data-science-and-ml

[toxic pilot]

Oh nvm lmao. What was it?

You don't need to persist it in the context window necessarily right?

Like I imagine you could provide the model with some functions, such as write_to_db or read_from_db, and have the Todo list retrieved istead

Ah smart

Hmm tru

How'd you do it?

Interesting

I was thinking maybe maintaining several databases 🤔

Maybe running ollama or smth locally?

Well not locally obviously but like hosting your own model

Oh yeah and it's fantastic at code generation

I use it to write all my latex these days 💀

Were u the person I talked to about Jax/Flax?

So maybe this isn't relevant to you but I've started using it

It's actually so good

Super speedy, and the syntax is a lot cleaner

Unfortunately I've had to implement some of the loss functions myself

But it's not too bad ngl

Like I had to implement Huber loss a few days ago

[broken seal]

hello guys, i am new to the server, i want to start with machine learning, can anyone help me with a good resource to start with.

[nocturne saffron]

how to use conda for data science and machine learning? (sorry I am a beginner)

[serene scaffold]

don't use conda. if you're following a tutorial saying that you need to use conda for data science, it's outdated.

[nocturne saffron]

ouh

and

what i have to use?

[serene scaffold]

just regular python.

[nocturne saffron]

ok thank you

i just use conda for ml

[serene scaffold]

don't do that.

[nocturne saffron]

what i have to use for ml?

They say that conda is ideal for machine learning and data science.

i just watch and i i belive it😅

[serene scaffold]

only outdated tutorials say this. it hasn't been necessary to use conda for machine learning for several years.

I have never used conda, and I've been doing ML since about 2018.

[nocturne saffron]

ohh

do you use regular python?

[serene scaffold]

yes.

[nocturne saffron]

ok thank you @serene scaffold

you are very helpful

[bold patrol]

hello guys i am kinda new to python and data science i kinda learned the basics of the python i still have some problem with the classes , lists and to add more extentions to the codes but what should i do and how can i learn better and i wanna become a data scientist if someone helps me i would be appriciated

[nocturne saffron]

you have to learn Descriptive statistics (mean, median, mode, standard deviation, distribution)Basic Probability (Chance, Normal Distribution, Binomial)Linear Algebra (vectors, matrices, basic operations)Basic Calculus (derivatives, gradients — for ML later).
and the library NumPy - numerical operations and arraysPandas - read data, clean, transformMatplotlib and Seaborn - data visualization

oh yea

@serene scaffold can i add you?

[serene scaffold]

I don't usually accept friend requests from server members. I answer questions here when I'm able.

[nocturne saffron]

oh

ok

[bold patrol]

i know every thing till here "... for ML later)."

[nocturne saffron]

do you know the library?

oh yea

i forgot to say scikit-learn

[jaunty helm]

i still have some problem with the classes , lists ...
then probably try to grasp those concepts first before trying more complex stuff

[short barn]

Hello

How and with what language can I run ai models on CPU efficiently?

[serene scaffold]

what kind of AI model do you have in mind?

there's no way to run an LLM on a CPU efficiently no matter what.

[sharp crow]

Had 2 columns , salary and log_salary , and I was passing log_salary directly into training set

This regression is driving me mad

No matter what I do I can't get R2 score above 50

[short barn]

Random Forest classifier

1MB approximately

[sharp crow]

Lads I have a doubt my regression model is pretty bad so I am thinking of converting it into classification, is this logical?

Or I am just dumb

[serene scaffold]

you can just run that normally and no special work is required to make it efficient

[sharp crow]

[toxic pilot]

ahhh so data leaking?

thats a big oof lmao

[sharp crow]

[short barn]

So Python is ok for it?

[sharp crow]

I was blind

[short barn]

I've heard that pandas is slow (and seen benchmarks)

[serene scaffold]

yes. python is the main language for AI.

[short barn]

And there's a high load expected

[short barn]

For developing

I need to run it with zero overhead

[serene scaffold]

there's no such thing as zero overhead.

[short barn]

Ok, with close to zero

Something that would run on a potato without major issues

[serene scaffold]

why?

what's the larger context here?

[sharp crow]

[short barn]

Cause cloud GPU is too expensive

[serene scaffold]

if a model depends on a GPU to be fast, that has nothing to do with Python. it can't be fast without a GPU in any language.

and random forrests don't need a GPU.

[short barn]

Yes, that's what I'm talking about

But I want minimal overhead

And probably want to avoid slow pandas

[serene scaffold]

idk where you're getting the idea that pandas is slow, but in either case, pandas is faster with the pyarrow backend than with the numpy one.

[toxic pilot]

altenatively xgboost

supposed to be faster i think?

but i feel like most of these tools are already optimized

[serene scaffold]

xgboost is a different algorithm than random forrest.

[toxic pilot]

oh is it?

i thought it was a gpu implementation of random forrest

mb

[short barn]

Benchmarks

Pandas spends 30 seconds on what polar does in 2 seconds

[short barn]

I've tried it, but sadly the metrics weren't good enough

[serene scaffold]

That must be an extreme example with many GB of data

[short barn]

Yeah

I expect somewhat high load

Isn't there something like model runner or so

[jaunty helm]

xgboost (extreme gradient boosting), as the name implies, is a gradient boosting machine; you can see more in sklearn here
and you can use a gpu with xgboost but you don't have to
both RFs and GBMs use trees as base learners so it's easy to confuse

and the other 2 popular libraries that you see around - lightgbm, catboost - are also GBMs and not RFs

[devout quartz]

hey guys

im a fresher as a cs major

i wanna choose data science in ai as my main goal

is it a bad option ,as i hear it from diff people about it negative sides

please give ur feedbacks

[coral pumice]

guys i have passion in ai where do i start learning

[sharp crow]

No, ds is one of the best domain you can choose

What negative sides are you taking about tho?

[broken seal]

it is growing very rapidly, but in my experience, it's not simple to break into data science role as a fresher without having masters or phd in ds or minimum 3 years experience is required if you are applying for job

[quartz turret]

ANYONE DOING JOB?

[serene scaffold]

yes

[quartz turret]

What's your role?

[serene scaffold]

computational linguist

[quartz turret]

i'm hearing it for first time , can you explain what's that?

[serene scaffold]

AI for applications that involve language

[quartz turret]

in the field of ds is front end dev is neccesary to learn,what's your pov

[serene scaffold]

no

[quartz turret]

what should i cover in my initial 3 months in my ds jounrny

[quartz turret]

??

[serene scaffold]

How to explore data

[quartz turret]

how can i do it ? i only know python and its some libraries

[quartz turret]

like num,pan and mat

[spring field]

The domain of AI is soooo broad, so many things you could do
Here's an idea I think would be pretty cool (though I'm slightly biased towards making games in a way, lol)

You could make a strategy game and integrate RULER into it, making the enemy AIs learn using RL, but with a more attractive way of crafting the reward functions
You could record player's actions, then use an LLM to summarize those actions, then use an LLM to devise a strategy to be fed into RULER and then make it learn based on that, then you can set up RAG to store previous player strategies as well or something
I'm sure something can be made in that direction, for example

but like, sooo many options out there

[runic parcel]

I need to wrap a custom function (like an API call) inside a PyTorch nn.Module. Is it hard to do?

[serene scaffold]

why are you wanting to do an API call in a torch module?

[worn atlas]

hey guys, im pretty new to all this. Im currently a rising sophmore in highschool and im wondering what college majors should i look into if im trying to become an ML engineer.

[serene scaffold]

a torch module is like, a layer of a network. I can't think of a situation where it would entail an API call. and it would make it take way way longer to train.

[serene scaffold]

you should probably plan to get a masters in CS.

[worn atlas]

would that be best over data science or engineering?

[serene scaffold]

in theory, there's a difference between ML engineering, data science, in data engineering.
in practice, you do whatever your job involves, and it doesn't fall neatly into any of those three.

[worn atlas]

ah thank you

[dry raft]

Alright, so I'm working on GNNs for molecular representation, and I found two main types, imo: GNNs that learn off 2d data, and GNNs that work on 3D data. Which should I use for my project? Should I use both or stick to one? (My project is on the understanding on organic molecules with LLMs, for context)

[worthy oasis]

U should get CS the reason almost every concept used on ML is learned on a CS career is really useful on ML I'm currently working on some ML stuff on my job and some AI concepts that i learned on College really help's to understant ML also College is not all u should read from books, tutorials and also very fucking important LEARN HOW TO READ DOCUMENTATION that really helps in work usually on school everybody just type on Google "Wtf is numpy" and just skip the official documentation, it could be a little overwhelming but the the sooner you do it, the better for you and Good Luck and wishing success bro !

and dont forget : "Understand very well the basics almost every thing in programming are just basic concepts on Complex problems "

[worn atlas]

Thanks man I appreciate the guidance

[quaint mulch]

I don't think those are two categories of GNN

generally, most GNN shouldn't care of the data is 2d or 3d

[odd meteor]

Math, Computer Science, or Statistics major will set you up for the journey ahead properly.

[smoky summit]

heyyy,does anybody have a github student verified account ???
I really need it

[grand minnow]

Why don't you enroll yourself into Uni or College?

[devout quartz]

so i need to have atleast 3 year experience as any post in a company to get into ML in AI

[smoky summit]

umm my uni is done few months ago, so I wont be able to get it

[grand minnow]

Re enroll ig

[smoky summit]

no I have to be in a uni in order to get a student account, do you have one??

[broken seal]

No, not any post it should be data scientist role , ml engineer role etc.

[runic parcel]

its the need

requirement

[serene scaffold]

okay, but what does the API do?

[modest vigil]

you should collect the API data before hand, then go through the data with torch (If you can)

[fresh sluice]

Thanks for this, thats a great idea and also in my niche

AI is really a vast field but thats the thing about it , you go deep in some areas which in future may not be that much beneficial and thus things can go south for you.....so i was also looking for a person who is in this industry so that they can guide me on which path shall i choose

I have some time so if the path is new , i will on it ....i just dont wanna be left out

[pseudo condor]

Is there any webistes that help with learning about neural networks?

[dry raft]

YouTube is a treasure trove of knowledge

3Blue1Brown videos are very intuitive

So I'd start there

[solar thistle]

Is anyone knowledgeable with GRU and transformer conceptually? I (this sounds stupid but it was good for getting the toes wet) have been working on exposing myself more to at least basic ML concepts, so I tried my best to implement an ML solution that can identify single word palindromes. I initially tried an LSTM, then I moved to working on GRU, and more recently have a GRU transformer hybrid, and I learned (probably nothing impressive) a bit about how to think about and abstract the basic ways to structure your dataset to improve a model of that kind. I’ve gotten it to a point to where the current model is 99.88% accurate on a data set of 800 randomly selected as well as generated inputs.

But now I’m trying to further understand how to identify what the model is doing over the course of its training, and I’m having a hard time making sense of the heat maps, I’m not sure if I’m generating them right or if I just don’t understand how to read them. Truthfully, I’ve done googling, YouTubing but also have used LLMs to try to get a better understanding and I’m still falling short.

I feel like I get how to identify when a model is overfitting and when it’s reached its peak for identifying over training. But I want to better understanding the process of what it’s doing, and how I can look at the training data to understand what correlations it’s drawing during the training.

If anyone’s got some insight they’d share, welcome to PM me if this isn’t seen until some later time

[calm cipher]

Hm a couple of things, first I guess there's nothing wrong with training a model to identify palindromes but it can be done very simply with a handwritten algorithm so machine learning is massive overkill

you're just experimenting so it is ok but it isn't something you would typically do with machine learning

I would expect massive overfitting with this problem, a dataset with 800 examples is extremely small compared to the size and complexity of the model

I'm curious where your 99.88% accuracy is coming from - when you're evaluating model performance, is it on a train/test/validation split of the data, or are you training and evaluating with the same data?

and finally I wouldn't expect looking at heat maps in general to tell you much, neural networks are famously black boxes and you won't learn much about what they're doing by peeking inside, unless you have a very specific problem where the model's attention directly corresponds with some explainable aspect of the problem

your model is overfitting if it achieves near-perfect results when evaluated on training data, but very poor results on test data that wasn't used for training

[serene scaffold]

@solar thistle are you tokenizing at the letter level? Otherwise, the model is certainly overfitting.

[quaint mulch]

Check the pinned message in this channel

[quaint mulch]

You can start by sharing us your plots (with proper description so we can understand it)

[quaint mulch]

Besides heatmap, you can also collect the model input and output pairs along training to see any patterns.

And in general, this is still a very active area of research.

[solar thistle]

Oh yeah I know that its an overkill solution for an already easily solved problem. But thats one of the reasons I picked it. Its an easy problem to model out and the possible ways to solve a palindrome traditionally is like, theres only 1 and its really easy to do lol. But That gives me an advantage because its not a problem that I have any difficulty understanding to the full degree how to solve it. I figured because of that the ML model that would solve it would for 1, be simple enough that I can use it as a self guided intro to ML, but also there wouldnt be any "magic" to how it works. That hasnt entirely really been the case though. Since this is my first time really trying to understand how ML works, I stumbled a lot to get to where I was. I had heard some of the terms before (LSTM, GRU, RNN etc etc) but wasnt sure how you identify a problem and which solution is most fit to be used to apply to it)

Really I feel like I learned more about the importance of well structured and valid data that represents the problem youre trying to solve lol.

Training intially wasnt that great, but I learned (for this specific application) things I hadnt really thought of before. Cuz youre right, its an easily solvable tradtional CS problem, but when I was working on the GRU and the LSTM both I noticed things about palindromes id never considered would be kind of important facets of what they are. For example the LSTM version I did would often misclass palindroms that were near-palindromes, consistently. Like wowowiw as an example often was mis-classed

so that lead me to generate data that specifically would expose the model to large amounts of data that included those kinds of near-palindrome edge cases, and improved the model success substantially. I think the orignal GRU model I started with would also fail a lot when you had the first 2 or 3 letters symetrically match the last 2 or 3, and it would basically ignore the middle section. Also got stumped when words had more than 2 repeated letters, no matrter where they appeared

Obviously what Im using ML for isnt impressive, but I dont expect to fully become ML capable, I just want to be able to better understand how they work and just be somewhat competent about recognizing how they work/what and how useful data is structred etc

[solar thistle]

I mean, yeah I think so, (forgive my ignorance on the proper terminology) but you mean like encoding each letter into an array of ints representing each letter right?

this is how im encoding

def preprocess(data, maxlen=MAXLEN):
    alphabet = list(string.ascii_lowercase)
    char_to_index = {c: i+1 for i, c in enumerate(alphabet)}

    def encode(word):
        return [char_to_index.get(c, 0) for c in word]

    X = [encode(word) for word, _ in data]
    y = [label for _, label in data]
    X = pad_sequences(X, maxlen=maxlen, padding='post')
    y = np.array(y)
    return X, y```

so X ends up as a numpy array (num_samples, maxlen) and y ends up being the binary representation identifier of if the word is a valid or invalid palindrome

I only use words of len 12 so pad the delta of word.len and 12, and then use the padding to normalize all words to the same array len

Heres an example of the attention heatmap for both classifications

I was reading that the heatmap represents the y axis is the query, so responsible for computing attention, and the x axis is the listener thats attended to or whatever, which sounds kinda fine, I guess the query token basically uses the grid to represent how much "attention" was paid on the listener

conceptually that sounds fine lol. But like. So why then in the first graph did each encoded letter position pay basically what I udnerstand to be like "max" attention, all to the same letter

wouldnt, if how Im thinking of that and said it, was correct, wouldnt you expect to see a distribution of max heat like this? - or rather, where you would expect the most "attention" to have been attatched

thats kinda why Im asking for help, cuz im not sure if im just not reading the graph correctly, or if the graph itself isnt being generated correctly

And I know the model works well, so its not like its just randomly outputting some invalid data

The data is a mix of predefined english palindromes and non palindromes, and randmomly generated just palindromes/nonpalindromes from random characters

[calm cipher]

I am not 100% sure about this but I suspect attention isn't right mathematically for palindrome detection

Maybe if you try adding some positional encoding it could work, but I think the principle of how attention works isn't going to help

If you're interested in interpreting how neural networks work, have you studied the xor problem?

Also here's a good blog post where someone is tracking the cell state of a RNN on different language tasks, this might be along the lines of what you're trying to do https://karpathy.github.io/2015/05/21/rnn-effectiveness/

[sharp crow]

Lads I want to make some good projects. Any recommendation?

[mossy pond]

Pi as function of angle of every number(0=0°, 1=36°, ... 9=324°) every number is one step forward in the angle of the number.
first 50000 second 200000 with the window of 50000

[analog leaf]

Hey, I hope that this is the right channel for my issue. I need to read some txt files with spectra Data. The thing is that both Excel and Origin can't really import the data. It just come out wrong. So I thought about doing it in Python. I got some old lecture skripts from a friend that goes in that direction but I don't really understand them. I either would like some help writing skript, which I can read all the files and have it build an Excel file so I can insert it into Origin properly, or recommendations for resources where I can do it myselve (preferably in a acceptable amount of time). I tried several AI Tools but the produced Excel Files all have major problems.

All the files look like above (I don't know if it acceptable to uploade one here, I would need to change some meta data in order to not post personal info). And I need an Excel file that looks somewhat like the following:

Mass (m/z) Value (counts)
100.3 281083.092750

can somebody help me?

Please also fell free to ping me

[cursive totem]

hey guys, i want to learn ml and now i am at a stage of learning pure python. Is asyncio worth spending time to learn or should i skip this step? i learned about coroutines and it says that it is useful in asynchronous programming, so i wonder if i even need it

and is algorithms knowledge (like in leetcode) needed for it? i will make graph neural network (some physics applied) as my diploma thesis, at this point i still didnt look up what graph neural network is (and neither my supervisor lol), so i wonder if i need to know algorithms like BFS or DFS

[neon owl]

Yes

[serene scaffold]

This "data format" doesn't appear to be intended for use in programs. You'll have to break it up into sensible units. Like one CSV for the mass and value rows at the bottom of the screenshot.

[serene scaffold]

asyncio is not important for ML

Neither are most leetcode questions. You should probably know the main sorting algorithms and understand their asymptotic analysis. Same for graph traversal algorithms.

"BFS and DFS" you need to understand these, and if you can't code them, you don't understand them

[serene scaffold]

See if any will let you be their research assistant

I think this is the most questions I've answered in my three-stop train ride

[analog leaf]

It's a txt file

[serene scaffold]

Yes

[calm cipher]

It looks like 3 separate tab-separated tabular files combined into one with headers separating them

[analog leaf]

It's two in that case, there are also other spectra with 2

[calm cipher]

If you manually separate them, pandas should be able to open them as is

Granted you'll probably have to do more cleanup, but at least you'll have it in memory in a format where that's possible

If you can't manually separate them you'll have to try to do it in code, which could be easy or hard depending on how much variation there is in the structure

Actually hmm the first two sections look like key value pairs

[analog leaf]

I tried to do something with ai assistantence. The main Problem are commas and that some numbers don't get recognized in Excel as numbers

[calm cipher]

Only the third section is tabular data

[calm cipher]

I don't think AI assistance would work well here, worst case it could misrepresent your numbers depending on how it attempts to break up the files

[analog leaf]

Oh you mean that with section. I only need the third in this case

[analog leaf]

It's better than me. But Manuel check ups revealed mistakes that are not allowed to happen

[calm cipher]

I mean it could completely make up numbers that were never there in originally

If you still want to use AI, try something like this in your prompt

You want to write a program that will read the contents of a text file into memory, but will ignore all lines that occur before "Raw Data:"

Once you have it in memory as a string, you can load it into Pandas

[analog leaf]

I don't know enough python to do it all on my own

I will try that tomorrow or the day after chatgpt told me that I hit the limit for data analysis

[calm cipher]

Yeah give it a shot

If you're intending any or part of this process to involve python scripts, even with chatgpt, you're going to need to know some python

[analog leaf]

I know enough so I can read most of what I encountered. Just the writing process is the problem

And I am working on refreshing some stuff with lecture notes and books from my Uni library

[cursive totem]

Thx for advice

[fallow coyote]

Atm, getting a bit bored with ai/ml. I feel like the projects I’m doing are not piquing my interest. I’ll still go through with them because I need the practice and experience in applying my knowledge but, I’d like suggestions in interesting machine learning projects

[solar thistle]

Thats kinda like, the only thing that makes sense right? Like. I said I wasnt sure If i wasnt reading it right, or if it wasnt being generated right, but maybe thats just not the way to conceptualize the model performace, I didnt really think about that so thank you so much for the input. And no I havnt heard of that but I will do some reasearch and hopefully that points me in a better direction that Im going now, thank you very much for taking the time to read and respond to me!

[calm cipher]

So the reason I suspect it is challenging is because dot product attention effectively uses the similarity between the key and query vectors to determine the score

And because determining palindromes requires considering position, that means the RNN has to produce a sequence of vectors where, to use the outer two characters as an example, the first vector is most similar to the final vector

Or at least it has to produce vectors that can be transformed into a key space and query space where that is true

But I also think that is probably very hard for a RNN to do because it goes into the input without knowing how long it is

I suspect it might be possible to make it work with a very deep multilayered bidirectional RNN, and it also might be possible that it is doing something unrelated to position that still generalizes, but at any rate it is probably just overfitting and memorizing the training data

[unkempt thorn]

https://www.reddit.com/r/LocalLLM/comments/1m3u23l/need_help_in_fixing_my_qwen25vl7b_ocr_script/

any help would be amazing

[nocturne whale]

We Just Build an AI Agent without a Big A$$ Prompt

Last year we tried to bring an LLM “agent” into a real enterprise workflow. It looked easy in the demo videos. In production it was… chaos.

• Tiny wording tweaks = totally different behaviour
• Impossible to unit-test; every run was a new adventure
• One mega-prompt meant one engineer could break the whole thing
• SOC-2 reviewers hated the “no traceability” story

We wanted the predictability of a backend service and the flexibility of an LLM. So we built NOMOS: a step-based state-machine engine that wraps any LLM (OpenAI, Claude, local). Each state is explicit, testable, and independently ownable.

NOMOS supports lots of llm providers including OpenAI, MistralAI, Groq, Gemini/Gemma, OpenRouter, Anthropic, Ollama and Cohere. and there is lots of functionality already there and more are coming everyday.

Open-source core (MIT)
• GitHub: https://github.com/dowhiledev/nomos
• Documentation: https://nomos.dowhile.dev/

Looking ahead: we’re also prototyping Kosmos, a “Vercel for AI agents” that can deploy NOMOS or other frameworks behind a single control plane. If that sounds useful, Join the waitlist.
https://nomos.dowhile.dev/kosmos

Would love war stories from anyone who’s wrestled with flaky prompt agents. What hurt the most?.

[nocturne whale]

yes, you can use nomos validate --config ... to check the validity

also can generate the schema easily using nomos schema and use it with your yaml

We introduced steps and transition between each steps. transitions are fully controlled by different routes and conditions. If the conditions are not met those routes will be not taken. Also we have introduced lots self healing techniques such as as soon the llm makes an mistake we will constrain the options it have dynamically, so next time it tries back it has fewer options.

[calm cipher]

This is an ad for a service, isn't it? <@&831776746206265384>

[south quest]

ehhh

it's OSS

but it's edging on advertisements

@nocturne whale showcasing projects is okay here, but not anything which has any sort of paid offering, just as a reference

i think what you're posting is okay, apart from the fact you're posting it in multiple channels. showcase OSS work but don't do it in the form of length walls of text, it's not appropriate for the space and violates the rules.

[calm cipher]

There's a waitlist for a control panel service

[south quest]

talking about OSS software is fine -- promoting paid offerings for upgraded versions of OSS suites is not okay and over-promoting the corporate side of OSS is not appropriate for this space

[hollow pagoda]

Do u do reinforcement learning

Those look more fun especially the game labs

More challenging aswell

[exotic star]

is why machines learn a good start for learning the math

or is khan academy better?

im already familiar with python and im currently learning pandas and numpy tho i was told here that learning the math is crucial for learning the libraries for data science

[fallow coyote]

Dont want to get into reinforcement learning until im comfortable i know more of the maths but i might give it a go anyway

[calm cipher]

there are some simple reinforcement learning problem formulations that might help you learn the math as you learn the RL

maybe try studying something like multi-armed bandits to start with

it's usually one of the earliest things you'd study in reinforcement learning anyway

[small wedge]

while a lot of RL algorithms require a lot of math like qlearning and gradient policy stuff, I find things like evolution strategies and genetic algorithms to be a very easy way to jump into some fun RL projects without needing much math. Or at least the math involved is very simple and intuitive; what it means to preform crossover and mutation is completely up to you as the dev.

[weak seal]

can anyone here proof chain a sigmoid activation function all the way to an equilateral triangle?

[neon owl]

I just an undergraduate

[sand herald]

i was still thinking about learning RL vs LLMs/agents... im personally more interested in RL than LLMs, but im not sure about the career opportunities

as im kinda mid-level already, I think it is time for me to specialize

and almost all jobs are asking for LLM experience, ugh

my current skillset is more towards ML/DL/optimization

[vocal zealot]

An LLM got gold in the IMO, something many thought would take years.

and it wasnt even allowed to use the internet, or external tools, and had the same time as the other contestants

PURE NATURAL LANGUAGE.

that is seriously impressive

It seems test time compute is indeed extremely effective. Of course, along with other algorithmic breakthroughs

[sand herald]

yeah i saw that news but uh... i dont know. seems to excel in advanced areas but struggle with other areas which are deemed more basic?

anyway, it seems like in industry, usually the applications of LLM are kinda... boring? like chatbots and shit

[vocal zealot]

Yeah but it's an unreleased model. The model capabilities can vary drastically between the models.

For example, gpt 4o fails at elementary math, while something like o3 gets like a 98% on AIME

[hoary jay]

hey, grade 11 student here. im lowk interested in learning abt data science and ai and i wanna be able to land a small internship next summer before going into uni for cs. any tips on how to get started?

[vocal zealot]

oh and also, what's even more crazy is that this was a GENERAL REASONING model. This was NOT A fine tuned model.

That is absolutely wild

unlike googles model

that was specifically made for this

this is fundementally different

[sand herald]

does anyone here understand LLMs on a deep level? i've been struggling to see it past a "next word guesstimator"

[sand herald]

i see, that's impressive

[vocal zealot]

A simple ""next word guesstimator" would absolutely NOT be getting scores like this

It uses lots of test time compute and some other breakthroughs which open ai has mentioned

It can think for many hours, that's the difference between a model like this, and say... gpt 4o, that responds almost instantly

[sand herald]

oh, so i guess it's very different than a vanilla Transformers then?

i have some basic understanding of Transformers as I'm going to need that for my next project. my "understanding" of LLMs is from that

i dont follow the LLM space/progress closely

[vocal zealot]

another important thing is this. Long horizon tasks, which LLMs have struggled with in the past. It seems really good progress is being made there.

[sand herald]

right. any idea how it's being done differently? is it a totally new model architecture or are they just adding on extra stuff to the core (which I assume to be still Transformers)?

[vocal zealot]

More test time compute, and other experimental things they are doing it seems. What that "other" is, im not sure.

[sand herald]

learn python? and how to handle basic files like excel/csv and do some data analysis stuff before jumping into the basic ML stuff. maybe look at statquest YT videos to learn and kaggle titanic dataset for a start.

[sand herald]

test time compute... hmm. wonder how it works.

because at the end of the day, it's all numbers isn't it? are they letting the LLM loop more...? against different parts of the pre-trained dataset? ugh

yeah

i know MLPs, LSTMs on a quite deep level

and basics for transformers, CVAEs

because to me, these models are learning a set of weights, which are just numbers, from data. i think that'd be good enough to guess the next word, and the next. or even a sentence or paragraph if given the right architecture and data. but to say that they can "understand meaning" and "reason", that's a bit of a stretch to me.

but it could be because i don't understand the leap from Transformers -> the current LLMs

i've seen some of the basic ML/DL models achieve surprising things though

[vocal zealot]

Quick reminder: Don’t end like this idiot. He was proven wrong a day later.

[calm cipher]

Until they publish their results and methodology in a way that is reproducible by other researchers, this is just marketing

[vocal zealot]

These all (obviously) Open Ai researchers.

[vocal zealot]

Yeah… we kinda already know the methodology…

If you think they are literally just lying like this… then you are beyond saving… you do you I guess

Yeah that will probably happen, as it literally has with every single Open Ai release…

But the model will be released likely by end of year

The most insane part is it’s a GENERAL reasoning 💀

This is not like Google, who creates their models specially for this.

THIS is what AGI is about

GENERAL intelligence. Google deepmind results are cool…. but far less impressive

What?

Open Ai said it passed the bar and then MIT said it didn’t?

It does easily pass the bar now btw lol

🤦‍♂️

I’m sure they are all lying bro.

But we’ll see

This insane distrust people have for literally no reason is so insane. Especially with a company like Open Ai which has a pretty good track record believe it or not

Every single open ai model ever released in the history of Open ai has had a paper along with its release.

This model obviously won’t have one as it’s not coming out yet.

[calm cipher]

They don't, they are extremely closed off and routinely hype themselves up and actively try to scare people

They are very closed off to the peer review process

[vocal zealot]

But I guess we’ll see. Im sure in 3 months it will come out that it was ALL A BIG LIE PERPETUATED BY EVERY SINGLE EMPLOYEE, LIKE THE LAST TIME THAT OPEN AI… did some imaginary major lie like this which apparently I’ve never heard of

It’s ok. We’ll see. For now, GPT 5 is incoming, which is exciting.

Anyways…

https://tenor.com/view/jim-carey-good-afternoon-good-evening-goodnight-i-dont-see-you-gif-15609107

Ima go eat pizza now

that movie was fucking awesome btw. ok im leaving now

[sand herald]

uhh i also view these news and benchmarks with distrust

to me, openAI needs investors and they tend to hype things up

kinda distrust sam altman as a person too

remember a few months ago when they were hyping AGI

and how close we were to AGI

actually, what is the deal about getting gold for math olympia?

what i mean is, how differently are the math olympia questions structured differently from ... i don't know, typical questions that the general public ask an LLM?

multiple layers of reasoning?

it's funny they mention that the model has no access to the internet or tools? i kinda thought it almost has the entire internet as its training data, isn't it?

I see

that's kinda strange though

i'd imagine

i'd think that math equations would be closer to code than say... natural language

for LLMs to perform better in natural language than math, it's interesting

yeah on that, has there been any research on why LLMs hallucinate?

yeah i've read something along this line as well. perhaps human-like reasoning is flawed, full of gaps and we tend to "hallucinate" too?

it could also be that the training data has much more natural language than code?

or that the underlying architecture, LSTMs and Transformers, seemed to be designed more for natural langauge rather than code or math

do you have any resources to quickly understand the leap from Transformers to the first-generation of GPTs?

i roughly understand how transformers work, but the Generative & Pre-trained part of it, I don't

haven't tried looking at it yet 😛

oh wow

this almost simulates how i add numbers quickly

but anyway, interesting, so there are multiple pathways

i guess that same set of weights has to be used across all functions/domains

not just simple addition/subtraction

oh yeah, i was about to suggest something similar, to have steps/layers at the start to figure the task

they probably already have it though

this is probably a simplified diagram

im kinda wondering about this part as well

do they use the same set of weights for all tasks? or different sets of weights for different tasks? or different PARTS of the weights for different tasks (i.e. can we activate/deactivate some of the weights depending on the task, as i'd imagine the other parts of the weight to represent the "reality" that is not specific to the task)?

ah this is making my brain hurt

hahahaha

😮

maybe i should go into LLM research

and get 1% of that sweet sweet 200 mil package

ahh this is too advanced for me. i don't have the background for it

hahahaha

i see

openAI is closed-sourced, how about the rest like claude?

right

"Kimi K2 is a state-of-the-art mixture-of-experts (MoE) language model with 32 billion activated parameters and 1 trillion total parameters." so 32B params is considered small?

in the world of LLMs

yeah prolly, I can run 1B params on 4 L40s

🤣

params is referring to the number of weights/ the scalar values that are in the matrices, right?

hmm that doesn't sound like a lot

i dont know, unless 1 param takes different amount of compute depending on if you're running a MLP vs LSTM vs Transformers, etc.

my own models had i think... uh

1GB of params

which was about 1B params if i remember correctly

which is why im a little surprised at the 50-100B number hmm

anyways

yeah

[thick heron]

Hey, I’ve been building a local AI assistant in Python — voice + text input, mood engine, memory with ChromaDB, and an LLM running via Ollama. It runs offline and uses a personality system that shifts tone based on mood vectors (sort of like a 12-band EQ). Also uses local TTS.

The problem is I’ve hit serious hardware limits (6GB VRAM). I’ve already optimized it to load models on-demand and split behavior vs task logic, but beyond this, the system can’t handle more development or testing. I’ve tried to keep everything modular and as light as possible, but even basic scaling breaks things.

At this point, I’m mostly just finishing up documentation. Not sure what to do next — cloud is not really an option. Has anyone here worked on something similar or found any hacks/tricks to keep these kinds of projects alive locally on limited hardware?

Appreciate any ideas. Not trying to show off, just genuinely stuck.

[thick heron]

Mood engine is a combination of different things the input is processed in 3 layers and you can find full documentation here https://github.com/SilverShadowHeart/silver_heart

[tidal bough]

I'd imagine it could have some steps to figure out what the question is asking, a step to encode the number and operations, but then why not just take advantage of the fact that it's running on a computer and that the network is just processing a bunch of vectors already ?
I don't think that's possible - remember that there's stuff like activation functions on each layer

the fact that everything is highly nonlinear harms it here - it has to find a combination of nonlinear operations that approximates multiplication

So it can't, just, say, have a subnetwork that turns a written number into a single scalar that it can then do operations on, because a single scalar can't represent a number - it'll instantly get truncated by the activation function

[tidal bough]

If you have a fairly new GPU, maybe experiment with vllm rather than ollama; you might be able to squeeze out some extra performance that way

(vllm is more hardware-demanding, e.g. it refuses to work on my Pascal-architecture GPU, but it's significantly faster and I suspect you can configure it for low RAM usage, too)

[tidal bough]

Kimi was already mentioned, but the technique is far older than it - Mixtral is what comes to my mind, which is a Mixture of Experts (MoE) model

oh, and the new Gemma models, too. It's popular for local reasoning because it lets you pack more capabilities into a limited amount of VRAM (or even RAM).

[jaunty helm]

could also try running a quantized version and/or a weaker model

[tidal bough]

Well, sure, but tt doesn't instinctively seem to me that there's simpler ones then the one that paper finds?

[thick heron]

Thanks a lot

[thick heron]

That would ruin the entire project's aim i tried

[tidal bough]

ollama is already using heavily quantized models, doesn't it?

it always downloads the Q4_M gguf (IIRC) unless you specify an exact huggingface file

[jaunty helm]

why would it "ruin the entire project"??
like I'm actually curious

[jaunty helm]

idk they were using ollama tbh, just saw they posted a link to the repo tho
and last time I took a look at ollama it was still using q4_0 or q4_1 as the default, idk if that has changed

[thick heron]

The model loses power and becomes too dumb and makes no sense and gives out rubbish in simple terms

[tidal bough]

ah, I think you're right

[thick heron]

Yes it does the model is q4 quantized

[jaunty helm]

well I mean that's the tradeoff you get for trying to fit more into the limited vram
anyway, quickly scanning through your readme, you're using this model:

nous-hermes2:10.7b-solar-q4_K_M
which is 2 years old at this point, you should definitely switch to something else

[thick heron]

I tried phi 3 tiny lama

[jaunty helm]

some models of similar size:

• llama 3.x 8b
• gemma 2 9b
• gemma 3 12b
• mistral nemo 12b
• qwen 3 14b

[thick heron]

They are too verbose or 1800 scholars

[thick heron]

Ohhh I will try these

[jaunty helm]

I haven't heard of "phi 3 tiny llama," but I know that the phi series is more meant for single turn reasoning / instruction following
they are extremely dumb when it comes to human knowledge
so maybe not what you want

[thick heron]

Yes i learnt it after using them
I wanted a model which is capable of good modern english and does some level of reasoning

[thick heron]

Phi3 is one model tiny lama was another model 😅 i forgot comma

[tidal bough]

maybe also see phi4-reasoning:plus, somehow I missed its release entirely but it's 14B and dominates benchmarks... though depending on the GPU it might be hard to fit into 6GB

[jaunty helm]

ah alright, so by tiny llama you mean this one?
then yeah I can definitely see it being dumb
personally I wouldn't go anywhere below 2b if you want to have a somewhat coherent chat without wanting to punch the model for being too stupid, and even then that's stretching it a bit

[thick heron]

Yes

[thick heron]

It is but I will do something about it and push this project forward

[tidal bough]

Your project is quite interesting to me because I also have only 6GB VRAM (and an old GPU that doesn't support a lot of important features like bfloat16), and I mostly concluded that all the LLMs I can run locally are too dumb to be useful, even as assistants

[jaunty helm]

there is offloading to cpu but it tanks inference speed obviously

[thick heron]

I have an rtx 3050 in a laptop 💻 overheating is a thing i should be careful with

[tidal bough]

Isn't that already part of the training process for most small models?

well, I guess they weren't finetuned on this exact environment, so it might help

[thick heron]

It was a multipurpose project

An assistant has to be balanced in all regions

[jaunty helm]

well but for someone to finetune a model you need quite a bit more compute than you need to run it
finetuning on 6gb of vram sounds sketch

[thick heron]

I tried a multi-model style. It's too much latency

Like brain swapping

Trying to make this fully local with colab u get high resources I get carried away puts me off the goal of making this possible

In any laptop with a gpu and 6gb vram

[jaunty helm]

well what they mean is, since finetuning is way more hardware intensive than inference, you can borrow the good hardware in colab to finetune a model to better suit your task
then the inferencing wouldn't take any more vram, but since it's finetuned the model will (ideally) perform better

[thick heron]

Ohh okayyy i misunderstood sorry

I will try that

I don't know much about fine tuning where do i learn more about it ?

Final question thanks for all the support guys

I will try youtube if I get any problems I will comeeee back 🤧 tysm for helping

[cursive totem]

Guys, is jupyter plugin in pycharm better choice than pure jupyter? I mean in pycharm it has much better tools, even fancy looking array display and i have pycharm pro cuz i am currently a student

[acoustic barn]

Hi all, what laptop would you recommend for starting with a data science and ai bachelor’s degree?

[serene scaffold]

Any laptop that isn't a Chromebook.

[serene scaffold]

That's a matter of personal preference. Try both and see how you feel

[cursive totem]

I tried google colab before because i needed it for university tasks to present some data visualization and i wasn't comfortable with it, but i like pycharm features overall so i really like this jupyter plugin so i can still use pycharm and its project management

So i guess i will stivk with it

[oak coyote]

I want to make an Project about the skills i learnt and how to integrate all things
I have understanding about - Python, numpy, pandas, streamlit, sql, matplot and powerBI

I am new at discord, finding peoples which can help...

[proven pier]

Is optical character recognition still practical tech?

[proven pier]

Pretty interesting. Thanks for giving me the rundown

[oak coyote]

I am just a beginner

And held on tutorials

But wanted to make some projects but not able to do so

I want to learn ML

[acoustic barn]

Thanks! I also have a macbook, but unfortunately I require a laptop with a minimum of nvidia 3070, so I need to buy a windows laptop

[oak coyote]

And somebody asked me to firstly learn about Data Analytics

[acoustic barn]

But I’ll just wait for the first lessons and ask a teacher if it’s really necessary

Yeah it’s probably for the machine learning course

[oak coyote]

and now i got stuck not able to find projects

[proven pier]

If you need a graphics card for AI, I have heard it is usually just better to rent a server and use that instead

[acoustic barn]

Ohhh that’s sick

[proven pier]

Yeah, that's the most practical thing to do...

[acoustic barn]

How does that work? (I’m sorry I’m very new to this)

[proven pier]

A server is just another computer

So you have a crappy cheap laptop, that you use to SSH into (remotely log in) another computer (the beefy server) that runs all of your costly computations on

[acoustic barn]

It is the same idea of logging into a raspberry pi from my macbook?

[proven pier]

Yes they are both computers

[acoustic barn]

Ahhhh I see

[proven pier]

Do you use SSH to get into your raspberry pi

[acoustic barn]

Yes

[proven pier]

Yes exactly the same

But instead of spending 2-3 thousand dollars on a laptop

You might spend 100 and then rent a server for like 10-20 bucks a month

[acoustic barn]

Where would I be able to do that?

[oak coyote]

Ohkay

[proven pier]

I'm not in the market, so I will literally just use my search engine and see what prices I see

[acoustic barn]

What should I search for?

[proven pier]

https://www.gpu-mart.com/pricing
Here's an example. There's a lot of companies providing this stuff
you gotta learn how to use a search engine 😆

[acoustic barn]

Also, does it work without flaws?

[proven pier]

Just look up VPS, gpu's, machine learning, do research from there

Dont straight up buy, read about what people have written

Not just about specific services, but the entire process

[acoustic barn]

Yess I see, thank you for your explanation 😊, I’m going to look into it

[proven pier]

Then you can weigh it against owning the GPU machine yourself

Idk how many semesters you're in class for. But if you can find the GPU in a laptop for $2k, divide the VPS monthly rental price by that and see how many months it would take to cost more

If rental fee is $50 a month, it would take 40 months of renting to cost $2k, or over 3 years

[acoustic barn]

The degree is 4 years

[proven pier]

You wont be in class during the summer

and maybe a month off during winter

[acoustic barn]

True

Oh no only 2 weeks ):

[proven pier]

Okay so what, 9 months of classes?

9*4 = 36 months

[acoustic barn]

I suppose yes

[proven pier]

$2k / 36 months = $55 a month

[acoustic barn]

Hmmm

[proven pier]

So if you get something beefier than that, it would maybe be better to buy a laptop at the $2k. Once again, you would need to weigh the option of what the laptop provides

Your laptop will also be running other graphical things when you're using it, so you probably wont have access to the entire GPU resources to program

[acoustic barn]

Is there a laptop you would recommend? If I were to go for a new laptop

[proven pier]

No I have no idea

[acoustic barn]

Ahhh okay thank you again for your replies ur amazing

[proven pier]

Just looking at a VPS, here's the rental price for 2 years (monthly rental price), and here's JUST THE GPU BY ITSELF

So you could buy the GPU at $2700 and own it, or you could rent it for 2 years which will cost $312

@acoustic barn ^ just things to consider. So once again, just look at hardware and VPS providers and see if things line up and how cost efficient it is to go one way or the other

It almost seems ridiculous, 18 years of renting until it costs more to rent? Yeah, I would just do a bit more research. I'm just saying, what I've heard is it's better to rent servers for this type of thing

[acoustic barn]

Thank you for sharing! Definitely gonna look into it

[proven pier]

Follow up - I have ZERO idea why it's showing.. indian? currency in the gpu price on the right? It thinks I'm in mumbai? 🤔
Whatever link I used did amazon.in which is india, I'm guessing

I was thinking what the hell is that currency symbol

[proven pier]

Sorry for spamming, I have learned the price of this specific gpu is $80. a much better price 😂 I would probably just purchase that

[glass carbon]

Yeah what about a recent gpu such as rtx 4080 or sth?

[acoustic barn]

Ohhhhh hahahaha I was so confused I thought it was dollars, so I was thinking “if it’s so expensive it’s best to buy a laptop 😂”

[acoustic barn]

Ohhh no don’t apologise I appreciate the information

[proven pier]

Yeah looks like after 6-9 months it's better to own. At that point, probably best to own a GPU somewhere and use it - if you're going to need it for 4 years.
I suppose the whole "rent dont buy" is for people who want to spin up a new application for their company during 1 month or something - train their models, then use it on "normal" computers

[glass carbon]

Ask older students what they use @acoustic barn

[proven pier]

Maybe you could have a cheaper gpu for personal use/more frequent lower end class projects, then when big projects come through you rent one of these bad boys

[acoustic barn]

Ohhh that’s also a good one

[proven pier]

Yeah if you could get ahold of a senior or junior and talk to them about how workloads are

[acoustic barn]

Yeah this will be the second year that this program exists here xd

I will ask someone in year two when uni starts

[proven pier]

Will they not lie about hallucination

If it already hallucinates wont it believe its information is true

Only when I directly "confront" it by saying it's wrong will it say so. And even then, it's because they want to agree so bad

Maybe some sort of roleplaying like "you are under oath, and perjury is a penalty that is enforced with jailtime" or some crazy shit 😂

idk how practical it is to test hallucinations. If hallucinating scenarios are even reproducible. I feel like a lot of LLM results aren't reproducible. Or maybe I'm ignorant, can you provide static seeds to LLM applications so they always respond the same to input?

That would be the only way I could see any sort of real test driven solutions bearing fruit. But then again, it would form an extreme bias towards that one seed, so 🤷‍♂️

what are you designing here

human memories are just in yaml format?

https://tenor.com/view/clueless-aware-twitch-forsen-emote-gif-25354609

I'm joking, I was just referencing that image I responded to

That's fun. I haven't gotten deep into AI, but I have wanted to because I have some of my own opinions on the human brain. It's cool how you're tying your conception of memory together in the process of information flow

Just gotta program the DNA that seeds the whole process, then it can be a reproducible learning agent with access to actuators and sensors 😂 maybe try to give a large reward mechanism towards friendly sociable, non psychotic selfish behavior 😂

LLM's are just trained models that predict statistically likely responses to queries. I still keep an open mind that a lot of tech people are skeptical of AGI possibilities. The brain exists in the physical universe, I dont see why it can't be recreated outside of our biological context

I think the human brain tries to put everything it interacts with into a sort of mental state. And we are always predicting where we expect those states to be. And as we interact with the world, it adjusts how our inner state representation looks and behaves. And as we further interact with the environment, our prediction of entities improves. You see a car driving down the street, you expect it to continue and it does. When it crashes, that is surprising. You know it can happen. Sometimes it's surprising. However, sometimes you see somebody driving erratically and you sort of expect an accident. It happens, still surprising, but not as much

You also expect your kitchen to be in a "state" of configuration. You leave your room, or enter your house, and are thirsty, so you move towards the state that should satisfy your quest

If it's a natural or instinctive decision, I can see it coming first. But some decisions you truly have to contemplate over

Then the act of deciding, it sort of doesn't matter which "happens" first. I would presume the "conscious" awareness must lag a bit

You can only be aware of something once it exists

Awareness is probably just your own ability to measure the state of your mind. Well, your mind reaches that state before it can be measured

Upon being aware, you are in a state of review, and you still have the option of "changing your mind"

You've trained your brain how to make decisions

It's been trained over a lifetime of you making decisions, right or wrong, then reviewing them

Even if you make a decision, you can review it and adjust it or change it completely

All of which requires the awareness and review process

You could choose to throw yourself in front of a bus, but you are not passive. It's always an option but your awareness is saying it's probably not a good one

Maybe sometimes you want to quit your job, but you still decide against it. Not passive at all

Yeah I mean we have a brain stem that breathes for us, or makes the heart pump. There needs to be an underlying order that we don't consciously focus on or we would get nothing done

If we had to be consciously aware of the prediction mathematics and calculus that goes into the process at all times it would be overwhelming. However, you can review and report on how you predict things using your words. It would be extremely verbose to lay it out in detail in realtime 24/7, but you can get quite introspective

When you say evolution, you mean neural network architectures predisposed from dna? Because babies still have to "learn" and train their models through experience

There's certainly some architectures already setup from dna. And initial seeding of its "training set" potentially which is why most babies act the same

I suppose that would come from evolution pressures, yes

Maybe no initial seeded training set, just randomized. Crying and all that is just lack of emotional control (training)

[dull radish]

Hey guys, I'm looking for a way to parse a pdf file into a format like:
{
"title": "This is the title",
"outline": [
{ "level": "H1", "text": "Introduction", "page": 1 },
{ "level": "H2", "text": "Main content", "page": 2 },
{ "level": "H3", "text": "Conclusion", "page": 3 }
]
}
pymupdf has been my primary basis to extract the content from the pdf but any idea, what can be done further from that, simple heuristics extraction doesn't always work for all the different types of pdfs

P.s apologies for disturbing the existing convo

[proven pier]

#data-science-and-ml message
Spoke somewhat about this earlier. But also you can look into the python module pdfplumber which can parse pdfs

[dull radish]

Okay that ocr model unfortunately sits quite a bit out of my size constraint and yes pdfplumber does extract the content of the pdf, but accurately classifying what are titles and what are headings has still been a challenging task

Especially when u take into account more complex pdfs

[topaz sail]

Hello so I am new in coding and I wanna learn data science and I wanna get the basics in math first where can I learn? Any recommendations

[sour blaze]

Given the mostly unstructured nature of the PDF format, everything would likely need to be done through heuristics. This could be things like detecting headers from font usage or the document outline or tables from the way items are grouped. Fortunately here, there are multiple existing tools for this, primarily for use with LLMs but perhaps useful here, such as docling, markitdown, and PyMuPDF4LLM to name a few.

Depending on where the document was sourced though, it's possible that it already includes similar information in the form of logical structure. Unfortunately, I'm not sure if anything supports a high-level interface for it though most PDF processors will let you access the StructTreeRoot key where this information is stored. So that would probably be a last resort.

[dull radish]

I have explored docling and pymupdf4llm as of now, docling does seem to do a decent job although I have to probably switch to an onnx runtime to fall under <200mb time constraint.

Just tried out pymupdf4llm earlier and it's pretty solid as well but not quite enough on its own, I'll try applying heuristics and see where this goes.

Haven't tried markitdown so will do that and see.

and I'm surprised this would be done through heuristics ideally, I thought a light weight classifier or a visual analysis tool maybe using yolo finetuned on a dataset would be a better approach.

and yeah I wouldn't bank on that, the ones I'm using don't have that yeah.

thanks though I'll definitely give this a shot

[tranquil jasper]

hi
how long would you say, on average, would take for someone to get a grip on computer vision?
someone who knows programming, but nothing about computer vision or ai in general

[tepid bluff]

claude 4.5

[grand minnow]

lol I didn't think I asked too much too often to hit these errors

[raven garden]

Hello guys, hope you are doing well. I have problem I cant solve. I try to deploy my ML model with streamlit. its the first time I do this. I ve been said that the model and other relevent files should be in the same directory than the streamlit app python file, which is what I did. however when I run the code I got the error message ( that I wrote in the code ) : "Model file not found! Please ensure 'los_model_complete.pkl' is in the same directory."

I do not understand whats going on, if someone can guide me on how to solve this I would appreciate it very much, thanks

P.S: I made a mistake in the code screen shot, the file should be los_model_complete.pkl and not los_best_model, I changed it but it still doesnt work

[raven garden]

issue solved, thanks anyways

[sage sparrow]

I'm looking for a realistic e-commerce dataset. Does anyone know any sources outside of Kaggle? My next bet was looking into synthetic data creation with Python functions since it's just for a showcase project

[stable flower]

#python-discussion

[modest badger]

I'll resume the convo: wtf are you talking about

Ai cannot "infinitely recycle energy" like a perpetual motion machine

[naive river]

I'll resume the convo: wtf are you talking about
is a good quote

[modest badger]

haha, well, I am excited to hear the thought

[tepid turtle]

hello, chat. I'm new here and wanted to share my data cleaning/ visualization project. Just looking for some feedback from you, so is it ok if i share a link on the gihub preview and a repo here?

[spring field]

sure

[tepid turtle]

Thanks) My project will be especially interesting for those who're interested in media, politics and journalism things

So I was looking at the Reporters Without Borders data estimates on countries' press freedom index, and thought that if we could see this data on the graph along the years - it could me much more informative than a simple timestamp that they have on their infographics in a form of worldmap with countries on it.

So here I've collected their data from 2002 and until now about the counties ranking in press freedom, their score and different factors (which are very valuable, but initiate only from 2022.

You can play around with it, selecting different countries, years and factors. Again, could be very informative for those who appreciate a truthful journalism.

And here's the link: https://vlad-gby.github.io/rsf_index_visualization/
And a repo with files and a notebook (readme is not ready yet, working on it): https://github.com/vlad-gby/rsf_index_visualization

[stable flower]

thank you

@modest badger the whole point of a PMM perpetual motion machine is to make a way to infinitely recycle and use heat energy to power machines or in general support civilization. the problem with every life form and physical object is its subject to entropy and loses either its structure or energy.

the only way we can avoid death or extinction is by finding a way to reverse or manipulate entropy and bypass the laws of thermodynamics

so far there doesnt seem to be a way to do it

it seems like we all will die with this bright staryy universe

whats even more terrifying than death is a cyclical universe. scientsits have said that our universe will never repeat again and this is likely to be the first lifespan or itereation of the universe. whether that sounds stupid since this is the only known universe as if there were more. well my point is scientists said that we will never be created again through another big bang or big crunch

it still worries me since there is such a possiblitiy for lifeforms on earth to exist once more and billions of animals suffering and being exploited and humans suffer as well

i dont wanna come on earth anymore

[modest badger]

so your angst is not just your own mortality, but the universes?

[torpid mirage]

Me when I'm not isent from transformation and entropy as a physical object within the universe

[serene scaffold]

But how is this related to data science and AI?

[serene scaffold]

But it can do the opposite

[mossy pond]

if you want that for all pdf, yes you should use heuristic ... simple analyze font type/size and bold-flag does only work for 50% of all cases.
in addition you need the position, is it a bit separated from other text-blocks max 10 words or such things you will run in a never ending story ^^
iv tried that for one week ^^

[vast thunder]

What are some good resources to get started with AI and ML? I barely know anything about it. And please a more "practical" approach. I couldn't keep reading the Microsoft AI tutorial on Github just because it was more theoretical than practical

[serene scaffold]

Don't fall into the trap of thinking practical means useful and theory means useless. They're two sides of the same coin

!resources data science

[arctic wedge]

Resources

The Resources page on our website contains a list of hand-selected learning resources that we regularly recommend to both beginners and experts.

[vast thunder]

o thanks

[vast thunder]

I just want the tutorial to be more interactive. I can't handle just reading 5 chapters of something and doing nothing

[serene scaffold]

Yeah, and you probably wouldn't learn anything from that either. Whatever resource you use should be interactive or give you an exercise to do on your own or something

[vast thunder]

mhm yeah that's what I meant by theoretical

[serene scaffold]

That's not really what theory means. See my prior message

[vast thunder]

yeah sorry, bad phrasing. I just wanted something more interactive 😆

[buoyant reef]

what module should i use to make an ai

[serene scaffold]

What are you trying to do

Like what is this AI supposed to do?

[buoyant reef]

i want the ai t be a very. low level catbot

chatbot*

[split olive]

Hello, im trying to make a suspicious activity detection model given as a challenge to me by my prof, my current plan is to make a shoplifting detection model then build upon other activities. Im going to use yolo11-pose.pt model to get the keypoints, label each frame and feed them into an LSTM model so it can predict whether or not there is a shoplift. Is this a good approach, i really want some advice. I havent started coding im only looking for approaches. Any help is appreciated :)

[reef latch]

hola

anyone know how i can get a start into machine learning through python?

[split olive]

learn the working of models like linear regression, logistic regression then implement them in python then work up to more complex algorithms

[dull radish]

yeahh I've been doing that, maybe an ml model might work

[stable flower]

thank you lisan. what a perfect answer you gave me about my issues. i really agree with #3 as well as the rest of what you said. I suppose there are ways to live much longer or possibly live nearly forever. although we may never defeat the laws of thermodynamics which governs us and energy exchange. i guess im just worried we will be trapped to do the same things over and over forever in a deterministic universe. i personally believe in free will but if time really is similar to a flat circle or even a clock i assume it would repeat but like you said thats incredibly far into the future and better not to worry about it

[sonic ravine]

not sure if this is the right place to ask but does anyone have experience with publishing applied maths papers?
and if so what plotting library did you use for graphs
i cant decide between plotly and matplotlib, but im open to other suggestions too

[serene scaffold]

matplotlib is probably more often used for academic papers than plotly, but it doesn't ultimately matter. either one can generate images of the plots.

I used matplotlib when I was in academia, and I hate it.

the plotly API feels less intentionally unintuitive than the matplotlib one.

[sonic ravine]

thats really helpful thanks a lot <3

[carmine nest]

You have this project and the accompanying thesis, which is one of the most advanced in the field (the model is under a non-commercial license). It’s up to you to rewrite and train the final model. https://github.com/TeCSAR-UNCC/PoseLift

[toxic pilot]

well he aint wrongl AI = 0 because thats the amount of value it provides in that specific usecase

[calm cipher]

Seaborn is a wrapper library around Matplotlib that has more complex built in graph types, native DataFrame support, and slightly better visual defaults, I always recommend it if you're going to use Matplotlib

[serene scaffold]

I just use plotly

[calm cipher]

But it doesn't make it less clunky to use unfortunately

I'm going to have to learn plotly sometime

[lapis sequoia]

Catbot

👀👀

[split olive]

Oh that helps a lot, thank you

[lapis sequoia]

Are GANS dead? Like, what’s going on? Pump the latent dim into robots so they can learn quicker and let’s make everything automated. Let’s go!

[iron basalt]

The answer for "why is X not used" is usually either: does not scale (it's not a bunch of copy pasted things you can just increase N on), too hard to train, or does not fit modern hardware (GPUs). GANs are too hard to train.

I guess I can add the fourth case of lack of software ecosystem (tooling) surrounding the idea.

[sour hamlet]

Hey all, I’m Ali from Code Craft — I made a beginner-friendly Linear Regression tutorial in Python. Excited to learn with you

[lapis sequoia]

Stable diffusion is better and GANS take forever. I don’t know they had their day.

[sour hamlet]

👋 Hey everyone!

I just finished making a short and beginner-friendly tutorial on Linear Regression in Python.

✅ You’ll learn how to:

Import and work with a clean dataset (included)

Train and visualize a simple linear regression model

Predict house prices using scikit-learn

🎯 It’s aimed at new ML learners who want a clear, step-by-step walkthrough with no fluff.

📊 Dataset (CSV): https://drive.google.com/file/d/1rZ5OhntQeJ5gA7WtWFAWbntgwznTx10X/view?usp=sharing
▶️ Tutorial (14–15 mins): https://www.youtube.com/watch?v=zBk72AV_weg&t=76s

I’d love any feedback or suggestions on what project I should do next. Thanks!

[iron basalt]

I lot of things in ML can work / are valid options, but not used due to simple practical issues, like being too annoying to work with.

The options are also narrowed down by available hardware (although FPGAs exist so you can put this under "too annoying to work with" too).

[lapis sequoia]

I don’t know, it used to be the big thing. But it’s just two NNs fighting. autoencoders are better by a lot.

[sour hamlet]

Totally agree, but why is it like that

[iron basalt]

Computers that we have these days are a miracle of modern science and international cooperation. It's a miracle that we have any, and as many options as we do. Ideally we would have way more types/options though.

Then on the software side it's driven by open source, which is driven by a few very motivated individuals (working for free) that are just rare, like 1 in a million rare (combination of skill, obsession and resources (free time mostly)).

[sour hamlet]

I know right, I cant even Imagine how they created stuff like this, but do you want to check out my chanel, I would for you to support

I started today, and made my first video

It is about LinearRegression

[next shard]

Yeah

[sour hamlet]

And if you do, suggest what I should teach next

[next shard]

I want tensor flow full tutorial

[sour hamlet]

Ok I will try!

although I am not that good

At tensor

But I will give a beginner

Tutorial

If that is ok

[next shard]

Ok brother

[sour hamlet]

If you want notifications subscribe!

You dont have too

[serene scaffold]

Self-promotion is not allowed

[sour hamlet]

Oh

Sorry

[serene scaffold]

Anyway, I don't recommend learning to use tensorflow. It hasn't been popular for several years

Just use pytorch

[serene dew]

I think i finally found the "book of everything"

[serene dew]

how old are you?

[gilded axle]

https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/ fascinating, really well written summary foe the layperson

[odd stratus]

does anyone know where i can find just a very basic premade ai network and weights to run on my home computer?
nothing too fancy, just something i can run on my cpu

[grand minnow]

https://docs.openwebui.com/

[gentle stone]

Hi Ali, I'm young coder too and Interested with ML, your awesome project interested for me, I want to know more about this project

[peak field]

Can anyone give me some advice on how to detect if a person is looking at their phone? Are there any libraries which could help me achieve that? I found this gazetracking library but its meant for webcams. Ig i could still use it but its not the greatest?

[sour hamlet]

Absolutely, my friend! I'm launching a free YouTube course called "Regressions" — a beginner-friendly series covering the top 5 regression models. I’ll not only break down how each one works but also show you how to choose the right model based on your dataset. Let’s make machine learning simple!

[restive roost]

Hi, anyone know how to get KitLive working with OpenAI? https://discord.com/channels/267624335836053506/1397500697842421862

[serene scaffold]

@sour hamlet your message was removed for advertising

[gentle stone]

Yes bro, hope we can work together after I learn enough about ML, keep share your knowledge

[sour hamlet]

I might do a live stream of an Ai "Hackathon"

[short moth]

anyone know how I can read a pdf table that is formatted like this?

i tried to use camelot but it gives me this

[short moth]

Ok ill put it in ai

yup 🤣

[sour hamlet]

Hey guys\

[next shard]

Hi guys please suggest to me what I should add in my GitHub profile

[short moth]

anyone have recommendation to have a more efficient code here (it works but I don't like creating another temporary dataframe I want it all in one statement):

import pandas as pd
def assign_id(df_,value_list):
    previous_value = ''
    for i in range(len(value_list)):
        if value_list[i] != '':
            previous_value = value_list[i]
        else:
            value_list[i] = previous_value
    return value_list
def tweak_df(df_):
    temp_df = df_.filter(items = [0,2,3,4]).rename(columns = {0:'END_ID', 2:'PART_NUMBER', 3: 'DESCRIPTION', 4: 'QTY'}).drop([0])
    updated_val = assign_id(temp_df, temp_df.END_ID.values)

(im basically trying to replace the values in one column based on the previous values before it since some values are empty

[serene scaffold]

for each value in the END_ID column, you want to replace empty-string values with values from a different series?

[short moth]

from the previous value of the same series

[serene scaffold]

if empty-string values were previously NaNs, and you replaced them with empty strings, do not do that.
replace the empty strings with NaNs and do ffill (which is forward-fill)

[short moth]

so for example if I have:
0 hello world
1
2
3 hello
4

[serene scaffold]

Always represent missing data with NaN and not with anything else.

[short moth]

oh i did not think of tht

my missing data is empty strings

so i replace ' ' with NaN

np.nan

and then use ffill

very good suggestion haha

did not think of it

[serene scaffold]

if the CSV is like ,,3,,, the actual empty strings should be interpreted as NaN
if the csv is like "","",3,"","", you can change the parameters of read_csv to interpret "" as NaN.

[short moth]

im reading pdf's so its different XD

pdf tables

[serene scaffold]

that sucks

[short moth]

its pain but I have no other choice

all the data at the company i work at is all in pdfs

i mean I could convert all the data to csv or excel but it would take me years cuz i would need to do it all by hand

its a weird ass and long ass table to parse

so using camelot helped me to put it in a more readable form

[lapis sequoia]

Have any of you ever used a transformer for time series? If so, was it encoder-decoder ish?

[magic dune]

what type of task you doing

classification

forecasting?

[lapis sequoia]

Forecasting. I just hate using RNNs all of the time.

[magic dune]

that's fair. have you looked into TFT?

[lapis sequoia]

No, this is the first time I considered anything out of RNNs, ARIMA, SARIMA, prophet, or regular ML models. People have had to use transformers for this right? Is TFT good?

[magic dune]

Well the one issue with time series data is that it is usually multivariable for most problems, which can lead to slow training times, but from my experience, TFT does well with forecasting. https://arxiv.org/pdf/1912.09363

@lapis sequoia also it depends on the nature of you data

[lapis sequoia]

Panel data , frequently updated, reliable sources, fairly consistent

[magic dune]

how many columns

[lapis sequoia]

Are you asking this because you want to know if it’s multivariate?

[magic dune]

ya

wanna know if its bivariate uni or multi

[lapis sequoia]

A good amount not over blown

[magic dune]

ok ya TFT should be good but I would start with a subset of dataset and experiment

with other models aswell

are you strictly stuck to a transformer-based model?

[lapis sequoia]

Is TFT anything like a HF transformer?

[magic dune]

if by HF you mean huggingface? than not really.

[lapis sequoia]

Yes, I mean hugging face. I am honestly more comfortable with that than really any RNNs or prophet .

[magic dune]

!pip darts

[arctic wedge]

darts v0.36.0

A python library for easy manipulation and forecasting of time series.

Released on <t:1751200606:D>.

[magic dune]

darts supports the TFT model

so you just got to import it

and you can play around with it

[magic dune]

it is inspired by HF transformer but it has modifications and additional features that make it specialized for its task

[magic dune]

one f the best libs out there is rn for ts forecasting

https://unit8co.github.io/darts/generated_api/darts.models.forecasting.tft_model.html?highlight=tft#module-darts.models.forecasting.tft_model

[lapis sequoia]

Is there any HF transformer that is used a lot for time series? I honestly cannot find much. And thank you.

[magic dune]

not that I have used but I can try and find some rq.

https://huggingface.co/docs/transformers/en/model_doc/time_series_transformer

found this in the docs

[lapis sequoia]

I will try something. It’s not a huge deal just waiting epoch after epoch and then boom, not even as good as an ensemble method. I appreciate your suggestions.

[magic dune]

ensemble methods are crazy slow

[lapis sequoia]

I know

[magic dune]

from my experience

took me 7 days to training on a simple dataset

@lapis sequoia

https://huggingface.co/docs/transformers/en/model_doc/patchtst

this looks promising for HF based

[lapis sequoia]

Thank you. I was about to use T5 for whatever reason

[sour hamlet]

🚀 Just Dropped: Build Your Own AI Chatbot in Python (Part 1)
Hey everyone! 👋
I just released the first video in a free series where I teach how to build an AI chatbot from scratch using Python.
✅ No libraries. No shortcuts. Just pure Python and real learning.

📹 Watch here: https://www.youtube.com/watch?v=2p9hr53iBYY
🤖 In this part, we build the bot’s brain + memory — and it actually learns from you as you chat!

If you’re into:

Python projects

AI / machine learning

Building real-world tools
…then this is for you. Would love feedback or ideas for future parts!

Let’s build smarter tools together. 💬

[restive roost]

Hi, anyone know how to get KitLive working with OpenAI? https://discord.com/channels/267624335836053506/1397500697842421862

Or any suggestion for low cost 'receptionist' AI agent?

[dense tulip]

I'm building a SQL Ai agent, And I'm a bit lost

Any public repository that can bring me some help?

[jaunty helm]

I feel like the sentiment for a good chunk of people still is that complex deep learning models aren't as great as advertised
for one they need a lot of data; while we have that for images and text that might not be the case for your time series
they're also computationally expensive compared to other options when the gain could be small

[fresh sluice]

Its really good but too basic for 2025

[gentle stone]

Hi I'm back, As I said I will learn this course after finishing previous courses. I'm going to take this course. How's your progress?

[cinder onyx]

Hey, i’m a student really into finance and quant stuff, and I’ve been thinking of starting a project in that space (something hedge fund-ish, infra/research focused). Just wondering if anyone else here might be interested in teaming up — could be a cool opportunity to build/research something legit and learn a ton along the way.
Nothing formal, just seeing who’s out there. feel free to dm!

[lapis sequoia]

I really like that you said that. I agree completely. Most people make this way too complicated when simple methods can be used and don’t take forever and are more direct and faster and accurate. Everything in data doesn’t need a Neural Net.

[sour hamlet]

I know that it is too basic my lan was to keep making it better, but in the end of my course I will teach how to make a chatbot that learns over time, more advanced

My friends requested it bbecaue they want ot learn about ai and machine laering, although this was not but my next video will be some ai to it

[toxic pilot]

wait arent transformers by definition designed for timeseries?

[toxic pilot]

well generally you'd use encoder for classification related tasks (i.e. given a time series, identify abnormalities or something idk) and you'd use decoders for autoregressive purposes

encoder + decoder is good for seq2seq

[jaunty helm]

transformers were originally designed for NLP, benchmarked on translation tasks

[magic dune]

nope

[toxic pilot]

yes right> because they were designed for sequence data, which is essentially a superset of time series

[calm cipher]

there have been variants for visual data but I'm not very familiar with them

[buoyant vine]

that is definitely a stretch of a definition imo 😅

[calm cipher]

attention doesn't necessarily imply timeseries but it's often used in timeseries data

I'm using timeseries and sequential interchangeably

actually to drive that point home even further, a lot of language models add positional encoding to the input data, because otherwise the transformer wouldn't know where timesteps exist in the input in relation to one another

whereas you'd get some kind of notion of position in a RNN for free just based on how it works

but the basic abstraction you use to think about transformers are queries, keys, and values, which has nothing to do with position

actually if you want an example of non-sequential attention, check out global style tokens for tacotron

they do but the attention layer would not consider position in weighing each step

positional encoding allows it to do this

right

yes exactly

it's a question of who thinks it's sequential data

without positional encoding, you're interpreting the data as sequential, but the attention layer doesn't care

with positional encoding, the attention layer can treat the data as sequential, or at least account for position when computing attention scores

this is only indirectly related to the attention talk but I thought it was cool

some TTS models perform a small one-dimensional convolution over the input data prior to giving it as input in a RNN

so if you're trying to synthesize "Hello", it has the effect of blurring nearby characters into each other

[magic dune]

they are built for time series as much as CNNs are built for timeseries

[calm cipher]

so instead of considering H, e, l, l, o separately, it would consider H-e, e-l, l-l, l-o together

and I think that's neat

yeah

I think it's loosely related to the n-gram concept from NLP

and it works really well with speech synthesis where a character by itself isn't enough context to know how to pronounce it

[toxic pilot]

This is a meaningless statement

[jaunty helm]

the thing is text data has some very specific properties that don't generalize to all time series, like:

• we have a lot of text
• text heavily correlates with each other on multiple levels
  etc

like I don't think there's even close to a silver bullet type architecture to time series just because how broad that entails
like the position of a pendulum swinging back and forth on a moving cart could be a time series, and so can the population of rabbits and wolves in an area

[toxic pilot]

ahhh that makes sense

[tepid bluff]

yeah

[iron basalt]

It's been shown that if you just scale it up enough and train long enough it basically learns to do positional encoding on its own. Having it helps by it not having to waste a bunch of time and resources on learning this though. Interestingly, the thing it learns resembles grid cells (which positional encoding does too) (hinting that this may be a universal solution to positioning problems that biology has also found).

(RNNs have been shown to learn the same thing too)

[calm cipher]

I wouldn't be surprised, but it would be based on patterns it finds in the data

mathematically any single attention layer has no concept of position or a local space around different timesteps in the input or anything like that

[iron basalt]

To perform well it needs to reinvent space.

(And grid cells it seems)

[lime grove]

does anyone know if cuda 12.9 plays nice with PyTorch?

[calm cipher]

I have 12.9 and torch is running fine for me

[lime grove]

which version torch?

[calm cipher]

whatever the latest is

[lime grove]

what GPU are you using?

[calm cipher]

actually to clarify I have 12.9.1 installed at the system level, but the cuda in my virtual environment is 12.6.4.1, but they are working fine together

it's on a 3090

[lime grove]

your PyTorch is using the 12.6?

[calm cipher]

it's whatever the PyTorch module has listed as a dependency

I wouldn't be surprised if it takes many more than one block

this paper tested it on models with 125 million to 1.3 billion weights

they saw it learning positions within 4 layers

[lime grove]

so you are using 12.9 with Pytorch or 12.6?

[calm cipher]

hm I guess it must be 12.6 but it is not incompatible with a system installation of 12.9

you might want to disregard what I said if you're intending to run it with 12.9, although I don't know if it's compatible with the current stable version

[lime grove]

sounds like you're not sure of anything

[calm cipher]

right, like i said, disregard what I said before

[magic dune]

It is not meaning less both algorithms work with time series data but neither was specifically built for it.

[thorny zealot]

book or online training to learn code ?

[serene scaffold]

!resources

[arctic wedge]

Resources

The Resources page on our website contains a list of hand-selected learning resources that we regularly recommend to both beginners and experts.

[brisk cave]

Hey guys!

My name is Romie and I am an aspiring Data Scientist!
No formal schooling and currently hold a position as a data analyzer!
My company is small and no one has degrees. The data literacy is very limited so I'm traveling through uncharted territory.

I just started Learning Python and I am great need of a mentor or someone to point me in the right direction. My co-workers have no interest in data, so most of my conversations are diluted to fit my audience.

Someone '@ me or message me. I'm hungry to learn.

What should I learn first. Mathematics? Fundamentals of Language? Someone point me.

Thanks!

Romie

[sour hamlet]

🚀 New Video: Smarter Python Chatbot That Understands Typos (Part 2)
Hey everyone! 👋

Just dropped Part 2 of my free YouTube course where we’re building an AI chatbot in Python — this time, it gets smarter.

✅ Now it can understand fuzzy input — like if someone types "helo" instead of "hello," it still gets it right.
This is where the AI magic starts!

📹 Watch here: https://youtu.be/gI4ftsWQSXk
🧠 Built with basic Python + fuzzy matching — no fancy libraries needed.

If you're into AI, Python projects, or building tools that learn, check it out and let me know what you'd like to see in Part 3!

[lime grove]

getting nvidia to install is fucking tedious

you'd imagine that a trillion dollar corp wouldn't be so inept

yeah

to be fair, Microshit is worse

and it has been worse for decades

like, why the fuck do I have to iterate over goddamn visual studio versions while installing nvidia?

slow mode

it has a "feature" called Nsight, and there are versioning issues. Added to that are whatever mysteries that happen in your own machine (GPU versions, which OS, etc)

Anthropic has an excuse. It is barely 4 years old

I just want a local CUDA for PyTorch (another versioning headache)

[lime grove]

done 🫡

had to install the cuda toolkit 12.9, then find out that Pytorch only likes 12.8, so uninstall, restart, redownload 12.8, install & discover that it hanged on the Nsights stuff.

so, I had to then wipe the computer clean of nvidia 12.8, Visual Studio 2022, then reinstall the same fucking Visual Studio version and nvidia version and then it worked

like WTF man

there is an Nvidia app that doesn't seem to work unless you have visual studio installed. The whole thing is bonkers

so, after installing 20 GB of Visual Studio, another 4 GB of nvidia, I can write a 3 line python script! Woohoo!

[wooden sail]

probably the visual studio c++ redistributables?

ah just noticed the message is pretty old

[lime grove]

not sure, tbh

these packages are so sprawling and over extended that it is probably unknowable

[last crag]

Hi does anyone happen to have a prompt script for sending prompts to a LM Studio llm with vectors from a qdrant database thats working? I cannot seem to get mine working after days of trying...

[eternal mist]

Hey from where should I learn numpy library

Thanks mate

[old rivet]

Hey everyone, I’m a Biochemistry grad diving into ML for bioinformatics. Just shared two Jupyter Notebook projects analyzing ALAS1 gene expression (1000-row Bgee RNA-Seq dataset) on GitHub (MIT License):

ALAS1 Expression Analyzer: Normalizes expression & read counts to compute Expression_Combined, with stats and histogram.
ALAS1 Read Count Visualizer: Plots trends over time with linear regression.

🔗 Code: https://github.com/Hi-Script/biochemistry_ml_mini_projects
Why: ALAS1 is key for heme biosynthesis, prepping for ML like LSTM.
Feedback: Ideas for ML models or code improvements?

[lime grove]

@old rivet I don't understand what that trend line is supposed to be trending on

like, is it appropriate to toss out those apparent outliers (like the one around 150 minutes)?

[old rivet]

the outlier might not correspond to an actual time point but rather a data point’s position I think it’s tied to a specific strain or condition, it could be biologically meaningful e.g., a strain with unusually high ALAS1 expression. What do you think?

[lime grove]

you could project the points to the y-axis, and thus build a sort of probability mass function. From there assume some sort of distribution, and remove outliers. But, that is just clumsy, surely they mean something physical?

I am not a biochemist. But from a pure time series perspective, that trendline is somewhat problematic. It doesn't seem to have any meaning

torch success after last night's hair pulling frustrations:

pretty cool that pip3 install ing pytorch brings in mpmath, arbitrary precision floating point computation. Also, sympy

[old rivet]

Thanks thanks for this, an outlier at ~150 minutes, which I kept after IQR analysis, suspecting it’s biologically meaningful (ALAS1 is key in heme biosynthesis). I want to:

1. Improve trend analysis (e.g., handle outliers better).
2. Predict future h.ALAS1_Combined values (time-series forecasting).
3. Classify strains by expression patterns (e.g., high vs. low).

ML Ideas: I’m exploring:

• Robust regression ( Huber) to reduce outlier impact.
• LSTM for forecasting ALAS1_Combined over Time_min.
• K-means clustering to group strains by expression.

Any suggestions for robust ML models or preprocessing? better ways to handle outliers or prep data for LSTM/clustering?

[lime grove]

forecasting time series is probably best done via ARIMA (classical, i.e.) approaches. I am not sure LSTM is interpretable enough. What do you plan using K-means?

But, again, not a biochemist.

[safe agate]

https://www.youtube.com/watch?v=NIBprn5cEZA

[lavish wraith]

Have use sql in data science or mostly use pandas ??

[lime grove]

SQL is mandatory

[barren path]

Bro can any one tell how much math is required for data science

[daring aurora]

I'm 14 rn and I really interesting in coding but I'm so scared and overthinking because of the ai, what do I do T-T

[proven pier]

I have a project in mind that should "leverage" ai, but I feel is more of a systems progarmming problem than anything.

I simply want to detect, through optical character recognition, the XYZ coordinates of a minimap. The minimap is not always on the screen, however, so I will need to also detect when the minimap is available

Will I need to manually take a lot of static images, and label them for where the minimap is and if it is available? What sort of process do I need to follow to have a classification/identification network for this?

[broken stirrup]

hey, i have a task and i dont know how shall i proceed
i have 700+ tables having polution related data, i stored them in bigquery and made excel modules to fetch them.
now i want if user give me a query like
tell me co2 emission of aniak electricity plant , then the source should be searched for in the tables.

i have made a graph rag for that (made metadata of tables and stored in json) and fed it to gemini and asking for response as in "which dataset u think this can be present"

but gemini response is very inaccurate, is there a way to make it accurate or is there any other way for me to get the source info

[proven pier]

Information overload

Dont take in everything at once. You need to learn how to program first

You didn't study calculus or geometry when you started school. You began with the basics, to give your mind context and a framework of logic to reason with. That will aid you in understanding more complex topics in the future

[spring field]

you could just run your OCR regardless of whether the minimap is on the screen and then just parse out what it gives you and see if part of the parsed text contains coordinates, if not, the minimap is not on screen, but at that point you don't really care about that anymore, lol

[broken stirrup]

let me send u the prompt

i think my metadata is abit loose but still im not sure

and i have these categorized based on country -> material -> sector

1.5 flash

uhm but wont that be time taking, as i may have 100s of inputs

but my approach is fine right?

i mean, it cant go more optimal

but how would i know which dataset my source input is

for example if someone say
tell me co2 emission caused because of crops in france

then my llm will see the metadata -> country (france)-> datasets in france -> oh its a crop (agribalyse dataset) -> give output as = your source can be in 'agribalyse'

then i perform normal fuzzy match in this dataset and show user the co2 emission

yes bcz i cant see each source in all tables as i have 700+ im using llm to narrow down it to one or two tables

so if i give wider description of all datasets and use a reasoning model, i can expect better results

is there any better approach or this llm thing sounds good

oh

😭 im short on tokens too or else i could have used multiple prompts

[errant bison]

Gemini?

[broken stirrup]

im not so worried about the data as im using openly available ones but still ill take care of it

thank you for your help

[woven prairie]

Has anyone worked on any application

Like which can do data analysis, and can do forecasting on the user data

[woven prairie]

What was the flow of your project

I have to make one , can you guide me ?

[errant idol]

hey everyone

i am into learning ML so i wonder if someone can assist me

like how to start and what should i learn is there any resources you recommand

[marble furnace]

i was going to say the exact thing aswell

[errant idol]

that is interesting

would you like if we start together ?

[marble furnace]

well im pretty much as lost as you but i dont mind at all!

[solar thistle]

Figured it out.

[solar thistle]

100% accurate now up to Len 100

[calm cipher]

oh nice! was that with a RNN with a transformer on top, or only transformers?

i'd be really curious to see the weights either way

and if you had to add positional encoding or it's working without it

also curious if you think two heads are required for this or if only one is necessary

[solar thistle]

uhhh

Its a transformer on top of a bi directional GRU classifier with like a little extra custom classafier injected in

idk if its necessary to have 2 really, but it helped a lot

it was getting caught up on palindromes made up of palindromes, and that seemed to help

like for example "wee kek kek kek eew"

[frail meteor]

Try to start with simple linear regression models.. Get the book, data, and compute the model on paper

PS: No one said it's going to be easy👍

[solar thistle]

I didnt add positional encoding like, directly

[calm cipher]

ah I see, I guess the GRU is learning some kind of position

[solar thistle]

I did give it this though

which like is kind of cheating but not really

def char_reflection_score(s):
    """Calculate symmetry score based on character reflection."""
    s = ''.join(c for c in s.lower() if c in string.ascii_lowercase)
    n = len(s)
    half = n // 2

    if n % 2 == 0:
        left = s[:half]
        right = s[half:]
    else:
        left = s[:half]
        right = s[half+1:]

    right = right[::-1]

    def embed(c):
        return (ord(c) - 97 - 13) / 13  # zero-centered

    vec1 = np.fromiter(map(embed, left), dtype=float)
    vec2 = np.fromiter(map(embed, right), dtype=float)

    diff = vec1 - vec2
    return 1.0 - np.mean(np.abs(diff))  # 1.0 = perfect symmetry```

it just splits a string basically, encodes it to a vector and then linear normalization

[frail meteor]

Whats an idea of the project?

[solar thistle]

i just wanted to explore ML, ive never done it before

so i trained a model to identify palindromes

I was joking with sdomeone I should write it in JS and release it as an NPM package lel

[frail meteor]

I would recommend to read about least squares algorithm, linear regression algorithm and try some simple data (about 50 lines of data)

[solar thistle]

oh im done now its 100% accurate lol. Can only beat this horse so much

[frail meteor]

Now get the multilinear regression with 10 predictor data))

[solar thistle]

What’s that mean

I’m an electrician lol

[frail meteor]

Oh no

[solar thistle]

I do wanna play with RAGs and LLMs though

[frail meteor]

Sounds scary tbh.. CS GO rules

[solar thistle]

Retrieval augmented generation. Like you can give it structured data that gets kinda injected into the prompt. It’d be cool to datastruct the whole electrical code book.

[frail meteor]

Idk.. I do like mine image recognition app with Telegram bot interface))

Based on on logistic regression

[solar thistle]

What’s it ID

[frail meteor]

Im trying to restructure it.. And I dont have server for 24/7

I also think if I have to split into 3×3 segments, rather than 2×2

[shadow viper]

hey guys, hows it going?

please anyone have an idea of how much it'd cost to build a chat bot?
i'm not asking for a job or offering one, i just want to confirm something

[serene scaffold]

From scratch? Millions of dollars

[solar thistle]

Free?

[shadow viper]

are you for real?
please i need to deliver an answer

[shadow viper]

not exactly

god why don't i know the marketing part of tech

[frail meteor]

At least he has to think about renting server, I guess

[frail meteor]

Chatbot for discord?

[spark kiln]

hi guys does anyone have a good/decent resource to learn ai? im kind of struggling rn

[serene scaffold]

What have you been trying to do?

[foggy jay]

Hey can anyone suggest me for numpy reference documentation, which I should use

Or some resource which I should use

[serene scaffold]

The numpy website has its reference documentation.

[foggy jay]

Which tutorial I should follow

I am in data science field

[woven prairie]

How things were working , was your input and how you were getting the output

[frigid crane]

how does one get into nlp

[steep raft]

guys,is 3blue1brown's website down?

@hot obsidian

[shadow viper]

let me confirm from my end, I only use their ytube

[shadow viper]

it's not from my end

[lapis sequoia]

I wanted to build a chat bot for discord just as practice since I’m new to python. Would you guys recommend that?

[serene scaffold]

building a chat bot from scratch is exceptionally complicated and expensive, and not something I would recommend. or which is possible.
making API calls to existing chatbots is just general software development and doesn't involve any actual AI.

[lapis sequoia]

What are some projects you would recommend for me to do as practice?

[serene scaffold]

making a basic classifier

[lapis sequoia]

You might face burn out and frustration like me 😅 , just start small and simple

[calm cipher]

This is a cool project but I don't understand how they're fixing the problem of reward hacking they describe in section 3.1 of the paper

I guess the idea is that they're qualitatively evaluating the architecture in addition to mazimizing performance, but that would just mean they're cherry-picking models that do well and look nice as opposed to just cherry-picking models that do well

something pretty easy to do would have been to hold out a few benchmarks that aren't part of the fitness function to show that the model generalizes as opposed to just optimizing for all the benchmarks

[calm cipher]

they're evaluating models on the training loss? wtf

[frail meteor]

I like it

Analysis for all life cases.. How do you input the data?

How do you guys keep your servers alive for 24/7?

Keep your desktop work for 24/7?

You use your desktop as server?

Ohh

How much do u pay?

Ohh.. Okok I see

If it works, you can make huge staff based on that data

Yes, we can discuss it on pm

I mean you can implement ml with it

For different apps

Are you good with the frontend programming?

Something like telegram bots?

Have you heard of telegram?

You can make a bot for telegram

Ohh okok

You write them from scratch or using specific libs?

Did you create discord.py by yourself?

Lmao ok😂

Yeah, in terms of bot creation the only one I made was in telegram

Yet

I like it😂👍

Good luck, man. Ill be back little later

[inner hemlock]

no it's actually community made

[marsh sage]

Can I share Logos

[calm cipher]

Against my better judgement I spent most of today looking into their results, lol

I have a lot of issues with the methodology of the paper but I put some visualizations together based on the lineage of different models

So first off this thing generated 1771 models and only 106 of them were actually selected as being good based on some iffy criteria

the vast majority of generated models performed worse than their parent

here's a visualization that shows the lineage of each model along with its score

it looks to me like this is maybe more akin to a genetic model that's just throwing stuff at the wall and seeing what sticks than anything that's actually making intelligent decisions about what to try next

Here's a better version of the lineage visualization, no change to the data, just better axis labels

[small wedge]

what issues do you have with their methodology?

[calm cipher]

there are a couple of things

all qualitative measuremeants of the model are done with LLMs

the qualitative LLM scores given to their models don't seem to be explained or explored in any depth and I don't know how it compares to something like a human rater, or code complexity measures, or anything that might show how it compares to other ways of evaluating code

if I take the language they use in the paper literally, they seem to be using the training loss rather than a validation loss when computing the score, which is useless

they don't use statistical tests to show if their results are meaningfully different from the human baselines

they only give a consolidated score for the 1,771 candidate models, so it's difficult to evaluate how the model performance changed as they adjusted the number of weights and training data size

it appears that some models might have performed worse as they scaled them up, which suggests that improvements are more due to random noise than actual improvements in the model

none of the training code or code that computes metrics are available in the repo

lol I think that's it

[small wedge]

hmm yeah that's not ideal

where are you seeing them getting worse as they scale?