ML for Curious
Open in Telegram
Here we eat ML for breakfast !
Show moreThe country is not specifiedTechnologies & Applications14 713
6 999
Subscribers
No data24 hours
No data7 days
No data30 days
Posts Archive
6 999
Submitted a PR on SD V1.5/V2 fine-tunning with diffusers. Regardless of the outcome, hope somebody will find it a helpful starting point
https://github.com/huggingface/diffusers/pull/6922
6 999
very sloppy work. It's better not to write a paper, rather than doing it this way to show competitiveness:
https://diffuse2choose.github.io/
6 999
very sloppy work. It's better not to write a paper, rather than doing it this way to show competitiveness:
https://diffuse2choose.github.io/
6 999
This new interactive guide, created by research teams at Meta, was designed to cover prompt engineering and best practices to help developers, researchers and enthusiasts working with large language models produce stronger outputs.
https://github.com/facebookresearch/llama-recipes/blob/main/examples/Prompt_Engineering_with_Llama_2.ipynb?utm_source=linkedin&utm_medium=organic_social&utm_campaign=llama&utm_content=video
6 999
New Year Resolution(2/2):
Technique-wise:
- How general-purpose is the method? Novelty is one thing, but general applicability is way more important.
- Does it follow the Bitter Lesson, i.e. scales well as the amount of compute scales?
- Is the idea simple and elegant? A pipeline with 10 modules almost certainly has no staying power.
- How future-proof? Does your algorithm's performance rise as the base models rise, or is it obsolete in a few months?
- Please don't solve toy problems (e.g. gridworlds for RL) and make grand claims. Just don't. It's fine in 2018, but that ship has sailed.
Author: https://www.linkedin.com/in/drjimfan/
6 999
New Year Resolution (1/2):
write fewer papers. When you feel the urge, look at Bill Freeman's chart below, and resist it. Instead, ask yourself:
Impact-wise:
- How many people does your paper inspire, both within the field and to a broader audience?
- 2023 saw the epic rise of AI hackers. AI is no longer an academic niche. How can we best service this new, augmented community? I'm much happier to see 100x more people building on my work, than 5 satisfied reviewers.
- How accessible is the research? I'm a big advocate of open-sourcing, but I understand that companies have their legal and business concerns. How to navigate hurdles and negotiate accessibility are now part of the job.
6 999
A great summary of Pinterest's talk at Qcon 2012 on how they scaled 11.7 million monthly active users with just 6 engineers, highlighting key lessons for software practitioners.
This journey emphasises the use of proven technologies, where Pinterest transitioned from a complex mix of newer technologies to a more streamlined architecture centered around MySQL and Memcached.
Key strategies included simplicity in design, manual database sharding over clustering to ensure better load balancing and high availability, and a focus on efficient data structuring.
This approach underscores the importance of proven, simple solutions in rapidly scaling tech platforms that are built through minimal engineering resources.
https://read.engineerscodex.com/p/how-pinterest-scaled-to-11-million
6 999
I know I have promised to give more updates from NeurIPS - but the amount of information is just overwhelming. I would need to compile everything into a concise reading list. Stay tuned
6 999
Cybersecurity team from Meta have found that the more LLM becomes proficient in code writing the more insecure code it writes. There is a clear negative correlation between code quality and code security.
6 999
https://about.fb.com/news/2023/12/purple-llama-safe-responsible-ai-development/amp/
Today they are releasing also the Llama safeguard - to put guardrails on the inputs/
prompts and outputs when the model is deployed to production
6 999
walkthrough of LLama, langchain and together.ai from Meta:
https://github.com/amitsangani/Llama-2/blob/main/Building_Using_Llama_Together.ipynb
6 999
Q2 2024 - Hardware that is specialised on purely on transformer architectures.
2.3x faster, 1/2 the cost of Nvidia DGX H100
6 999
Engineers from Qualcomm demonstrated quantization+knowledge distillation in two applications which are actually working:
1. Stable diffusion on the device
2. Privacy aware segmentation model training on the device
6 999
Fellows from Amazon use this technique to quantize models. It is also available on the Amazon Sagemaker
https://github.com/mit-han-lab/smoothquant
