UMAILab

@umaidotlab

Unified Methods of Artificial Intelligence Lab

Joined November 2024

21 Following

6 Followers

19 Posts

umaidotlab retweeted

almost 2 years ago

Introducing CerebrasCoder! An open-source app that generates websites with Llama3.3-70b from @cerebras as fast as you can type. 100% free and open-source. https://t.co/eaM2bWkEw3

59

2K

246

2K

241K

umaidotlab retweeted

Ivan Fioravanti

@ivanfioravanti

almost 2 years ago

Qwen 2.5 Technical Report is a great weekend read: - pre-training 18 trillion tokens with better data - post-training: over 1 million SFT samples - Long-context Pre-training interesting part on how they achieved 1M on Turbo - Comparison with other models in various sizes! Something other vendors keep missing in their papers 😜 7B and 32B base and instruct rock!

ivanfioravanti's tweet photo. Qwen 2.5 Technical Report is a great weekend read:
- pre-training 18 trillion tokens with better data
- post-training: over 1 million SFT samples
- Long-context Pre-training interesting part on how they achieved 1M on Turbo
- Comparison with other models in various sizes! Something other vendors keep missing in their papers 😜

7B and 32B base and instruct rock!

2

149

23

61

14K

UMAILab @umaidotlab

almost 2 years ago

Only fully open source is Qwen 2.5 🤔

almost 2 years ago

WebDev Arena Leaderboard is now live with 10K+ votes! #1. Claude 3.5 Sonnet #2. Gemini-Exp-1206 #3. Gemini-2.0-Flash #4. GPT-4o-2024-11-20 #5. Qwen2.5-Coder-32B #6. Gemini-1.5-Pro-002 Congrats @AnthropicAI topping the leaderboard by a significant margin🔥 Leaderboard link and examples below👇

arena's tweet photo. WebDev Arena Leaderboard is now live with 10K+ votes!

#1. Claude 3.5 Sonnet
#2. Gemini-Exp-1206
#3. Gemini-2.0-Flash
#4. GPT-4o-2024-11-20
#5. Qwen2.5-Coder-32B
#6. Gemini-1.5-Pro-002

Congrats @AnthropicAI topping the leaderboard by a significant margin🔥

Leaderboard link and examples below👇

36

803

99

207

192K

0

0

0

0

23

UMAILab @umaidotlab

almost 2 years ago

Bite 🧛‍♂️

almost 2 years ago

META JUST KILLED TOKENIZATION !!! A few hours ago they released "Byte Latent Transformer". A tokenizer free architecture that dynamically encodes Bytes into Patches and achieves better inference efficiency and robustness! (I was just talking about how we need dynamic tokenization that is learned during training 🥲 It's like fucking christmas!) I don't want to talk too much about the architecture. But here's a nice visualization from their paper. Let's look at benchmarks instead :) "BLT models can match the performance of tokenization-based models like Llama 3 at scales up to 8B and 4T bytes, and can trade minor losses in evaluation metrics for up to 50% reductions in inference flops!" This is basically a perplexity vs training flops chart - scaling laws with compute. BPB is a tokenizer independent version of perplexity. BLT is on par or better than LLama 3 BPE! Most importantly they scale this approach to train Llama-3 8B model on 1T tokens which beats the standard Llama-3 architecture with BPE tokenizer!

scaling01's tweet photo. META JUST KILLED TOKENIZATION !!!

A few hours ago they released "Byte Latent Transformer". A tokenizer free architecture that dynamically encodes Bytes into Patches and achieves better inference efficiency and robustness!

(I was just talking about how we need dynamic tokenization that is learned during training 🥲
It's like fucking christmas!)

I don't want to talk too much about the architecture.
But here's a nice visualization from their paper.

Let's look at benchmarks instead :)

"BLT models can match the performance of tokenization-based models like Llama 3 at scales up to 8B and 4T bytes, and can trade minor losses in evaluation metrics for up to 50% reductions in inference flops!"

This is basically a perplexity vs training flops chart - scaling laws with compute. BPB is a tokenizer independent version of perplexity.

BLT is on par or better than LLama 3 BPE!

Most importantly they scale this approach to train Llama-3 8B model on 1T tokens which beats the standard Llama-3 architecture with BPE tokenizer!

scaling01's tweet photo. META JUST KILLED TOKENIZATION !!!

A few hours ago they released "Byte Latent Transformer". A tokenizer free architecture that dynamically encodes Bytes into Patches and achieves better inference efficiency and robustness!

(I was just talking about how we need dynamic tokenization that is learned during training 🥲
It's like fucking christmas!)

I don't want to talk too much about the architecture.
But here's a nice visualization from their paper.

Let's look at benchmarks instead :)

"BLT models can match the performance of tokenization-based models like Llama 3 at scales up to 8B and 4T bytes, and can trade minor losses in evaluation metrics for up to 50% reductions in inference flops!"

This is basically a perplexity vs training flops chart - scaling laws with compute. BPB is a tokenizer independent version of perplexity.

BLT is on par or better than LLama 3 BPE!

Most importantly they scale this approach to train Llama-3 8B model on 1T tokens which beats the standard Llama-3 architecture with BPE tokenizer!

scaling01's tweet photo. META JUST KILLED TOKENIZATION !!!

A few hours ago they released "Byte Latent Transformer". A tokenizer free architecture that dynamically encodes Bytes into Patches and achieves better inference efficiency and robustness!

(I was just talking about how we need dynamic tokenization that is learned during training 🥲
It's like fucking christmas!)

I don't want to talk too much about the architecture.
But here's a nice visualization from their paper.

Let's look at benchmarks instead :)

"BLT models can match the performance of tokenization-based models like Llama 3 at scales up to 8B and 4T bytes, and can trade minor losses in evaluation metrics for up to 50% reductions in inference flops!"

This is basically a perplexity vs training flops chart - scaling laws with compute. BPB is a tokenizer independent version of perplexity.

BLT is on par or better than LLama 3 BPE!

Most importantly they scale this approach to train Llama-3 8B model on 1T tokens which beats the standard Llama-3 architecture with BPE tokenizer!

scaling01's tweet photo. META JUST KILLED TOKENIZATION !!!

A few hours ago they released "Byte Latent Transformer". A tokenizer free architecture that dynamically encodes Bytes into Patches and achieves better inference efficiency and robustness!

(I was just talking about how we need dynamic tokenization that is learned during training 🥲
It's like fucking christmas!)

I don't want to talk too much about the architecture.
But here's a nice visualization from their paper.

Let's look at benchmarks instead :)

"BLT models can match the performance of tokenization-based models like Llama 3 at scales up to 8B and 4T bytes, and can trade minor losses in evaluation metrics for up to 50% reductions in inference flops!"

This is basically a perplexity vs training flops chart - scaling laws with compute. BPB is a tokenizer independent version of perplexity.

BLT is on par or better than LLama 3 BPE!

Most importantly they scale this approach to train Llama-3 8B model on 1T tokens which beats the standard Llama-3 architecture with BPE tokenizer!

92

5K

681

4K

490K

0

0

0

0

23

UMAILab @umaidotlab

almost 2 years ago

Hatırlatma #LLMs

UMAILab @umaidotlab

almost 2 years ago

Haftaya LLM’leri yani Büyük Dil Modellerini konuşacağız.

umaidotlab's tweet photo. Haftaya LLM’leri yani Büyük Dil Modellerini konuşacağız. https://t.co/k3VlChjmF4

0

0

0

0

44

0

0

0

0

33

UMAILab @umaidotlab

almost 2 years ago

We all love ollama ❤️🦙🦙

almost 2 years ago

Ollama 0.5 is here with structured outputs! This makes it possible to constrain a model’s output to a specific format defined by a JSON schema. Some examples include: - Parsing data from documents - Extracting data from images - Structuring all language model responses - More reliability and consistency than JSON mode 🧵1/3

ollama's tweet photo. Ollama 0.5 is here with structured outputs!

This makes it possible to constrain a model’s output to a specific format defined by a JSON schema.

Some examples include:

- Parsing data from documents
- Extracting data from images
- Structuring all language model responses
- More reliability and consistency than JSON mode

🧵1/3

51

3K

351

1K

238K

0

0

0

0

28

UMAILab @umaidotlab

almost 2 years ago

Haftaya LLM’leri yani Büyük Dil Modellerini konuşacağız.

umaidotlab's tweet photo. Haftaya LLM’leri yani Büyük Dil Modellerini konuşacağız. https://t.co/k3VlChjmF4

0

0

0

0

44

UMAILab @umaidotlab

almost 2 years ago

Today Time Zone : Istanbul , UTC +3 8.00 PM Live session link is under the post. Come and upgrade yourselves 🖖 #architecture #architects #design #designers #AI

UMAILab @umaidotlab

almost 2 years ago

Free Workshop Alert ! Architects and designers 📢 If you’re interested in AI, this workshop is for you. Tomorrow we will share the live session link. #architecture #architects #design #designers #AI #transformation

umaidotlab's tweet photo. Free Workshop Alert !

Architects and designers 📢

If you’re interested in AI, this workshop is for you.
Tomorrow we will share the live session link.
#architecture #architects #design #designers #AI #transformation https://t.co/BWF5toR2RA

1

3

3

0

129

0

0

0

0

35

umaidotlab retweeted

UMAILab @umaidotlab

almost 2 years ago

Free Workshop Alert ! Architects and designers 📢 If you’re interested in AI, this workshop is for you. Tomorrow we will share the live session link. #architecture #architects #design #designers #AI #transformation

umaidotlab's tweet photo. Free Workshop Alert !

Architects and designers 📢

If you’re interested in AI, this workshop is for you.
Tomorrow we will share the live session link.
#architecture #architects #design #designers #AI #transformation https://t.co/BWF5toR2RA

1

3

3

0

129

UMAILab @umaidotlab

almost 2 years ago

The Transformation Academy For Interior Architects & Designers Using MJ & comfyUI 4 Aralık 2024, 8:00pm - 5 Aralık 2024, 12:00am · Saat dilimi: Europe/Istanbul Google Meet katılma bilgileri Görüntülü görüşme bağlantısı: https://t.co/Th8UzadWRc

0

0

0

0

22

UMAILab @umaidotlab

almost 2 years ago

Free Workshop Alert ! Architects and designers 📢 If you’re interested in AI, this workshop is for you. Tomorrow we will share the live session link. #architecture #architects #design #designers #AI #transformation

umaidotlab's tweet photo. Free Workshop Alert !

Architects and designers 📢

If you’re interested in AI, this workshop is for you.
Tomorrow we will share the live session link.
#architecture #architects #design #designers #AI #transformation https://t.co/BWF5toR2RA

1

3

3

0

129

UMAILab @umaidotlab

almost 2 years ago

We want to be a part of this transformation process without moving away from ethical values and offer alternatives against monopolization. Because the monopolization/cartelism of AI technology has the potential to open the door to a dystopia for humanity.

0

0

0

0

12

UMAILab @umaidotlab

almost 2 years ago

Hello everyone! What is UMAI Lab? UMAI: Unified Methods of Artificial Intelligence It is an AI R&D Laboratory established with the awareness that Artificial Intelligence is a multi-disciplinary technology.

1

1

0

0

16

UMAILab @umaidotlab

almost 2 years ago

The most important thing is that after a certain period of time, people will mostly move away from business life because of AI and Robots. This requires a transformation process.

1

0

0

0

13

UMAILab @umaidotlab

almost 2 years ago

Bu dönüşüm sürecinin etik değerlerden uzaklaşmadan bir parçası olmak ve tekelleşmeye karşı alternatifler sunmak istiyoruz. Çünkü YZ teknolojisinin tekelleşmesi insanlık için bir distopyanın kapısını açma potansiyeli vardır.

0

2

0

0

26

UMAILab @umaidotlab

almost 2 years ago

Herkese merhaba! UMAI Lab nedir ? UMAI: Unified Methods of Artificial Intelligence Lab Yapay Zeka’nın multi disipliner yapıda gelişen bir teknoloji olduğunun farkında olarak kurulan bir YZ ArGe Laboratuvarıdır.

1

3

1

0

101

UMAILab @umaidotlab

almost 2 years ago

En önemlisi de belli bir süre sonra insanların YZ ve Robotlar sayesinde iş hayatından büyük çoğunlukla uzaklaşacak olmasıdır. Bu bir dönüşüm süreci gerektirmektir.

1

2

0

0

29

Last Seen Users on Sotwe

Trends for you

Most Popular Users