Meta sees “an opportunity to introduce AI agents to billions of people in ways that will be useful and meaningful,” CEO Mark Zuckerberg told investors today. [link]
{autonomus, custom, local} agents // own infra, AI models can work locally
Funny guys play football // deepmind
Deploy open source LLM on custom infra
Dolly v2 was trained by Databricks, has 12B parameters and an MIT license.
Reasons to deploy own LLMs:
- Control, customization and ownership // no limits of big corps LLMs
- Data // privacy
- Ownership // you can sell/buy custom models/data
- Cost // optimize spends
Where to find LLMs? huggingface, replicate, github, twitter
Self-Instruct: Aligning Language Model with Self Generated Instructions
StableVicuna
How to build *TruthGPT*?
Self-hosted LLMs
Large language models (LLMs) generally require significant GPU infrastructure to operate. We’re now starting to see ports, like llama.cpp, that make it possible to run LLMs on different hardware — including Raspberry Pis, laptops and commodity servers. As such, self-hosted LLMs are now a reality, with open-source examples including GPT-J, GPT-JT and LLaMA. This approach has several benefits, offering better control in fine-tuning for a specific use case, improved security and privacy as well as offline access. However, you should carefully assess the capability within the organization and the cost of running such LLMs before making the decision to self-host. [link]
Some Models for Commercial Usage
Large Language Model (LLM) Primers
With the advent of ChatGPT, LLMs have been the talk of the town! We’ve recently seen a bunch of extraordinary advancements with GPT-4, LLaMA, Toolformer, RLHF, Visual ChatGPT, etc.
📝 Here are some primers on recent LLMs and related concepts to get you up to speed:
🔹 ChatGPT (http://chatgpt.aman.ai)
- Training Process
- Detecting ChatGPT generated text
- Related: InstructGPT🔹 Reinforcement Learning from Human Feedback a.k.a. RLHF
(http://rlhf.aman.ai)
- Refresher: Basics of RL
- Training Process (Pretraining Language Models, Training a Reward Model, Fine-tuning the LM with RL)
- Bias Concerns and Mitigation Strategies🔹 LLaMA (http://llama.aman.ai)
- Training Process (Pre-normalization, SwiGLU Activation Function, Rotary Positional Embeddings, Flash Attention)
- Visual Summary🔹 Toolformer (http://toolformer.aman.ai)
- Approach
- Sampling and Executing API Calls
- Experimental Results🔹 Visual ChatGPT (http://vchatgpt.aman.ai)
- System Architecture
- Managing Multiple Visual Foundation Models
- Handling Queries
- Limitations🔹 GPT-4 (http://gpt-4.aman.ai)
- Capabilities of GPT-4
- GPT-4 vs. GPT-3Notes written in collaboration with Vinija Jain. [link]
libraries for agents: LangChain MiniChain, examples
Your model is only as good as the data that goes into it.
