Thank you for highlighting several projects that we incubated and launched including @nvidia Earth-2, @nvidia Modulus and genome-scale language models with @argonne
MACHINE LEARNING
-
GPQA High Variance Issues and Evaluation Methodology Discussion
By
–
I believe we're running 4 repeats and taking the average, and also using the same zero-shot CoT prompt as before, see https://
github.com/openai/simple-
evals
… But i totally agree with you that GPQA is high variance, I wish there were like 3k examples instead of 300 -
VLMs Emotional Knowledge: Beyond Simple Smile Detection
By
–
Emotion recognition ≠ social signal detection – smiles don't always mean you are happy. But, my experience with VLMs suggests there's a lot of latent emotional knowledge and reasoning thanks to all the ingested novels. More testing and demos needed!
-
LLMs: Retrieval Capabilities vs. Reasoning Limitations
By
–
The always insightful Melanie Mitchell @melmitchell1 on what LLMs are actually doing (really good retrieval) and not doing (reasoning). https://
aiguide.substack.com/p/evaluating-l
arge-language-models
… -
Transformer Architecture: Foundation of GPT and Modern AI
By
–
The technology underlying GPT and LLMs. I was building Transformer based models and writing about them in 2020. Transformers have been revolutionising natural language and impacting computer vision and time series and will increasingly be going multimodal.
-
Faster AI Model Optimization and Scaling Potential Beyond GPT-4
By
–
This take feels very off. With the speed of this model, I’d expect it’s close to an order of magnitude smaller than older GPT-4 models. Take the learnings and approaches from building this, scale them up, and you have something way better than GPT-4.
-
What Made GPT-4 Seem Impossible at Release?
By
–
what about gpt4 on release felt impossible? i mean you ended up doing it ~1 year later so you figured it out somehow. just wanted to learn from benefit of hindsight
-
GPT-4o Launch: Major Leap Forward in AI Model Capabilities
By
–
Todo el jaleo de los modelos GPT2 en el arena era para tener testeado para el día de hoy a GPT-4o y wow. El modelo es un salto (literal) hacia adelante frente a modelos GPT-4 anteriores y competencia. Importantísimo también las mejoras en programación
-
AI model selection and management in workflow
By
–
It is located in the same selector along with other models
-
One Shot Stable Diffusion Fine-tuning In Context
By
–
5/ effectively one shot stable diffusion finetuning, in context!?