. .AI: More and faster AI everywhere (12.20.23)

Friends, more and faster AI is the theme of today's news. And yet, still nobody can beat GPT-4! These are the top stories the AI community is talking about, from yesterday and today.

Yesterday someone smart told me they using this newsletter in two ways: to track the state of the art in this sprawling world of AI, and to confirm or revise their geopolitical and market map thinking about AI. I thought that might be of use to others, but perhaps you find it useful in other ways. Let me know, if you can take a few minutes to shoot me a reply.

And now here’s today’s news.

First impacted: AI developers, Data center operators
Time to impact: Medium

A Cupertino firm called Etched has launched silicon chips with transformer architecture, which the company says results in powerful servers for transformer inference. According to Etched, this technology enables things like real-time voice agents to process thousands of words in milliseconds, improves coding with tree search, and facilitates the generation of new content in real-time, with the servers capable of handling future 100 trillion parameter models. On the home page, you can see the claim that Etched towers far above NVIDIA hardware in tokens per second.

[Etched | The World's First Transformer Supercomputer] Share by email

First impacted: AI researchers, AI developers
Time to impact: Medium

Researchers from the Institute of Parallel and Distributed Systems at Shanghai Jiao Tong University have developed PowerInfer, a tool designed to operate LLMs on a standard computer with a single consumer-grade GPU. According to the researchers, PowerInfer reduces the GPU's memory requirements and data transfers between the CPU and GPU by loading a small group of neurons onto the GPU for quick access, and it performs up to 11.69 times better than llama.cpp, while maintaining the model's accuracy.

* PowerInfer's design includes adaptive predictors and neuron-aware sparse operators, which optimize the efficiency of neuron activation and computational sparsity. [sjtu.edu.cn] Share by email

First impacted: AI users and developers
Time to impact: Short

A comparison of AI models Gemini Pro, GPT 3.5 Turbo, GPT 4 Turbo, and Mixtral (Mistral's new Mixture of Experts model) on general-purpose reasoning datasets GSM8K, SVAMP, and ASDIV was conducted in a recent study by Carnegie Mellon and BerriAI. The study found that GPT 4 Turbo consistently outperformed the other models, while Gemini Pro's performance was comparable to that of GPT 3.5 Turbo across all datasets.

[Gemini Mathematics | Zeno] Share by email

First impacted: Researchers in the AI field, Research institutions and funding bodies
Time to impact: Long

Berkeley professor Ben Recht writes about a Nature study alleging the AI community is wrestling with an overabundance of research papers, which some argue is diminishing research quality and placing undue pressure on researchers. It is suggested that research institutions and funding bodies should prioritize the quality of a researcher’s work over the quantity of their publications, thereby encouraging young researchers to concentrate on producing fewer, but more impactful papers.

* The post says some graduate student applicants have their names on as many as 50 papers, with over 10,000 citations. The author recommends 3 good ones, not 50 bad ones.
* It also points out the recent trend of longer papers, with junior candidates writing theoretical papers that are 50-100 pages long, often inventing a problem no one cares about and then spending the rest of the paper addressing it. [Too Much Information] Share by email

First impacted: Musicians, AI experts
Time to impact: Medium

Suno, a Cambridge-based team of musicians and AI specialists, says it is democratizing music creation through its new (and much-hyped) text-to-music AI platform, requiring no musical experience or instruments. You describe a song and it generates music and lyrics for a 1 minute long preview. It's pretty impressive, but it does sound like you might expect if I told you ChatGPT learned to play a guitar.

[Suno AI] Share by email

First impacted: Video content creators, digital marketers
Time to impact: Medium

VideoPoet, an LLM developed for video creation tasks, has been introduced, demonstrating capabilities such as text-to-video, image-to-video, video stylization, video inpainting and outpainting, and video-to-audio tasks. According to a user preference evaluation, Google says VideoPoet was chosen 24-35% of the time for better adherence to prompts than a competing model, and 41-54% of the time for generating more engaging motion.

* By default it generates 2 second long videos, but Google says it can generate longer videos by using the last 1 second of video to predict the next 1 second. [VideoPoet: A large language model for zero-shot video generation] Share by email

You might also be interested in …

First impacted: AI users and thinkers
Time to impact: Short

Digital futurist Brian Solis, in partnership with JESS3 design studio, has launched the GenAI Prism, an infographic they describe as a comprehensive representation of the generative AI sector. They say it "offers a mental model to mindfully and intentionally approach prompts toward more intentional outcomes and insights."

[Introducing The GenAI Prism Infographic: A Framework For Collaborating With Generative AI] Share by email

That’s it! More AI news tomorrow!