. .AI: Altman's back, DeepMind finds 800 years worth of crystals, Microsoft super prompt and more (11.29.23)

Friends, today’s newsletter is really interesting. Altman’s blog post about his return is the top story, but there are several other stories here that will leave you inspired and empowered.

These are the stories the AI community is talking about most today, based on a weighted analysis of community engagement. Big thanks to those of you sharing this newsletter with friends. I had a great phone call with someone just today who’d been forwarded a recent issue!

Ok, here’s today’s news…

Marshall Kirkpatrick, Editor

Altman Reinstated as OpenAI CEO, Murati Returns as CTO

First impacted: OpenAI customers, OpenAI employees
Time to impact: Short

Sam Altman has resumed his position as CEO of OpenAI, with Mira Murati returning as CTO, according to a blog post Altman published tonight. Altman identified three immediate priorities for OpenAI: progressing their research plan and safety investments, enhancing and deploying products for customers, and establishing a diverse board to strengthen governance. Despite difficulties, OpenAI did not lose any employees or customers over the recent period, Altman said. [Sam Altman returns as CEO, OpenAI has a new initial board] Share by email

Google DeepMind Discovers 2.2 Million New Crystals

First impacted: Material scientists
Time to impact: Short

Google DeepMind writes: "Modern technologies from computer chips and batteries to solar panels rely on inorganic crystals. To enable new technologies, crystals must be stable otherwise they can decompose, and behind each new, stable crystal can be months of painstaking experimentation. Today, in a paper published in Nature, we share the discovery of 2.2 million new crystals – equivalent to nearly 800 years’ worth of knowledge. We introduce Graph Networks for Materials Exploration (GNoME), our new deep learning tool that dramatically increases the speed and efficiency of discovery by predicting the stability of new materials." [Millions of new materials discovered with deep learning] Share by email

Powerful Prompting Strategy Yields Superior Medical Answers in GPT-4

First impacted: AI researchers
Time to impact: Medium

Microsoft has published a study that finds the general AI model GPT-4 can outperform specialized models in answering medical questions, if prompted well. They came up with a really sophisticated prompting strategy they say could be generalized into other contexts. Called MedPrompt, it combines (1) dynamic selection of training examples that are most similar to the specific question being asked, to narrow down the LLM's focus to high-relevance examples, (2) GPT-4-created Chain-of-Thought ("think this through step by step") explanations for answers in the training data set, which are then provided as examples in the new prompt, and (3) shuffling the order of choices in the prompt across multiple versions, then selecting the answer that is chosen and explained most consistently even when the questions are shuffled. Wow! I think I understood that correctly. Let's try it! :) [arxiv.org] Share by email

Perplexity Launches Online LLM APIs

First impacted: Developers Using AI
Time to impact: Short to Medium

Perplexity Labs has launched two new online large language models, pplx-7b-online and pplx-70b-online, with the aim of providing accurate and current information by making these LLMs fundamentally tied to the live internet. Perplexity's own mobile app has become my go-to search engine, though I continue to use GPT-4 and Claude for other purposes like conversations and data analysis. [Introducing PPLX Online LLMs] Share by email

New Course Launched on Advanced RAG Techniques

First impacted: Developers Using AI
Time to impact: Medium

Jerry Liu and Anupam Datta, co-founders of LlamaIndex and TruEra, have launched a free short course on Retrieval Augmented Generation (RAG) techniques, which enable search against an organization's own data. It covers advanced retrieval methods, evaluation best practices, experiment tracking, and the RAG triad: Context Relevance, Groundedness, and Answer Relevance. [Building and Evaluating Advanced RAG Applications] Share by email

"4D-fy" Method Advances Text-to-4D Generation

First impacted: App Developers
Time to impact: Medium

A new technique called "4D-fy: Text-to-4D Generation Using Hybrid Score Distillation Sampling" has been published that creates 4D scenes: impressive visuals with 3D structure plus movement. The method uses an alternating optimization process that combines supervision signals from several pre-trained diffusion models. [4D-fy: Text-to-4D Generation Using Hybrid Score Distillation Sampling] Share by email

UAE AI Minister Warns Against Overregulating AI

First impacted: AI regulators, AI technology developers
Time to impact: Long

Omar Al Olama, the UAE Minister of AI (reportedly the world's first), warned about the risks of excessive AI regulation, drawing parallels to the Ottoman Empire's decline following the early ban on the printing press in 1515. Meta's Yann LeCun picked up his comment on X and a debate ensued. [via @ylecun] Share by email

That’s it! More AI news tomorrow.