Nietzschean Ekko Enjoyer
Yet another thing solved by AgentV3N's prompt routing.
We're moving closer and closer to using 1B models for the first stage of prompt evaluation. A simple "thank you" could be met with an agent-level prepared response, or even just an emoji reaction.
Meanwhile, ChatGPT by default uses the last model you used in the conversation. Imagine blowing o1 reasoning on "thanks mate". Wouldn't be me.
SuperDicq
@r000t@ligma.pro The survey found that around 70% of people are polite to AI when interacting with it, with 12% being polite in case of a robot uprising.Holy shit people are stupid.
SuperDicq
@r000t@ligma.pro And people they are talking about this like it is about the environmental cost or whatever.
You know what I think? It's about fucking dignity.
I will not say "thank you" to a fucking machine. It's not a person.
Johnny Peligro
Johnny Peligro
Protoss
Yomiel
Yomiel
SuperDicq
@mischievoustomato@tsundere.love @r000t@ligma.pro I think anthropomorphizing chatbots is a really bad as the big AI companies will abuse this to create deceptive user experiences.
Johnny Peligro
Nietzschean Ekko Enjoyer
@Cyrillic
Bingo. That's also something we handle in the evaluation loop.
When a session is "closed" (released by the zone it was opened in), there's also what amounts to a "customer satisfaction survey" but it's performed in the background by the system loop. This is where the system can literally ask us for tools that would have made the request more efficient, more accurate (don't use a language calculator to calculate numbers), or possible at all.
@SuperDicq
Protoss
SuperDicq
@Cyrillic@lab.nyanide.com @r000t@ligma.pro You should only run AI models on your own machine and if that is not possible I definitely urge you against using any service that log yours prompts. Think about your privacy.
Nietzschean Ekko Enjoyer
@SuperDicq
What do you consider "deceptive" in this circumstance?
I can see a mihoyo-shaped object being like "Don't you love me enough to get the Professional plan?" but besides that I don't see a "persistent" customized character (what we call an Agent) being too terrible. It's basically a secretary at that point.
@mischievoustomato
Nietzschean Ekko Enjoyer
@SuperDicq
We wanted to do this, too! That's one of the reasons I got a laptop with an NPU (as a writeoff lmao) a few weeks ago.
The problem is that those "NPUs" are absolute fucking memes and basically can only do image classification. No good ways to run even 1B models on those.
But like hell yeah I want you to do as much as possible on *your* equipment that you pay for lmfao.
As for privacy, customers requiring HIPAA compliance (or similar) can force all inference to happen on our infra. The eval loop also looks for things that appear to be PHI and halt the request if the account is not enabled for HIPAA compliance.
@Cyrillic
SuperDicq
@mischievoustomato@tsundere.love @r000t@ligma.pro If it helps you add more context to prompts it's a good trick but I still recommend stripping out the "please" and "thank yous" after you have finished writing your prompt as all this does is add more tokens making the prompt harder to process while not improving results.
SuperDicq
@r000t@ligma.pro @Cyrillic@lab.nyanide.com I don't know where you work but gathering data on your customers in this sounds absolutely proprietary
SuperDicq
@r000t@ligma.pro @mischievoustomato@tsundere.love I am a firm believer that LLMs should return de-anthropomorphized responses only, like this example.
Why? Because normies are stupid and otherwise will start to think that there's some actual sort of sentience behind the scenes or something.
Johnny Peligro
翠星石
SuperDicq
@Suiseiseki@freesoftwareextremist.com @r000t@ligma.pro @mischievoustomato@tsundere.love I know this is a relatively small survery (1000 participants), but a staggering 12% even believe that if they are not polite to ChatGPT now they will be the first ones die in an unavoidable future AI uprising.
And I'm supposed to be the mentally ill because I have autism? Normies are not alright.
翠星石
Johnny Peligro
SuperDicq
@mischievoustomato@tsundere.love @r000t@ligma.pro That's just very cringe, ngl
Johnny Peligro
Nietzschean Ekko Enjoyer
@SuperDicq
It's just a 21st century Pascal's Wager. The part where they're clueless normies is that if you asked them now the information that your were nice makes its way to whatever is exterminating people, they won't be able to conceptualize that. It's all just "the AI" like there's "the cloud"
@Suiseiseki @mischievoustomato
翠星石
Nietzschean Ekko Enjoyer
@SuperDicq
rootworx Technologies.
And the testing realm puts a big banner on the client saying "Operators can view all messages on the testing realm."
@Cyrillic
Nietzschean Ekko Enjoyer
@mischievoustomato
Starfall makes hers pretend to be an exploding penguin from some weeb game and address her as the rule 63 protag from the same weeb game.
The default agent that will likely ship is called "Standard" and is somewhat close to being deanthropomorphized
Mine is the namesake of the platform itself and I can't wait to reveal him to everyone.
@SuperDicq
pettanko
Johnny Peligro
SuperDicq
@hakui@tuusin.misono-ya.info @r000t@ligma.pro I think having to say "thank you" to a kiosk is like one step further from having the kiosk say "thank you".
御園はくい
Light
SuperDicq
@forcedinductionretard@pl.pube.tk Even though it is an object that brings me joy it is still an object and I will not speak to it. I will also not use he or she pronouns like some people do with cars.
ForcedInductionRetard
Alexandre Oliva
SuperDicq
@lxo@snac.lx.oliva.nom.br @r000t@ligma.pro I never tip. I'm not American.
御園はくい
GNU social JP is a social network, courtesy of GNU social JP管理人. It runs on GNU social, version 2.0.2-dev, available under the GNU Affero General Public License.
All GNU social JP content and data are available under the Creative Commons Attribution 3.0 license.
