Let The AI Fill In The Blanks

🔷 Subscribe to get breakdowns of the most important developments in AI in your inbox every morning.

Here’s today at a glance:

• An AutoComplete For Your Imagination

• The Framework of Knowledge

• Things happen

• AI artwork of the day

📱An AutoComplete For Your Imagination

Generate high-quality images, on your mobile device, as you type. It’s that old feeling of magic once more, oh Google, how we have missed ye.

MobileDiffusion is the deployment of an array of optimization methods to address what the Google team states are the two major issues in image diffusion models:

• their inherent design is to reduce noise in images iteratively, which requires multiple evaluations of the model

• the network architecture consists of a large number of parameters making them computationally expensive

MobileDiffusion achieves a remarkable sub-second inference speed for generating a 512 × 512 image on mobile devices, establishing a new state of the art.

The whole paper is a great read on how individual small optimizations each yielding 10-30% reduction in parameter count or computation can total to something significant. Their big breakthrough though is in getting the diffusion down to a single step using a GAN (a Generative Adversarial Network), reaching a new level of sampling efficiency. In effect, the model takes a single wild guess rather than multiple iterative how-am-I-doing check-ins to form an image from white noise and a text prompt.

It is not often you see an order of magnitude increase in performance (2 billion parameters for 50 steps vs 400 million parameters for 1 step), but there you have it.

Now just think what happens when the input prompt is a brain EEG.

Share this story

💡The Framework of Knowledge

What Fortune 500 companies are really spending money on AI for right now:

• format their existing documents

• search through them

• explain them using a language model

This simplistic process goes by the rather clunky moniker of Retrieval Augmented Generation but has become very attractive as the search AI (embeddings) seems to understand context much better than simple keywords.

The main issue with this methodology is the key question of how many documents get fed to the language model to explain. This paper from Stanford provides a solution that beats state of the art, namely just repeatedly summarize small chunks of documents until the main themes are stored in the top layer. Notably, it ends up clustering the smaller chunks by meaning (semantic similarity).

And then when answering, performs a tree retrieval to assemble context.

And this framework is relatively agnostic to which language model is used.. after all the LLM is just an understanding engine applied to context, which is what is really important.

A truly elegant solution. Although I must admit, I’ve heard of more than one firm already using recursive summarization in production, so I doubt this is really new knowledge, rather than just proving what we already know to be true.

Share this story

🗞️ Things Happen

• MidJourney is shipping. Fresh off the Niji 6 anime model release last week, v6 beta might be released as soon as this week. Consistent styles are in testing as a precursor to consistent character. Once that hits, it’s all over for one generation of image-generation startups.

• Nat Friedman and Dan Gross’ AIGrant starts applications for their third batch. This is probably the leading incubator of AI talent right now, having backed Perplexity and Julius in earlier batches. With their own cluster to boot. It’s startling to me that YCombinator has chosen not to invest in GPUs, relegating startups there to become GPT wrappers.

• Stack Overflow questions and answers down by 50% in 2023. Hat tip @swyx.

🖼️ AI Artwork Of The Day

That’s it for today! Become a subscriber for daily breakdowns of what’s happening in the AI world: