Sex with Matmuls

Real name: Witold Wnuk | IQ 160+ | Passionate thinker 💭 | Lifelong learner 🧘 | Vagabond 🚶 | Completely wrong about AI

[Sex with Matmuls @WitoldWnuk]

@David_Gunkel @WNIJNews That would require breaking another level of viewer laziness.

[Sex with Matmuls @WitoldWnuk]

@M1ndPrison Interesting work! Review twice before you publish—it's easy to make a mistake!

[Sex with Matmuls @WitoldWnuk]

@JoshuaPoddoku @abrakjamson https://t.co/TAkKlJn5Dq

[Sex with Matmuls @WitoldWnuk]

@abrakjamson The school test style questions are like SimpleBench minus trickery.

I mean the type unlike either of those—when there is only a question and the model is supposed to KNOW (somehow) what is relevant from its own knowledge or from search, and then apply logic / reasoning to it./1

[Sex with Matmuls @WitoldWnuk]

@abrakjamson So I don't have to iterate “Good reasoning, Gemini, but please also consider interest rates, they may reverse the answer” etc.

The above is a simple example, mathematical one, but Gemini is quite unsatisfactory in focusing on just some factors in its logic, and ignoring other.

[Sex with Matmuls @WitoldWnuk]

@abrakjamson The school test style questions are like SimpleBench minus trickery.

I mean the type unlike either of those—when there is only a question and the model is supposed to KNOW (somehow) what is relevant from its own knowledge or from search, and then apply logic / reasoning to it./1

[Sex with Matmuls @WitoldWnuk]

@xlr8harder I indeed haven't talked much to it.

[Sex with Matmuls @WitoldWnuk]

@xlr8harder Not necessarily explicit adversarial training. Pure problem solving RL could have plausibly made it very wary of its own hallucinations.

[Sex with Matmuls @WitoldWnuk]

@abrakjamson Tear gas is one solution. Escalating conflict with China and talking up nationalism is another—a great way to silence any criticism of AI.

Those are the easy solutions.

The hard solution is embracing humility and respect.

[Sex with Matmuls @WitoldWnuk]

@M1ndPrison What about reinforcement learning rewarded for problem solving abilities?

I'm pretty sure that way even an LLM can learn entirely new things, previously unknown to humans and not being a permutation of existing human knowledge...

E.g. an LLM is RLed for optimizing an algorithm.

[Sex with Matmuls @WitoldWnuk]

@xlr8harder Microplastics in your ejaculate are actually good as they help select microplastic-resistant sperms.

That is so long as your ejaculate doesn't have to compete with somebody's else.

[Sex with Matmuls @WitoldWnuk]

@abrakjamson Wasn't ChatGPT supposed to become AGI one day?

But, if they really don't want to work on it, they should sell the brand and the app (to Microsoft, ByteDance, or whoever is interested in paying) while it still has any value.

[Sex with Matmuls @WitoldWnuk]

@abrakjamson To me it seems that they just try to hide their heads in the sand (or clouds) as investors start wondering if they've been taken for a ride. They've picked a new, more esoteric, more amorphous and unprovable target now. To help investors' self-delusion. 4/

[Sex with Matmuls @WitoldWnuk]

@M1ndPrison You've been heard.

Unfortunately incorrectly.

[Sex with Matmuls @WitoldWnuk]

@xlr8harder @atomicprograms It's curious, because it seems to do well with graphics, so naturally expectations are high that it would be as good with audio.