Видео

"Another Prompt" jaja

i hate him so much man. he is a total clown. and yeah that vocal is beyond pathetic.

lol

As someone who researches AI constantly for years, this perspective baffles me. I implore you - please - research the OpenAI o3 model. The smartest AI ever made. And how quickly China is catching up to the US

I want him to put up half a trillion, towards this, that's why I voted for him.

Can't cure a disease when CEOs need another yacht to go and visit a particular island.

weird tech oligarchy vibes with these announcements, then there is the OpenAI Economic Blueprint if you've read that... a bunch of us vs. china and we will control everything. Not saying this is sam... this for me was just a funny awkward moment where Trump catches Sam off guard with an out-of-domain question.

For confusion! Lol... no i just have limited screen space and i am using OBS to make a scene out of it

I made it to work with any openai compatible endpoint so should work fine with ollama, see the git

@@natecodesai OK. Thanks! Will give it a look.

Cool!

Yes, exactly! 😄♥ So yeah, the actual errant tokens, I have an idea for that: It means including "typing" on the variables so that the interpreter knows the best way to clean it: (ie: number, word, sentence, paragraph, etc.). For now, the best way to try to get rid of those errant tokens is through iterating on the prompt's instructions and the token count you expect from that variable. Currently there's a bug when I try to turn the max tokens down to 1, so fixing that error would make it an easy fix in code, just {yn_var(1, 0.0)} would do it, with temp at 0.0 and max tokens at 1. Now that the code is public, I need to start some issues in the git! (I use vLLM out the gate cause that is what we use in production) Also, on the idea of training, yes, this is a major thing that I was thinking when building this out. Now that Large Concept Models are a thing, I think PromptOwl could really help at creating very large amounts of specific and correct synthetic training data. At that point, if you pointed it in the direction of self-coding prowl scripts, you could get it to create declarative concept tokens which, iirc, could create a sort of internal template representation for schematizing it's own thought latents and then resolving them all within latent space... could be some trippy stuff to play with and might lead to more interesting capabilities on a concept model.

@@natecodesai I can imagine cases where simple typing may fail (e.g. an enum). This may require access to the logprobs instead. There's also a question of conditional prompting, where you may want to change the followup text depending on the response (e.g. if-else on yn_var). Exactly! The specific and correct data is what I was referring to. I'm am skeptical as to whether this would be useful to having a model prompt itself though (I suspect this will result in error accumulation). But simply getting a consistent template with logical reasoning to work could greatly improve this functionality within a model. As for the LCMs, perhaps you could extend the prowl language to mark "concept" regions. Essentially a group of tokens that the model should collapse into a single "thought". For example "is the text logically sound?" may fit into a single token (task token?). This makes me wonder if something similar to textual-inversion for diffusion models could work here, where prowl could be used to mark the conceptual meaning for training such an embedding.

yes sure, you on discord? Also, I will make a video about installation and use.

Yes please! 🙏🏼

Oh yeah, feel the intelligence! Ultra Mega Super Mechadon intelligence! oh, it's coming... 2030... I've seen it in the lab. The world isn't ready! 🤯🤯🤯

😆😆😍

Thank you! I totally am too, lol, just not toward crypto AI, simply because while it sounds cool those two concepts together, imho, you're adding a lot of complexity to solve a problem that doesn't exist.

@@natecodesai I agree. The whole crypto bros ruin it tho.

There are both typescript and javascript libs that will let you call openai endpoints, or ones at local sources, where you're hosting local models. check out my full api integration in raw js for an agent that controls the DOM at github.com/lks-ai/ibis open source, and all in one index.html file.

@@natecodesai that’s great stuff man, I’ll check it out

@@natecodesai so if I integrate this the agent will control the dom ? Also in the tutorial u made the agent with python so how’s that different,im a JavaScript merchant lol

@@dmh5220 On this agent I haven't put it up yet, if you want to see a js agent of mine that controls the dom, it's in raw js/html .. github.com/lks-ai/ibis ... it uses OpenAI compatible endpoints, it's open source, and you can modify what you want. all in one index.html file.

nope, this is a project for using LLMs to create new LLM based agents with their own sets of tools, etc. Meta released a bunch of bots onto fb, but yeah, what I'm talking about in this vid is more for using any model you want for this.

never has a video needed editing more.

yeah, essentially

Rightly my name on this video should be NateSeniorDevsAI, but whatevs, it's 2025 and I've been coding 30 years, lol.

@@natecodesai Got you beat. And I helped out at Harmonai for a year, if we are comparing provenance. Regardless, Gippity did the mental heavy lifting here. Feel like a Oobleck rehash. In other news your explanations were spot on, just soooo winding. No insult intended. Stick with orig comment

this comment is nuts 😂

I think that language models, as part of a larger system can achieve wonderful things. If you look at the current corporate definition, it's already AGI (cause it's whatever they say it is). Seriously though, I think they are getting diminishing returns for the scaling law, and that it is about as close as it can be (to the concept of AGI). There is something more energy efficient and elegant around the corner that will make "scaled LLMs" look like that behemoth prediction machine "Rehoboam" from Season 3 of Westworld... compared to Dolores. Which one is AGI? They are both obviously some form of intelligence, but one used a small country's worth of power to run, and part of a larger scheme. The robots in westworld represent the final and original concept of what robots were imagined as 100 years ago. Essentially, right now all of that stays science fiction while we mess around and experiment with the hidden properties of language and next-token-prediction.

I think its not about reaching that outer horizon, its about being able to see it and go for it. The horizon is infinitely far away. LLMs need outer structures to see beyond the known, just like we all do, but they are a fixed frozen structure ... so they never pull anything into their "known" space and it never grows beyond it's limits. We, (I won't even say as general intelligence, because I also don't really know what that means in a way that vibes with everyone), take journeys from the center in every which direction toward the unknown, and expand our knowns, integrating them into essentially a data fractal (one change affects our model of related things). Also, hahaha, I think with the word intelligence right now we fall into the idea that an entire group of things is intelligent and others are not. It's just another overfit box to try to explain the unknown in a mental model, but it's not at all accurate, clear, nor precise.

thanks for the feedback. you did state it more clearly.

This is a sort of deeper exploration into that: ruclips.net/video/7-BZQ1UQ5qQ/видео.html ... in poetry

My friend read it like that too! hahaha, no clickbait.

Excellent question

The fuzzy thing is where it gets funky. There is a point of failure but it is blurred between what you said and the inability of spoken languages to effectively convey precise meaning, even when it comes to spatiotemporal vocabulary it is very hard to describe a complex specific command with 0 ambiguity in just one phrase and with no back and forth clarification. Its not so much their ability to translate intent but our ability to exactly express intent using natural language. Just look at how many words we had to write here and how there is still the possibility that i didnt fully understand the intent or meaning behind your message.

Yes. The pattern is a mix of fractal equations, perlin noise and a kaleidpscope effect. I was very specific in my prompt. I will find the convo and share it

Yeah personally i am finding already a wealth of knowledge. Also, if you look at alpha fold by deep mind, its already figured out things we couldnt have in a century of testing. As a programmer i find neural networks are exactly that. If i cant figure out hos to write a function for a hard problem, i just get a bunch of input and output samples and have it learn the function. Its like if you want to make gold from charcoal, you would just gather the gold and charcoal as training data and tell your (alchemist?) AI to figure it out.

Until they can self reflect and store ideas that work for further training. Aka Strawberry and later versions.

@@OWENROTHLERNER AI just does pattern matching. It cannot do original thought.

I mean, what do we do? Is not every word we type into a sentence a set of symbolic patterns that just remixes phrases we already know? If you mix james brown samples with fatboy slim have you not created an original mix? If the scaling law holds (which looks to be the case) we can assume that parroting will become originality just like toddlers.

That is not an original thought. No offense. I'll put it another way... how many original thoughts do we have every day on average? Personally i dont think anything im saying here is original, as im not the only person who thinks this way. Im just the one who is saying it

Most coding is not original, most coders are doing the same things at different companies. We waste lots of time thinking through loops or statements. I'm worried about my job security honestly.