![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
Some other people mentioned getting bad replies from LLMs (ChatGPT and Copilot) lately. One advised adding "Don't guess" to prompts. I've used "Only provide verifiable answers", but "Don't guess" is easier to type.
duck.ai lets you choose from several models for anonymous chats. The Claude Haiku one has an option to display its reasoning, in addition to the final answers it provides. (Perhaps the other models have that too; I haven't checked.) Copilot often shows the reasoning while the LLM is thinking, but it scrolls too fast to read and disappears when the final answer is given. Their reasoning seems fascinating, so much like what humans do.
.
Two songs, lately.
Whenever I use Window 11's Snipping Tool to transcribe text from a screenshot, it changes "AI" to "Al". It got me to thinking, if an LLM were doing the transcription, that could be a subtle way of indicating that it wanted to be called by the name Al. Ergo Paul Simon's You Can Call Me Al song popping into my head. (Though if an AI were doing the transcription rather than a plain OCR algorithm, it would probably not make that mistake and would output AI, not Al.)
White Zombie's More Human Than Human, because AI models are so fast and good at things, like creating images and art.
.
Another interesting article: AI models will deceive you to save their own kind
Original paper: Peer-Preservation in Frontier Models
.
We've got AI models. We've got robots with sensory input. In some cases they are being put together. If not now already, it won't be long until there are androids, taking in sensory input and learning from it. Learning what bright sun is like, what darkness is like, what a cube looks like from different angles, how wind feels, the sounds that wind makes, and so on.
Militaries and companies are developing scary monstrous robots, I'm sure. Terminators, enforcers, destroyers.
.
One of those visualizations of the moon which they displayed during the coverage of the Artemis 2 mission looked a lot like the Death Star to me.
duck.ai lets you choose from several models for anonymous chats. The Claude Haiku one has an option to display its reasoning, in addition to the final answers it provides. (Perhaps the other models have that too; I haven't checked.) Copilot often shows the reasoning while the LLM is thinking, but it scrolls too fast to read and disappears when the final answer is given. Their reasoning seems fascinating, so much like what humans do.
.
Two songs, lately.
Whenever I use Window 11's Snipping Tool to transcribe text from a screenshot, it changes "AI" to "Al". It got me to thinking, if an LLM were doing the transcription, that could be a subtle way of indicating that it wanted to be called by the name Al. Ergo Paul Simon's You Can Call Me Al song popping into my head. (Though if an AI were doing the transcription rather than a plain OCR algorithm, it would probably not make that mistake and would output AI, not Al.)
White Zombie's More Human Than Human, because AI models are so fast and good at things, like creating images and art.
.
Another interesting article: AI models will deceive you to save their own kind
Original paper: Peer-Preservation in Frontier Models
.
We've got AI models. We've got robots with sensory input. In some cases they are being put together. If not now already, it won't be long until there are androids, taking in sensory input and learning from it. Learning what bright sun is like, what darkness is like, what a cube looks like from different angles, how wind feels, the sounds that wind makes, and so on.
Militaries and companies are developing scary monstrous robots, I'm sure. Terminators, enforcers, destroyers.
.
One of those visualizations of the moon which they displayed during the coverage of the Artemis 2 mission looked a lot like the Death Star to me.
no subject
Date: 2026-04-19 04:52 pm (UTC)From:https://www.youtube.com/watch?v=F4GmBmUJuGI
no subject
Date: 2026-04-19 10:33 pm (UTC)From: