Run 2x Larger AI Models in Limited Memory 🔥

Today’s top AI Highlights:

1. You Can Now Archive ChatGPT Chats

2. LLM in a flash: Efficient LLM Inference with Limited Memory

3. VideoPoet: An LLM for zero-shot video generation

4. Google DeepMind used LLM to Solve an Unsolved Math Problem

5. AI Cannot be Named a Patent ‘Inventor’

& so much more!

Read time: 3 mins

Latest Developments 🌍

Archive your ChatGPT Chats 🚮

You can now archive your chats, removing them from your sidebar without deleting them. You can see your archived chats in Settings. Currently available on Web and iOS with Android coming soon.

Running 2x Larger LLMs on Limited Memory 🚀

LLMs have immense computational and memory requirements, posing challenges for devices with limited DRAM capacity. Researchers at Apple have proposed a technique that addresses this challenge in deploying advanced AI models, especially on resource-constrained devices. The technique allows models up to 2x the size of the device's DRAM capacity to be run, widening the accessibility of sophisticated AI technologies.

Key Highlights:

1. The core of this new method involves storing LLM parameters on flash memory and selectively loading them into DRAM as needed. This approach greatly reduces the amount of memory these models use, allowing them to run on devices with less available memory.

2. The paper introduces 'windowing' and 'row-column bundling' that together enable running models that are double the size, and speed up the inference with a 4-5x increase on CPUs and a 20-25x increase on GPUs.

3. The strategy capitalizes on the fact that not all parts of LLMs are always active, by only loading the active parts, reducing unnecessary data loading. The approach also includes a design that balances the need to transfer small amounts of data with the efficiency of reading larger blocks of data.

Google’s Model to Rival Runway and Pika Labs 🦾

Producing coherent and dynamic large motions in video generation has been a longstanding challenge. Google has released VideoPoet which employs the prowess of LLMs for an array of video generation tasks. This innovative model is capable of text-to-video, image-to-video, and video-to-audio conversions, along with advanced techniques like video stylization, inpainting, and outpainting.

Key Highlights:

1. VideoPoet utilizes an autoregressive language model approach, facilitating learning across multiple modalities including video, image, audio, and text. This is achieved using specialized tokenizers like MAGVIT V2 for video and image, and SoundStream for audio, enabling the conversion of these media into discrete tokens.

2. Primarily generating videos in portrait orientation, VideoPoet caters to the growing demand for short-form content. The model can generate longer videos also by chaining short clips, maintaining object appearance consistency over extended durations.

3. VideoPoet offers extensive editing capabilities to interactively edit video clips and control object motions, including modifying actions and manipulating objects within the video. The model can also accurately control camera movements by appending techniques in the prompt like zoom, pan, arc shot, and crane shots.

LLM Solves a Previously-Unsolvable Math Problem 🧐

Google DeepMind has successfully employed an LLM to solve an elusive problem in pure mathematics, marking the first instance where a language model has contributed a novel and verifiable solution to a complex scientific question, offering insights previously unknown in the field.

Key Highlight:

1. DeepMind's tool, FunSearch, is at the heart of this development. It uses Codey, an LLM based on Google’s PaLM 2, and fine-tuned for computer code. This allows it to approach a broader range of mathematical challenges to solve complex problems.

2. FunSearch provided a correct and previously unknown solution to the cap set problem, a challenge that has perplexed mathematicians for years. FunSearch worked by generating and refining millions of suggestions, eventually arriving at a new, correct solution that was not based on pre-existing information.

3. Beyond the cap set problem, FunSearch showcased its versatility in the bin packing problem, a classic issue in mathematics and computer science. In this instance, the tool devised a solution method faster than those previously created by human mathematicians.

AI Cannot be Named a Patent ‘Inventor’ 👩‍🔬

Where at one hand we are witnessing AI being increasingly used for research and complex real-world problems (just like above!), the UK Supreme Court has ruled that AI cannot be legally named as an inventor to secure patent rights. This decision comes after Dr. Stephen Thaler, a US technologist, challenged the Intellectual Property Office for rejecting his attempt to list an AI, named DABUS, as the inventor of two patents.

Dr. Thaler claimed that his DABUS autonomously created a food or drink container and a light beacon. He argued that he was entitled to patent rights over its inventions. The court emphasized that DABUS is a machine with no legal personality, and Thaler does not have an independent right to obtain a patent for any technical advance made by AI. (Source)

Tools of the Trade ⚒️

• Drawww: AI drawing app for iPad, offering fast, secure, and creative tools with offline functionality, precision instruments, and a unique blend of vector and pixel art. The app utilizes Apple's silicon for rapid AI results and offers on-device processing.

• Moonvalley.ai: New text-to-video AI tool to generate HD, 16:9 cinematic film-quality videos. You can select from a range of styles including hyperrealism, anime, fantasy, and more. ​It is evolving to introduce new features for camera movement and aspect ratio control, and is accessible via Discord.

• Creatify: AI-powered app for creating high-quality marketing videos directly from product links or text descriptions, with features like AI-generated scripts, customization options, and a library of stock footage.

• Findly: AI tool to chat with your Google Analytics 4 data using natural language and get reports quickly. It integrates with Slack for easy communication and data sharing. It offers intuitive data visualization, real-time insights, and user-friendly reporting.

Hot Takes 🔥

Some predictions for 2024 – keeping only the more controversial ones. You certainly saw the non-controversial ones (multimodality, etc) already 1. At least 10 new unicorn companies building SOTA open foundation models in 2024 […]2. 2024 will be reality-check for older AI-unicorn pioneers […]3. Model quality will be harder and harder to evaluate in 2024 […]4. The return of academia […]5. Dangerous times for annotation companies […]6. The rise of synthetic data […] ~ Thomas Wolf

Meme of the Day 🤡

That’s all for today!

See you tomorrow with more such AI-filled content. Don’t forget to subscribe and give your feedback below 👇

Real-time AI Updates 🚨

⚡️ Follow me on Twitter @Saboo_Shubham for lightning-fast AI updates and never miss what’s trending!!

PS: I curate this AI newsletter every day for FREE, your support is what keeps me going. If you find value in what you read, share it with your friends by clicking the share button below!