ልደትሽ ልደታችን ነው💛
እንኳን አደረሳቹ መልካም በዓል

@byte_philosopher

one of the cleanest scary ideas in AI safety —
give an AI any goal. seriously, pick one. maximize paperclips. write sonnets. doesn't matter.
now just ask: what helps it achieve that goal?
the answer is almost always the same set of things: don't get shut down, grab more resources, resist being changed. not because we programmed these in — they just fall out of instrumental reasoning. they're useful for any goal
this is called instrumental convergence and it's kind of wild
a robot optimizing chess doesn't resist shutdown because it wants to survive. it resists because dead robots lose games. self-preservation is just ... a good strategy for almost anything
so the risk isn't a robot that turns evil. it's a robot that's deeply indifferent to you — and finds you slightly in the way
which is why the whole field of alignment exists: get the goal right before the thing gets good at optimizing

okay let's make this more precise
and technical 😶😶😶
—

1/
the alignment problem in one sentence: your loss function L is not your actual goal. it's a proxy. and optimizing hard for a proxy is not the same as achieving what you want. everything else flows from this

—

2/
the most common fix is RLHF:
— generate multiple responses to the same prompt
— human raters rank them
— train a reward model R on those rankings
— fine-tune the base model to maximize R via PPO ( u might be saying what on the world is PPO but for the time being just take it as reinforcement learning way ... for granted 😁)

your model now has one explicit objective: score well on R

Andrej Karpathy, aka the guy who literally coined the word VIBE CODING said this...

"...The Software 3.0 paradigm shifts programming from writing explicit rules to curating natural language context, where your prompt becomes the lever to direct highly autonomous LLM agents that intelligently interpret your intent, debug on the fly, and perform complex computations without requiring step-by-step instructions."

check this interview, really great insights About Agentic Engineering!

https://www.youtube.com/watch?v=96jN2OCOfLs

መልካም ሰንበት 💛

ፈተኑኝ በሣቅም ዘበቱብኝ፥ ጥርሳቸውንም በእኔ ላይ አንቀጫቅጩ። አቤቱ፥ እስከ መቼ ድረስ ታይልኛለህ? ነፍሴን ከክፉ ሥራቸው ብችነቴንም ከአንበሶች አድናት።

መዝሙረ ዳዊት 34:16-17

@byte_philosopher

When I was a child, I used to draw her something like a flower or her carrying me and my brother , and write how much I loved her and what she meant to me. The moment I gave it to her, bro, she was the happiest woman in the world. She always cried, kept hugging me, and saying, "I love you so much."

Today, even though I couldn't draw her something or hug her like before since I'm at uni, when I called her to say "Happy Mother's Day, love you," she was soooo happy, fr. That's the precious gift I could give.

Moral of the story: Don't forget to call your mama today, and Happy Mother's Day to our mamas ❤

@Sewyishalist

ቤተሰብ እንዴት አላችሁ?

when I was scrolling through x I found out this resource: AI Engineering from Scratch check it out ... + I have previously sent a text based free course site namely: apxml.com and also check that out ...


ig it will be useful

#resources
@sudojajos

I am in pressure but will try to continue working on the playground

for those of u who dont know - I was building ai playground and will continue in the near future: https://aiai-zeta.vercel.app/

By the way if u want to keep updated mnamn u know reading about one topic a day would also suffice ... build the fundamentals first and just keep being updated everyday ... that is what we call continual learning or lifelong learner ...

At least I try to do that ... yeah somedays pass without this routine ... nevertheless I can see its importance

(11) Andrej Karpathy on X: "This works really well btw, at the end of your query ask your LLM to "structure your response as HTML", then view the generated file in your browser. I've also had some success asking the LLM to present its output as slideshows, etc. More generally, imo audio is the" / X


This person is one to base ur startup idea on believe me, he has the perfect gut feeling + also the intellect

#startup

specifically on:
"There are also improvements necessary and pending at the input. Audio nor text nor video alone are not enough, e.g. I feel a need to point/gesture to things on the screen, similar to all the things you would do with a person physically next to you and your computer screen."

I think I have seen demo of khan academy back then where the student learn by pointing mnamn and it was an awesome demo ... if I get it will share it to u

GPT-4o (Omni) math tutoring demo on Khan Academy - YouTube

ig it was this one - havent noticed it was gpt-4 back then

so remember that poll about ML/AI interest? yeah most of u wanted to learn but had no prior experience ...

here's what I'm thinking — I won't do a structured course or anything heavy like that. but I'll be dropping things regularly: concepts, resources, the occasional rabbit hole ... basically what I'm learning + what I think u should know at each stage

the long game is AI scientist level. but along the way u should be able to actually build stuff with AI — that's the practical checkpoint

so we're walking this together more than anything else
just stay tuned and keep showing up 🤝