LLMS

–

25 January 2025 23h01

Custom backends now support reasoning tokens in the latest version of Apollo. Just change the model type in settings.

DeepSeek R1 7B: Powerful Reasoning AI on Local Networks

By

–

25 January 2025 23h01

Here’s deepseek r1 7b (qwen) beaming from my laptop to my phone at 60 tok/s.

Wild that you can serve such powerful reasoning AI models to your whole network with just a few apps. pic.twitter.com/RTHnrjmzRh
— Aaron Ng (@localghost) 25 janvier 2025

Here’s deepseek r1 7b (qwen) beaming from my laptop to my phone at 60 tok/s. Wild that you can serve such powerful reasoning AI models to your whole network with just a few apps.

→ View original post on X — @goodside

RL Model Achieves Superhuman Snark on Human Cope Chains

By

@goodside

–

25 January 2025 19h41

applying large-scale RL to chains of cope written by human seethers, we observe an “lmao moment” upon which the model spontaneously exhibits superhuman snark

Future Mobile AI Models: Shift Toward Specialized Smaller Systems

By

–

25 January 2025 19h16

near term future for mobile models might be many small specialized small ones

Apollo Adds Text-to-Speech Support in Voice Mode

By

–

25 January 2025 17h58

apollo has some tts support in voice mode

→ View original post on X — @mjasay

LLMs Training Data Dependency: Stack Overflow Risk

By

@mjasay

–

25 January 2025 16h38

@GergelyOrosz Given the LLMs' dependence on training data from Stack Overflow (and other Q&A forums), what happens when SO dies and that information source dies with it? How do LLMs replenish their sources?

→ View original post on X — @sakanaailabs

Sakana AI Proposes Transformer² with Two-Stage Reasoning

By

@sakanaailabs

–

25 January 2025 15h48

東京発Sakana AIが新たな「Transformer」を提案 https://
xtech.nikkei.com/atcl/nxt/colum
n/18/02801/012200014/
…
Transformer²は2段階の推論プロセスを採用している。
１段階目では、プロンプトを試験的に実行してAIモデルの挙動を観察し、実行に必要なスキルを理解する。

→ View original post on X — @soumithchintala

DeepSeek AI breakthrough: genuine innovation or geopolitical psyop?

By

@soumithchintala

–

25 January 2025 15h47

What's more likely? 1 – small group of AI engineers at @deepseek_ai figures out how to beat all of the top researchers in the world as a side project 2 – Chinese government has 100k GPUs they shouldn't have and releases open source models claiming $6m training cost as a psyop