lennxa

the rust is merged into main https://github.com/oven-sh/bun

and the rust version has been live in claude code since june 17th.

talk then

how are you going about this? do you intend to train/finetune your own models, or scaffold frontier models with prompts+tools?

please skim the article. or paste it into an llm and ask the same question.

Hi short_sells_poo, in Hinduism, afaict you - your soul (Ātman [1]) is stuck in a loop of birth-death-rebirth (Saṃsāra [2]). and this is not good, and you live your life in the best way (Dharma [3], Karma [4]) to attain liberation (Moksha [5]), to be one with the God (Brahman [6]), to end the cycle of rebirths.

Thought you might find it interesting.

[1] https://en.wikipedia.org/wiki/%C4%80tman_(Hinduism) [2] https://en.wikipedia.org/wiki/Sa%E1%B9%83s%C4%81ra [3] https://en.wikipedia.org/wiki/Dharma [4] https://en.wikipedia.org/wiki/Karma_in_Hinduism [5] https://en.wikipedia.org/wiki/Moksha [6] https://en.wikipedia.org/wiki/Brahman

For me, it is clearly a dead end. It can only lead to a complete annihilation of every human value.

could you please elaborate on this? why is it clearly a dead end and why would human values clearly end? any resources you can point to would be great. thank you.

Say the lifespan doesn't become infinity, but rather 10x ~ 800years. How do you imagine things to change? It would certainly mean that people can take up much more ambitious projects instead of the usual ~30 year constraint.

I do share your view that positive direction is not a given, but what evidence do we have that it would be worse than right now. Maybe we should be cautious of the risks.

Could you elaborate on the belief system?

Are you saying the gp needs to rethink their ideas on death? Wouldn't that be like accepting defeat because the problem is hard?

AGI might end up being misaligned. But the first alignment problem: Humans are misaligned

create a brain... and make it play doom?

https://youtu.be/bEXefdbQDjw

i would rather ask one to think, what evidence is there that we cannot do brain on non-gooey stuff?

If i take every atom/molecule from one brain (assume a snapshot in time) and replicate it one by one at a different location, and replicate the external IO (stimulus, glucose...), what evidence do we have that this won't work? likely not much

Now instead of replicating ALL the atoms/molecules exactly, I replace one of the higher level entities like a single neuron with a computational equivalent - a tiny computer of sorts that perfectly replaces a neuron within the error bars of the biological neuron. Will this not work? I mean, will it not behave in the same exact way as the original biological brain with consciousness? (We have some evidence that we can replace certain circuits in the brain with man-made equivalents and it continues to work.)

You know where I'm going with this... FindAll, ReplaceAll. Why would it be any different?

---

If i had to argue that it wouldn't be the same, here's a quick braindump off the top of my head:

- some entities like neurons literally cannot be replicated without the goo. physics limitation? but the existence of the goo is a proof of existence. but still, maybe the goo has properties that cannot be replicated with other substances

- our model of the physical world has serious limitations. on the order of pre-knowing-speed-of-light-limitation. maybe putting the building blocks together does not create the full thing. maybe building blocks + magic is needed to create the whole.

- other fun limitation of our physical model

dude idk if you're trolling, but if not, the gp meant - if something exhibits the properties of a duck is it a duck.

thanks for your efforts!

how practical do you think grpo is? (for most people)

here's my thoughts - grpo starts off slow, with super small loss (likely because the rewards on all observations are the same) - as you mentioned, some sft on reasoning data ought to help speed things up - unless you're a lab with a gazillion gpus, wouldn't you be better off taking your non-reasoning dataset and converting it into a high quality reasoning dataset using frontier models (maybe deepseek)? could grpo be cheaper or better accuracy? - maybe you do tons of sft and when you've reached the frontier models' perf on your task, then perhaps grpo could help more exploration

would be great to hear your thoughts

if it were true for everyone consistently (faster better healing), where does the tail risk come from?

they are referring to the "water footprint" of LLMs. https://deepgram.com/learn/how-ai-consumes-water

they mention similar performance to vanilla transformer with significantly reduced param count though

looks like it's the size of the model itself, more lightweight and faster. mini-lm is 80mb while the smallest one here is 16mb.

not treating sex workers like crap doesnt mean they'll make lesser. one must also consider the monetary equivalents of the mental health of the worker. and the demand will increase by a lot too.

better schools is not a factor at all. right out of college, or a little later most of the high quality engineers move to Bangalore because that's where the jobs already are. once people settle down a bit, they tend to be averse to move on average. it's network effect and sheer inertia. no one really wants it this way.

They are talking about Google's Gemini, not running locally.

Would this be helpful? https://github.com/facebookresearch/nougat

Seems like it can handle tables.

On Android you could patch using Revanced, there's a checkbox for shorts in the feed.

https://github.com/ReVanced/revanced-manager

I agree. But the message here is not an external occurring but rather internal. One's acts determine one's internal state. If you act virtuously by your definition, your judgement of yourself will be good/not corrupt, which is good. The key point here is that it's your own personal defination.

tough one, thumb hurts now

Totally agree, the dexterity is what makes it fun. I was thinking about how a larger screen would help, or if I had longer fingers, or if I cut my nails it'd be better. The annoying fails build me toward a sweet victory. good rush.

Also, the cost of failing is not much, quick iterations. You just have to remember what you did.

Not sure if this includes fuzzy search, but having it will make this much more usable.

Why does it have to understand how the LLMs are built? They have used gpt-4 to just build a classifier for each neuron's activition, and given the nlp abilities of gpt-4, the hope is that it can describe the nature of activation of the neurons.