Parse notes
How do you align a model much smarter than humans? OpenAI has been working on this question across teams for some time, but now they're bringing 20% of the entire org (compute-wise) to bear on this challenge. They're hiring extensively and many top scientists internally are working on this problem.
The popular AI powered image editing suite raised money to continue pushing forward text driven creativity expressed via pixels.
It seems like dilated attention now finally works! This is an exciting advancement if it holds at scale, because it means that we can essentially treat entire corpi of documents as context. All of the Python code on GitHub is only a few hundred billion tokens. LongNet is a linear attention variant, but they seem to have stabilized training, and it doesn’t sacrifice performance on shorter sequences.
This study presents a technique that improves AI's understanding of image-text tasks by focusing on fine details and keywords, enhancing performance, especially when faced with unseen scenarios.
One of the top multimodal models today was developed by the UniLM team at Microsoft. They have now released demo code, weight checkpoints, and more for you to play with and integrate into your systems.
There's a lot of excitement around using vector databases to augment the memory of different language models. This vector database has a simple interface and is primarily used for ChatGPT plugins. It runs many times faster than competitors while maintaining simplicity. It doesn't have the bells and whistles of many other offerings, but it performs the task well for plugins.
This research introduces SAM-DA, a new system that improves real-time tracking of unmanned aerial vehicles (UAVs) at night. By using the Segment Anything Model (SAM) to generate detailed training samples from limited nighttime images, it significantly enhances the tracking ability while using fewer resources.
WebGLM is a new online tool powered by artificial intelligence designed to answer questions by searching the web and providing the most relevant information. It improves upon previous models by offering increased accuracy and efficiency, and studies show that it performs as well as, if not better than, more complex systems.
If you’re doing anything with generative AI and playing internally with your own LLMs, you should check out this post about Streamlit.
Celebrities are partnering with brands to create AI-generated duplicates of themselves that can be used in marketing campaigns. This gives celebrities more control over their own images and allows brands to use the stars in ways they never could before.
We’ve covered this in previous newsletters, but it just got an upgrade. Inspired by Andrej’s great NanoGPT repo, the NanoT5 repo can train a model to equivalent T5 paper performance 150x faster. It’s definitely worth checking out if you care about training models from scratch.
Ryan Moulton discusses the various methods through which digital minds can acquire knowledge, such as direct programming, machine learning, simulation, and training.
The article explores the issue of AI model weights and their availability as open-source. It argues that while the source code of AI models may be open, the weights—which contain the actual learned knowledge—are typically not shared openly due to various reasons such as intellectual property concerns, privacy, and commercial interests.