Is AI deflationary?

welcome welcome! and Bienvenue to the 161 new folks joining since yesterday.

I need to work on my process a bit as it's getting a little longer to get these out - I'm like the cook in the back being shouted at for not having my prep work on time. Idiot sandwich anyone 🥪. At some point soon I'll switch on a proper domain to make this look all proper. If you've got any feedback or suggestions on what I'm up to, please let me know.

Let's get to it, you can also join us on Discord if you want.

Prompt: List the topics in today’s email:

🫦 BEN'S BITES

Midjourney v4 is awesome, but, it’s pricey too. So PromptHero built a tool for Stable Diffusion fine-tuned on Midjourney v4 images.

An interview (podcast + written) with Dylan Field, (a BB sub too 👋), where he talks about the Figma acquisition (obviously) but he also talks about AI at the end. Here are some notes I took from it :

- There is a lot of hype around AI at the moment, which may lead to a bubble.- AI could be deflationary, as it may make many existing jobs obsolete- Inspiration can be drawn from lots of places. How do you work with an AI agent versus an AI agent replacing a designer?- The bear case is what we have now in AI. If we can take that and just apply it to new areas of the economy, that is already going to be very disruptive.

p.s. we should totally do our interview, right? - Dylan, are you up for it?

I stopped including the different Diffusion models as they are so many now, but, for this paper-cutting model, I’ll make an exception.

Stable-Diffusion with DPM-Solver can generate high-quality samples by diffusion models within only 10-25 steps.

SEScore - a reference-based text-generation evaluation metric that requires no pre-human-annotated error data.

Runway introduced Text to Color Grade - Use natural language to colour grade your videos.

Hugging Face’s store is now open!

Twitter has had a search problem forever, you can’t find anything! Luckily Kiril built a semantic search for Twitter, after seeing the ebook search tool.

AI Assistants that talk like a human - Quickchat

DeepMind released TAP-Vid, a new benchmark for tracking any point in videos.

Deploy a Gradio app to Spaces from a notebook without having to refactor your code into a script (which Hugging Face currently requires you to do).

Why generative AI for music is harder than images…TL;DR - training and it’s not visual!

This magician has been conjuring up AI-generated potions on Twitter.

ReAct by Google - a combo of reasoning and acting advances that enable language models to solve various language reasoning and decision-making tasks.

A sneak peek at Capsule's new tool that magically adds b-roll to videos. How does it work? An ASR model transcribes your video into text THEN a diffusion model generates images from text AND THEN Capsule's video markup language automates editing the video right inside your browser. No apps needed. Video-to-Text-to-Video 😅.

A tutorial for creating a low poly bedroom in Blender.

A deep-dive on Musk’s move to eliminate the responsible AI team at Twitter.

I asked for this on Twitter! A simple AI-terminology breakdown. AssemblyAI wrote an explanation of 10 AI terms you should know (in plain English).

Contrastive search can generate human-level text using off-the-shelf language models across 16 languages. Generating human-level text with contrastive search in Transformers. A demo that directly compares contrastive search with other popular decoding methods.

A physics-based method for assembling physical items. Think manufacturing, 3D objects, product assembly, recycling etc.

Mysteries of mode collapse due to RLHF is a pretty interesting read, but a little technical. So here’s Ben’s Bitesize Summary:

It discusses mode collapse in GPT models that have undergone RLHF (reinforcement learning from human feedback) training.

Key takeaways:

1. Mode collapse is a phenomenon where the model produces very similar responses to similar prompts.

2. Mode collapse can happen out of distribution, meaning that the model can produce the same responses even when the prompts are very different from what the model was trained on.

3. The reason for mode collapse is not fully understood, but it may be due to the model overfitting to patterns present in the training data.

Self-conditioned Embedding Diffusion, a continuous diffusion mechanism that allows for the flexible and scalable learning of diffusion models for both conditional and unconditional text generation.

A method for removing artificial reverb from music without requiring pairs of data for training.

🖼 AI IMAGES OF THE DAY

Knolling Gandalf's bag

🤗 SHARE BENS BITES

Send this with 1 AI-curious friend and receive my AI project tracker database!

or copy/paste this link: https://bensbites.beehiiv.com/subscribe?ref=PLACEHOLDER

👋 SEE YA

⭐️ HOW DID WE DO?

⭐️ REAL REVIEWS