To follow up, I had a chance to try it btw:
before: 212ms / iter
>>> model = torch.compile(model)
after: 135ms / iter
nice 🙂
MACHINE LEARNING
-
PyTorch Compile Optimization Reduces Iteration Time by 36%
By
–
-
AI Model Conditioning: Prompting vs. Fine-Tuning Techniques
By
–
Generally I’d call that “conditioning the model”, which can take the form of either prompting or fine-tuning, but not sure if that’s exactly the right term. They’re two different techniques to make it respond how you want.
-
Fine-tuning for AI Specialization and Performance
By
–
Not sure I understand fully re: alignment, but fine-tuning is usually the most performant way to specialize when you have the data for it.
-
Improved prediction function for broader interest and impact
By
–
I think i have developed a better prediction function of what is likely to be of broader interest or higher impact, but I can't be 100% sure obviously
-
GPT-3 Inference Non-Determinism at Temperature Zero Explained
By
–
An edge-case in GPT-3 with big implications: Inference is non-deterministic (even at temperature=0) when top-2 token probabilities are <1% different. So temperature=0 output is *very close* to deterministic, but actually isn't. Worth remembering.
-
Self-Review of Research Work: Fulfilling Projects and Limitations
By
–
New blog post, self-reviewing my research over the past few years: https://
jasonwei.net/blog/research-
i-enjoy
… In addition to work I found fulfilling, I also critique why I felt some papers I've written didn't go so well, or were inherently limited. Apologizes in advance for hot takes -
AI Today Podcast: Content Summarization and Sentiment Analysis Episode
By
–
In this @Cognilytica #AIToday #podcast episode 'Content Summarization and Analysis, Sentiment Analysis' from the AI Glossary Series, hosts @rschmelzer & @kath0134 define Content Summarization and #Analysis and #sentimentanalysis. Full episode: https://
cognilytica.com/2022/12/28/ai-
today-podcast-ai-glossary-series-content-summarization-and-analysis-sentiment-analysis/?utm_source=dlvr.it&utm_medium=twitter
…
#AI #nlp #ML -
PaLM 540B Architecture Code Released Open Source
By
–
Mola, pero… ¿quién tiene capacidad de entrenar un PaLM de 540 billions (x3 el tamaño de GPT-3)? Aquí están compartiendo el código de la arquitectura y no un modelo pre-entrenado. Con lo que es un entrénalo-tú-mismo de ChatGPT.
-
Invitation to learn embeddings to understand AI models
By
–
What you are saying is fundamentally wrong. I invite you to get closer to the concept of embeddings, to start your learning, to understand a bit more how information is represented in certain AI models. https://towardsdatascience.com/neural-network-embeddings-explained-4d028e6f0526 …
-
DeepMind GraphCast Beats World’s Best Weather Forecasting System
By
–
DeepMind & Google’s ML-Based GraphCast Outperforms the World’s Best Medium-Range Weather Forecasting System