Yeah I’d suspect understanding English grammar takes up some reasoning capacity that could assigned elsewhere in a simpler language
RESEARCH
-
Qwen vs Llama 3: Programming and Math Capabilities Comparison
By
–
Efectivamente el punto fuerte en el que hacen hincapié desde Qwen es en las capacidades de programación y matemática frente a Llama 3.
-
New 7B Model Outperforms Llama 3 8B in Programming
By
–
También el modelo 7B que la mayoría de vosotros utilizaréis también tiene muy buena pinta, con mejor rendimiento que Llama 3 8B -sobre todo en programación y matemáticas-. Si acostumbrabais a usar el modelo 8B de Meta, pues ahora tenéis un buen upgrade que probar 🙂
-
Language Character Systems Impact Model Training Properties
By
–
I feel like model training on languages where the individual characters have meaning (eg chinese) would have different properties than on languages with purely phonetic alphabet, but i’m not sure what
-
Designing Token Efficient Language Model Architecture
By
–
brb designing a new token efficient language
-
AI breakthroughs in antibiotics, audio models, and workplace tools
By
–
Top stories in AI today: -AI predicts nearly 1M new antibiotics
-Stability AI releases open audio model
-How to use Perplexity Knowledge Cards
-Asana debuts AI teammates
-6 new AI tools & 4 new AI jobs Read more: http://
therundown.ai/p/ais-antibiot
ic-breakthrough
… -
INT4 Model Achieves Zero Accuracy Loss Breakthrough
By
–
"Zero accuracy loss of INT4 model" super cool!
-
Adversarial Techniques Work Across Languages and Model Architectures
By
–
Yes, I can confirm that it also works in other languages (tested in French with NeuralDaredevil). I haven't tried to apply it to MoE models but it should work too. It may be trickier to choose a refusal direction because there are more blocks.
-
Glaze and Nightshade: Evidence effectiveness and ethical communication dilemma
By
–
Esto es así, y la efectividad de técnicas como Glaze o Nightshade no están sustentada por más evidencia que la de los papers de quienes la promueven. Y aquí el dilema está en si es mejor avisar ante posibles falsas expectativas, o dejar que medidas como estas sirvan para
-
Microsoft GigaPath: Vision Transformer for Cancer Diagnostics
By
–
@Microsoft introduces GigaPath, a pioneering vision transformer for cancer diagnostics! In collaboration with @UW and Providence Health Network, GigaPath utilizes dilated self-attention for whole-slide modeling, pre-trained on over 1 billion 256×256 pathology image tiles from