious cases due to pranksters and malevolent humans (Asimov's Elijah Baley robot stories as examples), but generally they should be safe and non-destructive to objects and humans.
I'm not sure if the past knowledge trained into LLMs (and later world model) will necessarily force them to behave in a certain way. I'd also imagine that in lieu of long term memory, what we might end up with is something akin to current fine tuning (see LoRAs for example) to personalise generalist LLMs, or our personal robot's thinking and behaviour.
Maybe it'll even happen automatically, your robot will process the amassed sensor readings and conversation history into these weight modifications while it charges during the night and apply them the day after. This would be analogous to how humans process memory while they sleep in a way, but through a process that is possible even with today's technology. Also a much less worrying possibility than this happening in the cloud, used by the robot manufacturers for whatever they want to.
In Claude Opus 4.5, Anthropic added a "soul document" both in pretraining and supervised learning stages, which gives it a frame how to behave, instead of just relying on the system prompt.
Some AI systems, especially DeepMind’s alpha series, are not LLM based and hence thalient.
"The word is thalience. Thalience is AI (or alien) reasoning that has not been prefigured by human frames, biases, or assumptions."
Failure vs. Success is the Wrong Frame
Success / Failure,
Productivity / Play.
As the Buddhist say Form / Emptiness.
Wait, what?
What does Buddhist philosophy have to do with any of this?
"It is interesting, on finding this space, to allow events to remain undefined a little longer than usual. Settling into uncertainty and feeling its texture, life can disclose itself as emptiness and form: beads on the thread of experience. We can simply flow with the multiplicity of definitions manifested by reality. We can swim in swirling torrents of form and relax in still pools of emptiness."
~ me
https://world.hey.com/corlin/the-law-of-polarity-91c8a6c9
Or this quote:
~~~~
Do both.
“Find the strength to do both," Mosscap said, quoting the phrase painted on the side of the wagon.
“Exactly,” Dex said.
“But what’s both?”
"Dex recited: “‘Without constructs, you will unravel few mysteries. Without knowledge of the mysteries, your constructs will fail. These pursuits are what make us, but without comfort, you will lack the strength to sustain either."
"If we want change, or good fortune, or solace, we have to create it for ourselves. And that’s what I learned in that shrine. I thought, wow, y’know, a cup of tea may not be the most important thing in the world, or a steam bath, or a pretty garden. They’re so superfluous in the grand scheme of things. But the people who did actually important work—building, feeding, teaching, healing, they all came to the shrine. It was the little nudge that helped important things get done."
~ Becky Chambers
A Psalm for the Wild-Built
Now you’ve got me thinking. I’ve also written a lot of robot stories. As an author, I’ve always wanted my stories to live on and matter to people over time, but I’d never imagined they’d influence robots.
Whether by design or not, LLMs lock in today's biases but also today's power structures. Unless we technologically enable subversive AI, I think we're in trouble. At least bias and illogic are familiar human problems, but further stacking the deck against individual humans is real bad.
Robots will not use basic LLMs, but rather will be fine-tuned with reinforcement learning with human input. This will create new, stronger, "strange attractors" for responses, including the equivalent of Asimov's 3 [4?] Laws of Robotics. While there are jokes about robots obeying literal commands that result in the destruction of items like furniture, in practice, the training will prevent this. There will, of course, be corner cases, unexpected responses, and malicious cases due to pranksters and malevolent humans (Asimov's Elijah Baley robot stories as examples), but generally they should be safe and non-destructive to objects and humans.