AI Everyday

Matt Wallace talks about the latest OpenAI announcements regarding the release of GPT-4, its improved speed, cost, and performance compared to other AI models. He explores features like Elo ratings, real-time interactions, and potential applications in coding and vision processing.  He also mentions the upcoming release of a desktop app and the free availability of GPT-4.

Matt talks about the big things this week - Gemini 1.5 Pro with 1m token context, and OpenAI text-to-video model Sora, which is generating incredible high-quality videos up to 1 minute long.

7 wild updates from this week in about 8 minutes. #AI moving at crazy speed!

Hands on and discussion around vLLM, high performance inference engine supporting continuous batching and paged attention.

Matt reviews bootpig, a paper/model that provides DreamBooth-style modification of images without fine tuning.

Matt discusses his results from testing Phind-CodeLlama-34B-v2, which beat the GPT-4 zero-shot HumanEval test by a good margin.

Matt discusses Llama2 with GPTQ quantization, which is much more powerful than previous methods of quantizing model weights, and demos text-generation-webui.

Matt gives ChatGPT access to root on a Linux box. It can read and write files and run commands. He tells it to give itself long term memory, and it installs weaviate as a vector db. What's next?!

Matt discusses the paper "How Is ChatGPT’s Behavior Changing over Time?" and the crazy headlines resulting from its publication, that are misleading a lot of folks.

Discussing Llama 2 - where does it stack up, and what does it mean for open LLMs?

On this episode, Matt reviews the ChatGPT Code Sandbox, which just rolled out to ChatGPT Plus subscribers. He goes through a half dozen example use cases from bulk text analytics of PDFs to rapidly adding to a flask app.

Matt discusses databricks acquisition of MosaicML, the new Falcon-40B model atop the HF OSS LLM leaderboard, and some resources.

In this episode of AI Everyday, Matt reviews some code and output from a simple Langchain application that uses 2 prompts in a map-reduce chain to digest 3+ hours of podcast transcripts into pages of high-level notes, then digest those notes to 20 bullets.

Matt discusses ChatGPT Functions, their relationship to plugins on the chatgpt web service, and how functions in the API can be used as a hack to get the GPT API to constrain its output to a tightly fixed format.

Matt discusses nVidia blowing up after raising expectations a ton, and about the future of AI in general. It's longer, but it's a must-hear.

Matt discusses a paper quantifying emergent behaviors in LLMs, and looking at Microsoft CodeT and how it applies to general use of LLMs at an app layer.

AutoGPT, Dolly v2, AWS Bedrock, Vicuna

In this podcast, Matt discusses an incredible paper from Stanford that introduces a tiny virtual world with 25 generative agents that can come up with realistic behavior. The agents use GPT as a brain and have short-term and long-term memory. They converse with each other and propagate information, changing their intent. The implications for simulated worlds and gaming are profound, and the paper is definitely worth checking out.

Matt plays around and demonstrates meta's new Segment Anything model, which can zero-shot segment just about anything in an image.