Kiril Morozov βishop

Don't try. | Physicist | Lead Data & AI Scientist at Convotis | Photographer

[Kiril Morozov βishop @AntrhaxXx]

@systemdesignone Playbook

[Neo Kim @systemdesignone]

I just created a playbook to help you learn AI engineering.

It gives you:

• Core concepts behind modern AI systems.
• Condensed notes to understand AI engineering stack.
• Must know techniques to build, deploy, debug & scale AI apps.

(24 hours only!)

To get it for FREE:
1. Like, Retweet & Follow @systemdesignone
2. Reply "Playbook"

Then I'll DM you the details.

[Kiril Morozov βishop @AntrhaxXx]

@Gsnchez Con que has montado el esquema si no es secreto? Gracias. Enhorabuena por el trabajo.

[Kiril Morozov βishop @AntrhaxXx]

@paulabartabajo_ @liquid Hasta el nabo de que viváis del hype promocionando cosas por las que os pagan. Sois como la prensa comprada por el gobierno. Este tipo de modelos vale para POCs que le puede enseñar a tus colegas o colgar en X. Uso real? De cara a cliente? Ninguno.

[Kiril Morozov βishop @AntrhaxXx]

@calcsam Book

[Kiril Morozov βishop @AntrhaxXx]

@calcsam Copa

[Elon Musk @elonmusk]

Accurate

[Kiril Morozov βishop @AntrhaxXx]

@NVIDIAGeForce GeForce Day

[Tom Dörr @tom_doerr]

multi-language OCR toolkit with custom recognition model and Python API

[Kiril Morozov βishop @AntrhaxXx]

@albrommar1 @MarinaMasss @WillyTolerdoo ¿Tan difícil es entender, tú qué aspiras a ser profe de políticas, que 6-7% de 1080€ y 6-7% de 2080€ es no es la misma cuantía? Estando el nivel de vida relativamente homogeneizado en toda la UE. ¡Qué bárbaro, así nos va!

[elvis @omarsar0]

Say goodbye to Chain-of-Thought.

Say hello to Chain-of-Draft.

To address the issue of latency in reasoning LLMs, this work introduces Chain-of-Draft (CoD).

Read on for more:

[Ben Pierron @Ben_escrito]

@feedcoyote Si te ha gustado este hilo, por favor:

1. Sígueme @Ben_escrito

2. Retuitea el siguiente tweet para difundirlo

[Akshay 🚀 @akshay_pachaar]

Microsoft has released its own document parser for LLM use!
.
.
Introducing MarkItDown, a 100% open-source, one-stop solution for effortlessly converting any file to Markdown—perfect for text analysis, indexing, and more!

Here’s what makes it special:

↳ Converts PDF, Word, Excel, PPT, images, audio to markdown
↳ Extracts EXIF, OCR, and transcripts automatically
↳ Available via CLI, Python API, or Docker
↳ Offers LLM-based image descriptions
↳ Supports batch conversions

Link to the repo in next tweet!
_____
Find me → @akshay_pachaar ✔️
For more insights & tutorials on AI and Machine Learning.

[Rohan Paul @rohanpaul_ai]

Self-attention is actually doing kernel PCA under the hood - now we can make it robust

So how does Self-attention works: it's kernel PCA in disguise, as proposed in this paper.

🤔 Original Problem:

Self-attention in transformers has been developed through heuristics and experience, lacking a systematic theoretical framework to understand its inner workings and improve robustness against data corruption.

-----

🔧 Solution in this Paper:

→ Derives self-attention from kernel Principal Component Analysis (kernel PCA), showing it projects query vectors onto principal component axes of key matrix

→ Introduces RPC-Attention (Attention with Robust Principal Components) that uses Principal Component Pursuit to handle corrupted data

→ Implements PAP (Principal Attention Pursuit) algorithm that iteratively recovers clean data from corrupted inputs

-----

💡 Key Insights:

→ Self-attention mathematically performs kernel PCA in feature space

→ Value matrix captures eigenvectors of Gram matrix of key vectors

→ Number of principal components used must be ≤ number of data points

→ Different value matrix parameterizations lead to different attention architectures

-----

📊 Results:

→ RPC-Attention outperforms baseline on ImageNet-1K by 1% accuracy

→ Shows 3% improvement in AUPR on ImageNet-O

→ Better performance against PGD, FGSM, SPSA adversarial attacks

→ 1 PPL improvement on WikiText-103 language modeling

[elvis @omarsar0]

MegaParse is an open-source tool for parsing various types of documents for LLM ingestion.

Supports text, PDF, PowerPoint, excel, csv, and Word documents. It can convert these into a format ideal for LLMs.

It can parse content of different types such as tables, TOC, headers, footers, images, etc.

I am also building a similar tool and I think the most important feature at the moment is the ability to customize the format of the transformed data as different LLMs prefer different formats.

[Rohan Paul @rohanpaul_ai]

The paper list of the 86-page paper "The Rise and Potential of Large Language Model Based Agents: A Survey"

[Trending GitHub Repositories @trending_repos]

Trending repository of the day 📈

awesome-remote-job

A curated list of awesome remote jobs and resources. Inspired by https://t.co/T0LH6gvkp3

Last 24h: 561 ⭐
Total: 30991 ⭐️
https://t.co/7IBLX4JfXq

[Club de los Viernes @clubdeviernes]

Siempre ha sido más fácil robar la riqueza ajena que producir la propia

[Santiago @svpino]

Large Language Models are not "intelligent."

More importantly, it's abundantly clear we won't get to general intelligence by simply scaling these models.

We need new breakthroughs. We need a different architecture.

(Those waiting for GPT-5 to be AGI will probably be disappointed.)

Attached is yet another paper dunking on the hype.

An important idea from the paper:

"Our findings suggest that purported emergent abilities are not truly emergent, but result from a combination of in-context learning, model memory, and linguistic knowledge."

In English: The intelligence you see is coming from memorization and the person writing the prompt.

Another one:

"The ability to follow instructions does not imply having reasoning abilities, and more importantly, it does not imply the possibility of latent, potentially-dangerous abilities."

Yeah, we knew that already: LLMs are not going to destroy the world.

[Python Programming @PythonPr]

GPT-4 Vision API | Web Scraping with GPT-4 Vision API and Puppeteer
#gpt #python #puppeteer
https://t.co/HKvWHxN7QC