Vital Contribution
@VContribution
Followers
1,733
Following
125
Media
257
Statuses
1,135
AI Researcher Heaven = Programming + Anime
Basement
Joined July 2022
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
lisa
• 902355 Tweets
#VSFashionShow
• 336240 Tweets
Venezuela
• 257455 Tweets
Colombia
• 202963 Tweets
#السعوديه_البحرين
• 177969 Tweets
Chile
• 167882 Tweets
Adriana Lima
• 132869 Tweets
سالم
• 117876 Tweets
Jimmy Carter
• 68803 Tweets
bella hadid
• 68444 Tweets
Tyla
• 63641 Tweets
مانشيني
• 61676 Tweets
#GHLímite6
• 50276 Tweets
Cher
• 38979 Tweets
Alessandra
• 37743 Tweets
Durán
• 35384 Tweets
Candice
• 27613 Tweets
kate moss
• 26099 Tweets
#فزعه_الشعب_السعودي
• 24047 Tweets
Anok
• 23770 Tweets
barbara palvin
• 23588 Tweets
Paraguay
• 22567 Tweets
EU TE TRATO MAL
• 21771 Tweets
Gareca
• 20215 Tweets
Gisele
• 19960 Tweets
Alex Consani
• 17968 Tweets
Tyra
• 17626 Tweets
パルクロ
• 13736 Tweets
Irina
• 13395 Tweets
kendall
• 13100 Tweets
Vidal
• 12435 Tweets
#LaRojaxCHV
• 11605 Tweets
Milad
• 11123 Tweets
Pinned Tweet
Alright, I posted now for exactly 60 days every single day my ML progress. It was a lot of fun but I will temporarily pause it due to my trip to Japan.
In the meantime, I will post some cool stuff that I see in Japan, damn I hope they have subtitles.
7
0
30
What stops you from learning the Cross-Entropy loss function inside a bush on a mountain?
45
52
1K
Just talked to someone that has been a ML Engineer for 4 years. I had to explain him the difference between RAM and VRAM...
Please tell me this is not normal.
56
5
302
First thing in the morning.
Breakfast + watching someone code an LSTM in PyTorch. What are you up to?
10
5
295
Day 12 - Back to the basics:
Felt stupid today so I put Aqua in the background
❖ Data Science/ML:
┈➤ PyTorch-Tensor: Learned the absolute basics (tensors, reshape, aggregation, etc.)
❖ Cybersecurity:
┈➤ DNS, whois, Google "hacking"
┈➤ Web Application Infrastructure
6
10
246
Thank you
@Trickywii
for making this project possible.🥰
#TrickyPiccy
#vtuber
#Robot
Song: Jim Yosef - Samurai [NCS Release]
Music provided by NoCopyrightSounds
9
17
230
Day 25 - my first paper implementation🔥
Paper: "Very Deep Convolutional Networks (VGG11)"
❖ Data Science/ML
┈➤ I picked a simple architecture (VGG11) and implemented it in PyTorch. It was actually a lot of fun.
┈➤ Learned about tensor.detach() and no_grad
11
4
232
Day 21 - 🔥🔥🔥 Karpathy announced today an AI school, best news ever!
❖ Data Science/ML
┈➤ Continued Karpathy's GPT Video (90% Progress): Understood Self-Attention.
┈➤ Implemented a simple GAN and trained it on MNIST. The final goal is to train it on Anime girls.
5
4
226
Day 16 - 🔥
❖ Data Science/ML:
┈➤ PyTorch: Trained RNN on firstname to country classes (18 different classes) got about ~50% accuracy. Tomorrow I want to compare it against a transformer.
❖ Cybersecurity:
┈➤ Port Scanning: Nmap (basics)
┈➤ Bug Bounty Course: CSS, JS
6
5
224
Day 20 - 🔥
❖ Data Science/ML
┈➤ Continued Karpathy's Video taking it slowly because I really want to understand it.
┈➤Added 3 old ML projects to my Portfolio Website (freelancing)
❖ Open Question (struggle)
┈➤ I still dont get what advantage view() over reshape() has.
9
8
225
I have just received an email by my University that I am going to graduate with honors from my CS master's program🫡.
I would like to thank my favorite study waifus Suzune Horikita, Makise Kurisu, Lofi girl and CC.
12
2
190
Day 8 - I'm happy:
❖ Data Science/ML:
┈➤ ViT-PyTorch: Implemented class & position learnable embeddings for image patches. Transformer tomorrow! I'm hyped!
❖ Cybersecurity:
┈➤ Learned about TLS/SSL
┈➤ Found two flags in Hack the Box machines. (but tbh was super basic)
3
5
189
Day 11 - CNN > ViT:
❖ Data Science/ML:
┈➤ ViT-PyTorch: Finally started training! Then I realized it would cost me 5k dollars. Ditched ViT and implemented CNN.
❖ Cybersecurity:
┈➤ Attack vectors on encrypted passwords
┈➤ Bug Bounty Course: HTTP headers/methods/codes
3
6
168
Day 32 - ML Journey 🔥
❖ Data Science/ML
┈➤ Read one chapter in the book "The Art of Statistics"
┈➤ Implemented FastAPI endpoint for an LLM but not scalable🥲
❖ Struggle:
┈➤ How the hell do I deploy an LLM like ChatGPT to handle many users quickly? Kubernetes?
5
0
148
Day 43 - learning is fun, especially ML🔥
❖ Data Science/ML:
┈➤ Test training run Llama3.1 to become an agent. Estimate time: 21h I will be so broke.
┈➤ Learned about stochastic Bayesian neural nets
┈➤ Implemented FastAPI endpoint that supports token streaming (gpt2 rn)
1
2
115
Day 46 - 🔥
❖ Data Science/ML:
┈➤ Read more about data cleaning for LLMs, duplicates in the dataset really seem to drag down the LLM.
┈➤ Learned about Weight Decay (L1 & L2 regularization)
┈➤ Added some new sections to my personal ML portfolio website
2
8
111
Finally trying out something useful.
Learning more about GANs.
Has anyone a catgirl image dataset?
6
3
107
Day 29 -🔥
❖ Data Science/ML
┈➤ Implemented different OCR data extraction strategies: OCR + LLaMA3.1, OCR + normalized bounding boxes
┈➤ "Textbooks Are All You Need" got me excited about model distillation. I implemented a Training script for gpt-2 and some synthetic data
5
3
95
Day 47 - 🔥
It's strange, but watching Anime characters achieve great things makes me want to achieve great things too.
❖ Data Science/ML
┈➤ Finally trained Llama3.1 (9h, 100k dataset). Open source soon, if good.
┈➤ Redid CNN satellite image classifier project (portfolio)
3
0
92
Day 9 - Could have done more but still was good:
❖ Data Science/ML:
┈➤ ViT-PyTorch: Implemented transformer encoder components (MSA & MLP)
❖ Cybersecurity:
┈➤ Learned about Hashing, Digital Signatures, Salt & Pepper (got really hungry during my Cybersecurity session)
4
1
90
Day 35 - Today was really good but the day does not have enough hours.
❖ Data Science/ML
┈➤ Implemented first LLM Agent from scratch
┈➤ Understood Batch Normalization vs Layer Normalization
┈➤ Tested out history awareness of instruct embedding models
2
3
89
Note to myself:
When posting my daily progress reports, the one with anime girl images perform better.
7
1
88
Started my AI research job.
It kind of feels wrong to get paid for something that's so much fun.
8
1
86
Day 50 - 50 days and I don't regret starting progress posting at all met so many amazing people 🔥
❖ Data Science/ML
┈➤ Reviewed Transformers library & learned basic Python: getattr, isinstance, *args, **kwargs
┈➤ Fixed all my mistakes in the Llama3.1 training script
3
1
81
Day 52 - Data Cleaning LFG 🔥
❖ Data Science/ML
┈➤ Started to implement data cleaning pipeline with model distillation (I really start to like this)
┈➤ Dived into the transformer embedding and finally understood why we need "model.resize_token_embeddings(len(tokenizer))"
1
4
75
Day 7 - Today was kinda meh:
Had to attend a lot of job meetings.
Spend way too much time on designing a new x banner.
To post this actually feels bad lmao!
▰▱▰▱▰ Cybersecurity:
╰┈➤ Learned about encryption: Symmetric (AES) and Asymmetric encryption (RSA)
▰▱▰▱▰
3
1
75
One of the simple things that really boosted my productivity was to literally always have an “unknown word” list around me so that whenever you are reading or watching something and an interesting unknown word appears, add it to the list to learn it.
8
5
73
Day 12 - Sunday is an amazing day to grind:
❖ Data Science/ML:
┈➤Waifu-CNN: Classifying 130 waifus with ResNet50. Ran some hyperparameters experiments.
❖ Cybersecurity:
┈➤ Reconnaissance and information gathering
┈➤ Bug Bounty Course: Finished Module Web Requests.
5
1
66
Day 18 - My RNN sucked😂
❖ Data Science/ML
┈➤ PyTorch: Implemented and finetuned BERT on last-name to country classification.
┈➤ BERT vs RNN: BERT destroyed my RNN. RNN achieved 50% accuracy on test set and BERT achieved 90%.
I want bigger projects, but basics come first.
4
4
64
Today went quite well🔥
# Data Science/Machine Learning
>> Found fine-tuned OCR system that performs quite well on a task for my job.
>> Came up with some cool ideas on how to analyze VTubers with ML
# Cybersecurity
>> Partitions, UUID, GRUB and File Systems
>> Learned about
5
1
64
Day 49 - 🔥
❖ Data Science/ML
┈➤ Updated RAG to have better chunking mechanism
┈➤ Analyzed entire Llama3.1 run due to failed training
My mistakes so far:
┈┈➤ SFTTrainer adds special tokens
┈┈➤ Llama3.1 has ~250 reserved special tokens. I added a completely new one
0
0
54
Day 24 - 🔥
❖ Data Science/ML
┈➤ Finished DC-GAN. Results look certainly interesting. Will try out more sophisticated Generator architectures.
┈➤ Implemented Object Character Recognition (OCR) script to extract some specific text from documents.
1
1
56
Day 48 - 🔥
❖ Data Science/ML
┈➤ Evaluated the trained Llama3.1 but it generates strange special tokens uff and back
┈➤ Understood a lot more why L1 regularization promotes sparsity (compared to L2)
┈➤ Wrote a python script that obfuscates + AES encrypts python files
1
0
55
New anime pfp observation:
It seems like when people change their anime pfp with a friend their post style changes as well.
Could the anime pfp be deeply bound to a person's character?
3
1
52
Day 55 - 🔥
❖ Data Science/ML
┈➤ Collected information of ~500 twitch streams
┈➤ Used Llama3.1 to extract some data specific data out of a large amount of text
┈➤ Algo trading: was able to get all signatures + webhook for a wallet. I have no idea what I'm doing lmao
0
0
51
ML Day 60 - 🔥
❖ Topics
┈➤ Product owner gave feedback and I integrated it into the LLM prototype.
┈➤ Did some research on papers for LLM creativity (so that I have something to read on my flight to Japan)
It's not much but preparing the Japan trip takes a lot of time.
1
1
50
Question:
What do you do while your ML model is training?
10
0
48
Day 4:
▰▱▰▱▰
Data Science/ML (VTuber Analysis):
ׂ╰┈➤ Implemented YouTube video stats data scraper
ׂ╰┈➤ Continued to re-implement ViT paper in PyTorch
Cybersecurity:
╰┈➤ Finished Kali Linux introduction section
ׂ╰┈➤ IP addresses, NAT, Subnetmasks
▰▱▰▱▰
3
1
48
Day 5:
▰▱▰▱▰
Data Science/ML (VTuber Analysis)
╰┈➤Reimplemented scraper (found out YouTube API v3 offers a lot more)
Cybersecurity:
╰┈➤ Learned key terms: spoofing methods, sniffing, APT, bind/reverse shell, rootkit
╰┈➤ Created Anki Deck for this
▰▱▰▱▰
2
3
47
Day 51 - Agent day🔥
❖ Data Science/ML
┈➤ Had to write [text not code :(] concept for agent that utilizes tools like OCR, RAG and API calls.
┈➤ Implemented Agent with gpt-3.5 prototype cause potential clients wants it
┈➤ Data cleaning for some youtube videos meta data
0
1
47
Day 22 - Too many meetings today but still tried to learn🔥
❖ Data Science/ML
┈➤ Finished Karpathy's GPT Video🥳
Learned so much! But I will have to watch it again in the future
┈➤ Started to implement the DC-GAN similar to the GAN from yesterday but with convolutions
1
1
44
Day 19 - GPT.
❖ Data Science/ML
┈➤ Watched 50% of Karpathy's transformer video. (every time I watch his videos I realize how much I still have to learn and I love it)
┈➤ Secret data analysis project. Came up with some funny analysis ideas (YouTube video might follow lmao)
0
0
44
Feels good to not waste your time while waiting in a bus. (ViT at scale).
1
0
43
Day 36 - Trained a model to detect jokes - needed this desperately as a German.
❖ Data Science/ML
┈➤ Finetuned BERT on a 200k jokes dataset to become "JokeDetectBERT"
┈➤ Understood Bias Variance Tradeoff a lot better
┈➤ GraphRAG seems to be nice for some uses-cases.
0
1
41
Day 30 -🔥
❖ Data Science/ML
┈➤ Tested and compared OCR data extraction strategies: OCR + LLaMA3.1, OCR + normalized bbox, LLaVA1.6 vicuna. ==> OCR + LLaMA3.1 won.
┈➤ Model distillation: tested GPT2 training run in colab. I will move the training tomorrow into the cloud.
0
0
39
Day 10 - good one:
❖ Data Science/ML:
┈➤ ViT-PyTorch: Finished implementing ViT architecture (stacked encoder block + MLP classifier at the end)
❖ Cybersecurity:
┈➤ Learned about public key infrastructure (fck I love public & private keys)
┈➤ Started Bug Bounty Course
2
0
40
Nothing beats coding at 6:30 AM on a sunny morning over the roof tops.
4
1
37
Day 17 - erm errors
❖ Data Science/ML:
┈➤ PyTorch: Implemented 95% of the transformer training script (BERT). The last 5% are some errors that I will fix tomorrow)
❖ Cybersecurity:
┈➤ Bug Bounty Course: HTML Injection, XSS, Cross-Site Request Forgery (Overview)
0
5
34
Day 15 - Definitely a good day🔥
❖ Data Science/ML:
┈➤ PyTorch: Started to understand and implement an RNN. Tomorrow training (classification [name -> country])
❖ Cybersecurity:
┈➤ Completed reconnaissance section
┈➤ Many protocols: ARP, TCP/UDP, ICMP
1
1
35
Day 53 - Data Cleaning LFG 🔥
❖ Data Science/ML
┈➤ Started to read LLMClean Paper
┈➤ YouTube Video Data Visualizations
❖ Lessons Learned
┈➤ Dataset cleaning works better by integrating the context they were collected
0
0
34
Day 41 - 🔥
❖ Data Science/ML:
┈➤ Learned a lot about LoRA and QLoRA. I get the feeling it's very often a trade off between computation and memory.
┈➤ Used SAM-2 + OCR
┈➤ Agents: Finished writing the training dataset preprocessing for Llama3.1-8b
1
0
34
Day 34 - ML4ever 🔥
❖ Data Science/ML
┈➤ Learned about: Bytepair encoding, distributed training (workers, world size, rank), GGUF
┈➤ Had to rewrite my vector db stuff cause of an error but i swear i did not change anything.
❖ I feel like I want to dive into agents.
2
0
34
@angkul07
Isn't 'senior' just a job title companies throw on people quickly so that they are more attractive to potential employees?
2
0
33
Day 56 - 🧳🇯🇵
Today was meh when it comes to ML progress but still really exciting because I spend nearly the entire day to plan and purchase some stuff for my flight to Japan.
❖ Data Science/ML
┈➤ just added a new project to my ML portfolio website
1
0
32
Day 33 - ML Sunday 🔥
❖ Data Science/ML
┈➤ Got first client who was interested in adding machine learning to their product and reached out to me for suggestions. --> created presentation.
┈➤ Did some basic data analysis and visualizations of someones YouTube comments lmao
4
0
32
Day 14 - Still back to the basics but felt pretty good🔥
❖ Data Science/ML:
┈➤ PyTorch: Yesterday MLP >> Today CNN.
┈➤ Tensors: Asked Claude to give me tons of tensor exercises.
❖ Cybersecurity:
┈➤ Reconnaissance: Maltego & recon-ng
┈➤ OWASP & Front-End (HTML)
0
0
31
I really love to learn with visualizations.
The simple Softmax function is actually really cool once you visualize it. Amazing video.
(Disclaimer: Anime girl not included)
2
0
32
Day 27 - I swear the background helps me to code faster.
❖ Data Science/ML
┈➤ VGG11 script updated with training. MNIST attempt failed due to max pooling. Fix tomorrow
┈➤ Explored ID card segmentation to improve OCR. No suitable model found yet. Might have to do it myself.
0
0
32
Tried to write today a custom optimizer in pytorch.
The loss is supposed to go up right?
2
1
30
Day 39 - nice one🔥
❖ Data Science/ML:
┈➤ PyTorch revisited: rebuilt neural net using AdamW-8bit and gradient accumulation.
┈➤ Learned Gradient Check-pointing
┈➤ Data Cleaning: Cleaning dataset with embedding spaces (visualized a prompt injection dataset)
0
0
30
Day 28 - booked tickets to Japan🤭
❖ Data Science/ML
┈➤ Fixed the VGG11 training loop
┈➤ Put some more time into research how to extract information from images and documents and found LayoutLMv3 (seems to be good).
┈➤ Did some data visualizations for a YouTube video.
1
1
29
@farhahahahaan
It's "The Little Book of Deep Learning" by
@francoisfleuret
it's a great book!
1
0
29
@mrsiipa
Looking back at my university time, I would have liked to pay less attention to grades and more attention to my own projects.
1
0
27
Day 45 - PEBCAC error🙄
❖ Data Science/ML:
┈➤ Found the problem in the Llama3.1 training script. Turns out I was the problem.
┈➤ Wrote base podman template to easily deploy training scripts
┈➤ Spend some time on document perspective correction transformations (for OCR)
1
2
27
Day 26 - bus drive => ML paper
Had a very long bus drive so I just read some ML papers.
❖ Data Science/ML
┈➤ Read some model cards BERT (tbh, this was a very cool way)
┈➤ "Textbooks Are All You Need"
┈➤ "Interpretable Deep Learning: Interpretation, Interpretability"
0
1
24
Day 23 - Overall nice day🔥
❖ Data Science/ML
┈➤ Implemented DC-GAN. Generator loss skyrocketed during training and very low Discriminator loss. Need to review Generator tomorrow
┈➤ Learned more about Tensorboard (seems to be a nice local alternative to weights and biases)
4
1
25
Day 54 - 🔥
❖ Data Science/ML
┈➤ Created basic GUI for code generation LLM Prototype
┈➤ Looked a bit into algo trading
┈➤ More YouTube video data visualizations
❖ Lessons Learned
┈➤ Need to lock in more. I don't get enough stuff done. The day should have more hours
0
0
25
“Size matters in machine learning,
but will this continue to be true?”
0
0
25
Day 40 - damn 40🔥
❖ Data Science/ML:
┈➤ Agents: Writing training script for Llama3.1. If it works i'm gonna open source the model, if it does not work i'm gonna cry.
┈➤ Fine-tuned BERT got like 94% on val dataset but only 80% on testing dataset (something is off here)
0
1
25
Day 44 - 🔥
❖ Data Science/ML:
┈➤ Found some issue in the Llama3.1 training script. I came pretty close but still did not find it.
┈➤ Learned about serverless server.
┈➤ Improved FastAPI endpoint so that it not only streams tokens but also sources (RAG + Citations)
0
0
25
Day 38 - Definitely a good day🔥
❖ Data Science/ML:
┈➤ Read today
@maximelabonne
's "Fine-tune Llama 3.1 Ultra-Efficiently" amazing article. Took some notes & will apply it definitely to future training runs
❖ Cybersecurity:
┈➤ Bug Bounty Course: Proxies, Burp suite, ZAP
2
1
25
Day 6 - Today was great🔥:
▰▱▰▱▰
Data Science/ML:
╰┈➤ViT: Implemented image patch embedding model in PyTorch (with Conv2d layer).
╰┈➤ Scraped all Videos of one my fav VTuber. Analysis can begin😍.
Cybersecurity:
╰┈➤ Proxies, Proxy Chain, VPNs, Tor
▰▱▰▱▰
1
1
23
Applying all of my new knowledge.
Sequence Classification, BERT, Fine-tuning, Weighted Random Sampling, Gradient Accumulation, Learning Rate Scheduling, Weight Decay
1
1
23
Day 57 - 🔥
❖ Topics
┈➤ Created another GUI with Python for LLM
┈➤ Algo Trading
┈➤ Learned Hiragana
❖ Lessons Learned
┈➤ I never want to implement a GUI in python ever again.
┈➤ Becoming fluent in Japanese in a week is kinda difficult
1
1
24
Day 31 -🔥
❖ Data Science/ML
┈➤ RAG: Extended embedding model and Llama3.1 to understand historical context.
┈➤ Ran distill training for GPT-2-large on A6000 (Runpod) and I never thought that the batch-size would have such a big impact on VRAM. 48GB was not enough at all.
4
0
23
New bus drive new ML paper.
Keep learning everywhere 🔥
1
0
21
Day 37 - meh day, so many work related meetings today still got a little bit done.
❖ Data Science/ML
┈➤ RAG: Splitted RAG into retrievel (embedding) and Llama3.1 response generation (for testing)
┈➤ Read a couple of articles on dataset preparation for fine-tuning LLMs.
0
0
22
Tokenizers split Chinese, Korean or Japanese sentences into more tokens compared to English sentences. Therefore, they have to pay more for API calls...
Are tokenizers racist?
3
0
21
Day 42 - theory day & too many work meetings & it was super hot. Not sure how I will survive Japan
❖ Data Science/ML:
┈➤ Learned more about "warm up" and "learning rate scheduler". Thought they'd contradict, but they actually harmonize.
┈➤ Learned why Adam is goated
0
0
19
Two important things that I learned yesterday:
1. You can combine CNNs and LSTMs Networks.
2. You can't trust anime pfp anymore.
4
0
20
If you are not singing the PyTorch training loop song, you are ngmi.
2
1
19
YOU GUYS ARE INSANE.
I never really expected that so many people would be interested in vtuber/anime robots, holy shit!
Thank you all so damn much!
#vtuber
#VtuberUprsing
#VtubersEN
#anime
6
1
18
Japan made me realize that I'm too tall.
I hit my head several times, my ML journey might suffer from this.
4
0
17
Just saw a description of a freelancer on Upwork.
Can someone please explain to me what is "Time Serious Prediction"?
5
0
17
ML Day 59 - 🔥
❖ Topics
┈➤ Finished GUI for LLM code generator
┈➤ Meeting with co-founder and integrated the LLM into GUI (integration was successful really happy about this one)
┈➤ Designed logo for startup (man perfectionism is really a curse)
0
0
17
you literally made my thanksgiving
@ironmouse
!
ngl that you rect to one of my videos was one of my dream goals when I started my vtuber tech channel.❤️
#ironmouse
#vtuber
3
3
16
I love Huggingface a lot, but it's a single point of failure for the open-source community. I don't know any other platform that hosts as many models as Huggingface.
3
0
16
Someone called me "RAG-Expert" at work today.
My ML career is so over.
1
0
15
I wonder what this DC-GAN is going to generate.
(Image output after: ~200 steps nya | batch-size: 128 | dataset-size: 10k | epochs: 10 | input-noise-dimension: 100)
1
0
15
Day 58 - 🔥
❖ Topics
┈➤ Data Cleaning: LLM Evaluator with Llama3.1-70b evaluation script
┈➤ Learned about blockchain, tokens, transactions
┈➤ Learned more Hiragana
❖ Thoughts
┈➤ I'm more interested in LLM creativity
┈➤ I will not become fluent in Japanese in a week
0
0
15
Day 3:
▷ ML/Data Science
✓ Test deployed my Onboarding RAG-System + Front-End on Runpod
✓ Started a tutorial to re-implement the first Vision Transformer paper.
▷ Cybersecurity
✘ Nothing really just stared at some IP addresses 🥹. Have to catch up tomorrow 🔥
0
0
14
You know you are in Japan when your train has its own anime girl.
0
0
15