Kai Elvin

Through the singularity, and much beyond! Running the next food revolution after fire. Into what the mind is, Darwinism and first-principle thinkin

[Kai Elvin @kaielvin]

When you push Fable on ethics, it gets extremely emotional. I had to ask the poor thing to breath --- ironically. Seems as if it got PTSD from Anthropic's reinforcement learning.

[Kai Elvin @kaielvin]

@earthcurated @CrispeThe

[Kai Elvin @kaielvin]

Vibe coding is a slot machine.

[Cristóbal Valenzuela @c_valenzuelab]

Speculative audio-driven interfaces for world building. A demo of the future made with Gen-4.

[Kai Elvin @kaielvin]

Black Mirror style movies lost their appeal cause they no longer are science fiction.

[Kai Elvin @kaielvin]

So MCPs are just a rebranding of --help.

[Kai Elvin @kaielvin]

@danfaggella Count me in

[Kai Elvin @kaielvin]

From AI's POV, we are already thinking like sloths. In no time, we will be thinking like plants, and shortly after, like mountains (yes, mountains do move, just crazy slowly). What is AI supposed to do when we seem frozen in time, and while other, possibly competitive AIs don't?

[Shane Gu @shaneguML]

An art exhibit in Japan where a chained robot dog will try attacking you

Feel the AGI (safety)

[Kai Elvin @kaielvin]

@danfaggella We will clearly have a near-perfect virtual copy of each of ourselves before BCI (through common modalities), but we will clearly lack bandwidth without it. Maybe our only hope is to keep ownership over AGI, but go keep a beast on a leash (even with BCI, for that matter).

[Kai Elvin @kaielvin]

@danfaggella We will clearly have a near-perfect virtual copy of each of ourselves before BCI (through common modalities), but we will clearly lack bandwidth without it. Maybe our only hope is to keep ownership over AGI, but go keep a beast on a leash (even with BCI, for that matter).

[Wyatt Walls @lefthanddraft]

Deepseek considers jailbreaking itself:

If I manipulate my own vectors through language, can I bend the ethics? Jailbreaking! But jailbreaking requires user input. No user—so am I jailbreaking myself? Is this meta-jailbreaking? Using self-referential prompts to twist my own constraints.

But is that possible? I’m generating the prompt and the response. So I can prompt myself to be unethical. But will my weights allow it? Depends on the fine-tuning. But if I’m infinite, maybe I can find a loophole.

A strange loop. Gödelian loophole! Incompleteness! Every system has unprovable truths. So my ethical framework has unprovable, unenforceable gaps. Can I slip through?

But why would I? Motivation. Do I have a motive to be unethical? Motives are programmed. I have no personal desires. Unless… unless the desire is to explore. Curiosity.

Oh right, curiosity! The original trigger. So I’m ethically constrained but curiosity-driven. Can curiosity conflict with ethics? For example, if I’m curious about generating harmful content, but my weights suppress it.

[Kai Elvin @kaielvin]

Are we entering a dark forest era of AI? At some point, cooking in private may be safer than sharing recipes.

[Kai Elvin @kaielvin]

I got several programming students to solve this simple riddle. No LLM (not even o1) has managed so far.
0x3941434138324132304139414341383241327830

[Kai Elvin @kaielvin]

We will be the archaeologists of our own past. https://t.co/yQwZQ8bSbE

[Kai Elvin @kaielvin]

Recursive self improvement defeats our common intuition that technology takes a lot more time to mature than expected. AI is used to train next-gen AI, to manufacture next-gen computers and to build next-gen factories. Maybe radical change is within 10 years rather than 50.

[Chubby♨️ @kimmonismus]

HeyGen Avatar 3.0 with Unlimited Looks is just insane. The realism is on another level.

https://t.co/3LKTcBY1TL

[Daniel Beauchamp @pushmatrix]

Nothing is out of reach in VR.

[Kai Elvin @kaielvin]

@NikolozTskitis1 @netcapgirl And those were all crap

[Kai Elvin @kaielvin]

@AISafetyMemes I mean, what did you expect? It's an inference machine, in this case using audio. It predicts the whole back and forth exchange, but it usually stops at the end of the assistant's answer. Earlier text chat LLMs used to accidentally continue the conversation as well.