TLDR AI 2023-11-22

OpenAI is welcoming Sam Altman and former President Greg Brockman back following the boardroom coup, a move seen by investors as a win for the firm's mission. This shake-up includes a new board aimed at resetting OpenAI's governance and a planned investigation into the recent turmoil. Both Microsoft and Altman wish to occupy seats in an expanded board.

Microsoft has announced Windows AI Studio, a hub for developers to access and configure AI models, with tools from Azure AI Studio and Hugging Face and features like model configuration UI and performance testing templates. Nvidia updated its TensorRT-LLM for PCs with GeForce RTX 30 and 40 Series GPUs, enabling local running of large language models and enhancing privacy.

Google has postponed the launch of its large language model (LLM) and GPT-4 competitor Gemini until Q1 of 2024 due to concerns about its ability to match or surpass OpenAI's GPT-4 and a focus on strengthening its consumer offerings before allowing external developer access.

Trained to generate 14 frames of video, Stable Video from Stability AI is a glimpse into the future of content creation.

One of the innovations of the original Stable Diffusion model was the use of a smaller latent space for diffusion. This means the diffusion doesn’t happen on pixels but on some compressed image representation. This post dives into some ways to interpret that space for SDXL.

There are many ways to speed up autoregressive language model inference. One way that people are excited about is using a draft model. This requires two models but it can be slower. However, by generating related ngrams from the same model, you can alleviate the need for a draft model and speed up generation linearly.

It’s now possible to easily integrate language models into your Go programs with LangChain. This includes running local models with frameworks like Ollama.

You can very quickly deploy an endpoint that serves HuggingFace-hosted models using llama.cpp and GGUF server.

Developing LLM applications is one thing, deploying them successfully in production is another. This post cuts through the complexity, offering a developer's guide to maximizing accuracy as you transition your LLM applications from the dev environment to the rigors of production.

ChatGPT with voice is now available for all free users. OpenAI released it with a joke about its CEO drama.

OpenAI’s dominant place in the developer ecosystem is suddenly at risk, presenting an opportunity for smaller companies to fill a new void. Microsoft will likely walk away as the AI winner, but perhaps other model-layer companies like Anthropic will take advantage of the disturbance.

A study published in the journal Science revealed that Google DeepMind's AI meteorology model, GraphCast, significantly outperforms traditional weather forecasting methods, accurately predicting global weather conditions up to 10 days ahead.

Amazon is gearing up to enter the AI race with Olympus, a large language model boasting 2 trillion parameters, aiming to surpass OpenAI's GPT-4 and Google's Bard in capability.

Interim OpenAI CEO Emmett Shear is reportedly threatening to leave his position at OpenAI if the board can't show evidence of Sam Altman’s wrongdoing.

Reverb delivers AI-powered but hallucination-free answers to online forms like RFPs and questionnaires with training from past completed forms.