looks like openai is getting a design overhaul 
somethings cooking
https://x.com/bwalkin/status/1762182300769653057?s=46

✔️ 

https://github.com/Mozilla-Ocho/llamafile/releases/tag/0.8.10

you know what impresses me? making a small model work well.

undertrained behemoth…but why? i’ll have to dive into this one later ☺︎

Tele-FLM-1T an open LLM released by BAAI and TeleAI with 1 trillion parameters.

- supports chinese and english
- apache 2.0 license
- cost-effective progressive pre-training
- enhanced with input/output scalers, RoPE, RMSNorm, and SwiGLU

undertrained behemoth…but why? i’ll have to dive into this one later ☺︎

Tele-FLM-1T an open LLM released by BAAI and TeleAI with 1 trillion parameters.

- supports chinese and english
- apache 2.0 license
- cost-effective progressive pre-training
- enhanced with input/output scalers, RoPE, RMSNorm, and SwiGLU

https://huggingface.co/CofeAI/Tele-FLM-1T

the intelligence explosion concept doesn’t necessarily hinge on achieving AGI.

i’m leaning towards AGI being defined as an algorithm capable of proving theorems in a proof assistant with the same expertise as a skilled human.
few.

food for thought  
https://warpcast.com/gm8xx8/0x13fafc1d

SearchGPT

https://openai.com/index/searchgpt-prototype/

it’s no surprise that OpenAI is getting into search.

the rumors were spot on, and the publisher partnerships gave it away.

😈

let the search wars begin!

this is a significant breakthrough, marking the first successful use of MCTS with LLMs for theorem proving.

😈

this is a significant breakthrough, marking the first successful use of MCTS with LLMs for theorem proving.

😈

https://deepmind.google/discover/blog/ai-solves-imo-problems-at-silver-medal-level/

strong visual. 

(note: missing deepseek and only mmlu)

1st time we see an open source model in the top spot on EQ-Bench

1st time we see an open source model in the top spot on EQ-Bench

https://eqbench.com

sqlite-lembed: A SQLite extension for generating text embeddings locally

i’ll leave this here…

https://alexgarcia.xyz/blog/2024/sqlite-lembed-init/index.html

sqlite-lembed: A SQLite extension for generating text embeddings locally

i’ll leave this here…

Mistral Large 2 features:

- 128k context window
- supports 80+ coding languages
- 123 billion dense parameters
- matches Llama 3 405B in coding with 1/3 the parameters
- superior function calling compared to GPT-4o and Claude
- released under the Mistral Research License

and it’s only wednesday 🔥 

https://mistral.ai/news/mistral-large-2407/

(more links) 

Mistral Large 2: Quality, Performance & Price Analysis: https://artificialanalysis.ai/models/mistral-large-2

Mistral Large 2: API Provider Benchmarking & Analysis: https://artificialanalysis.ai/models/mistral-large-2/providers

Mistral Large 2: 

- 123B parameters
- 128k context window
- natively multilingual
- strong code & reasoning
- SOTA function calling
- open weights for non-commercial use

its also smaller and cheaper to host than models like Llama 405B. 

looks like this continues Mistral’s trend of smaller, specialized models.

🐵 ← 🙈🙉🙊 → phil.cockfield.net, db.team →  ƒ(trust / time) →  ƒ(open-commons) → ƒ( /sys )  ← @cell primitive<T> 🧫

llama 3.1’s license is similar to previous versions, one key difference: it allows model outputs to be used for improving other LLMs, enabling synthetic data generation/distillation w/ different models—while it permits redistribution, fine-tuning, &  creating derivatives, it requires derived models to include “Llama” in their name and mention “Built with Llama” in any services. also any derivatives, including models & datasets, must be distributed under the llama 3.1 license. 
https://huggingface.co/meta-llama/Meta-Llama-3.1-8B-Instruct/blob/main/LICENSE companies with over 700 million active users at the time of release cannot use the model. 
(great for individual use not so much for large organizations) 

also adding this link from nathan worth the read and i agree its better but not perfect (kind of expected from meta): https://x.com/natolambert/status/1815768837926842842?s=46

who wants to see Mistral-123B evaluated on SEAL against Llama 3.1? 

🙋‍♂️

mistral large 2 model checkpoints

https://huggingface.co/mistralai/Mistral-Large-Instruct-2407

Mistral Large 2 features:

- 128k context window
- supports 80+ coding languages
- 123 billion dense parameters
- matches Llama 3 405B in coding with 1/3 the parameters
- superior function calling compared to GPT-4o and Claude
- released under the Mistral Research License

and it’s only wednesday 🔥

run sota at home, 2 MacBooks is all you need!

https://x.com/ac_crypto/status/1815969489990869369

☺︎

https://warpcast.com/swabbie.eth/0x7c10c023

𝚐𝔪𝟾𝚡𝚡𝟾

@gm8xx8