Want to learn how BloombergGPT was built? Watch Snorkel AI's Alex Ratner and Gideon Mann, Head of Machine Learning Product and Research at Bloomberg discuss the challenges and triumphs of building a doman specific Language Model. https://
buff.ly/3TXflT0
LLMS
-
Building BloombergGPT: Domain-Specific Language Model Development
By
–
-
Groq and Llama3: 24 Hours of Development Updates
By
–
24 hours for Groq and #Llama3. Read more about today's developments at https://
groq.link/llama3blog. -
400B Model Expected to Surpass GPT-4 Performance
By
–
they were “only” trying to match gpt4, and now 400b is likely to beat it
-
llm-gpt4all 0.4 Release Notes Published
By
–
Release notes for llm-gpt4all 0.4 are here: https://
github.com/simonw/llm-gpt
4all/releases/tag/0.4
… -
Llama 3 8B Now Available via llm-gpt4all Plugin
By
–
I upgraded the llm-gpt4all plugin to support running Llama 3 8B Instruct (thanks, @nomic_ai
) llm install –upgrade llm-gpt4all
llm -m Meta-Llama-3-8B-Instruct 'Write Python code to print 5 great names for a pet pelican' The quantized model is a 4.34GB download, needs 8GB RAM -
Handling Chatbot Self-Help Queries with RAG and Prompting
By
–
If you ship a chatbot, it's now indisputable that people are going to ask it questions about how to use it Handling this isn't trivial but it's not unsolvable either – use RAG, fine-tuning or a meticulous system prompt, it's important to anticipate this use-case
-
GPT-2 Activation Memory and GPU Cache Analysis
By
–
Makes sense, in GPT-2 (124M) case we're currently doing B=4, T=1024, C=768 => 3M activations @ float32 => 12MB. A100 L2 cache is 40MB, and even L1, at 192KB/SM with 108 SMs => ~= 20MB (wow, that's more than I expected). The pleasures of smaller networks and caches…
-
Anthropic increases Haiku tokens amid API provider competition
By
–
5-10x times more Haiku tokens per day from @AnthropicAI – hard not to assume this is in reaction to the flood of new API providers selling cheap access to Llama 3
-
Flow Engineering Webinar Recording with LangChain Leaders
By
–
The recording from our Flow Engineering webinar with @hwchase17 and @itamar_mar is up! https://
youtube.com/watch?v=eBjxz7
qrNBs
… "Flow Engineering" is a term that has been gaining in popularity recently. The first time it was mentioned as term was in CodiumAI paper on AlphaCodium, where they -
Kernel optimization attempts yield minimal performance gains
By
–
added under kernel4 https://
github.com/karpathy/llm.c
/commit/cb791c4ef58d45d58e5af624b0ed41439ac7aeff
…
a bit surprised to only see ~1-2% out of it, which then washes out in training, as the layernorm is not a top-ranking time kernel. Also tried float4 and unrolling but that didn't improve it too much bleh