LLMS

Bypassing LLM and Embedding Model Weaknesses with Open Source Solutions

By

–

01 July 2023 17h18

I'm not aware of a startup working on this directly, but people are talking about the problem. And it's an interesting way to bypass some issues with weaknesses that embedding models/existing LLMs have, especially if there's broader adoption of OSS LLMs.

Prompt Routing: Selecting Optimal Models for Different Prompts

By

–

01 July 2023 17h17

I've also heard *very* recently (like, the past few weeks) about interest in what you could call Prompt Routing. The idea being that different models hallucinate in different ways, and there could be an opportunity to pick the right model for the right prompt algorithmically.

AI Frameworks as Business Models: LangChain and AutoGPT

By

–

01 July 2023 17h13

Investors and industry professionals are still very split on how frameworks are going to play out as businesses. LangChain is the obvious one here, but there are a lot of other emerging frameworks w/ companies forming around them. AutoGPT comes up a lot in this context.

LLM Benchmark Debate: Comparing Model Performance Standards

By

–

01 July 2023 17h12

That's led to this very weird debate over how to explicitly benchmark how these perform against each other. There's no great consensus on how to compare one against each other, and many (like Falcon 40B) are using leaderboards as their selling point.

Meta’s LLaMA Supercharges Open Source LLM Ecosystem Growth

By

–

01 July 2023 17h11

First, the obvious: the explosion of OSS models. Meta supercharged the OSS LLM ecosystem with the launch of LLaMA, which has both powered _and_ inspired a new generation of OSS models like MPT, INCITE, etc—and many startups building them like @MosaicML and @togethercompute
.

Transformers Paper Anniversary: Six Years of LLM Progress

By

–

01 July 2023 17h10

This also coincides roughly (June 12, to be exact) with the six-year anniversary of the release of the seminal Transformers paper, which made all of the developments in LLM possible in the first place—even if we're starting to run into some of the limitations it presents.

→ View original post on X — @deeplearn007

LLM Pruning: Scaling Edge AI with Reduced Energy Footprint

By

@deeplearn007

–

01 July 2023 15h44

Excellent work on pruning LLMs! This is where the next generation of #AI LLM models need to go in order to scale across the edge of our networks on devices, reducing energy consumption and carbon footprint.

→ View original post on X — @yitayml

Scaling Challenges in Universal Transformer Parameter Sharing

By

@yitayml

–

01 July 2023 13h06

It's hard to scale UT in terms of param sharing because of the shared params across all layers (similar to ALBERT). If you want to have like a 1B UT model, it's going to be super slow. That said, we didn't try scaling non-shared UT which could be okay. More about

→ View original post on X — @bamitav

Managing Generative AI Risks: Ethics and Safety

By

@bamitav

–

01 July 2023 12h38

Managing the Risks of Generative AI https://
hbr.org/2023/06/managi
ng-the-risks-of-generative-ai
… #GenAI #ChatGPT #BingAI #ChatGPT #opensource #EthicalAI #Python #tech #chatgpt4 #AI #ML #AIEthics #chatgpt3 #code #GPT3 #GPT4 #GPT4AI #aichat #chatbot #ChatbotAI #BardAI #chatgpt5 #LLMs #LLM #GenerativeAI #GoogleBard