Sunday Rundown #61: Mega Meta Model & Sloppy Batman
Sunday Bonus #21: Combining --sref and "describe" in Midjourney
Happy Sunday, friends!
Welcome back to the weekly look at generative AI that covers the following:
Sunday Rundown (free): this week’s AI news and an AI fail for your entertainment.
Sunday Bonus (paid): a goodie for my paid subscribers (a guide, tip, walkthrough of a tool, etc).
Let’s get to it.
🗞️ AI news
Here are this week’s AI developments.
👩💻 AI releases
New stuff you can try right now:
Adobe introduced a slew of new AI features to its Photoshop and Illustrator products.
ElevenLabs released its Turbo 2.5 model, bringing “high-quality low-latency conversational AI” to 32 languages.
Google brought its fast and competent Gemini 1.5 Flash model to free
gemini.google.com and expanded its context window to 32K tokens.
Kling, the Chinese text-to-video model that’s been compared to OpenAI’s Sora, is now available to everyone. You get free daily credits to generate 5-second clips from text or a starting image.
Luma introduced a Loops feature to its Dream Machine video generator, making it easy to create endlessly looping clips.
Meta delivered on its promise and released the upgraded Llama 3.1 family. It comes in three sizes: 8B, 70B, and the massive 405B model, which goes head-to-head with the big guns like GPT-4o and Claude 3.5 Sonnet. It’s the best open-source LLM and—in some areas—the best LLM overall. (Visit llama.meta.com to learn more and try the models.)
Microsoft introduced its own version of Google’s AI Overviews called Bing generative search.
Mistral rolled out Mistral Large 2, an upgrade to its flagship model. It shows improvements in code generation, math, and reasoning, as well as better multilingual support.
Perplexity had two releases this week:
Fixed navigational searches. Searching for e.g. “Coca-Cola” will now show a direct link to the www.coca-cola.com homepage. (It frustrated me that Perplexity didn’t have this before.)
Voice Mode for Android and iOS. (You’ll need Perplexity Pro.)
Suno now lets Pro and Premier users split AI-generated tracks into Stems to separate vocals from instrumentals.
Udio is one-upping Suno every step of the way and has released version 1.5 of its text-to-music model. It returns higher-quality tracks and performs better in multiple languages. Udio also added the ability to split tracks into stems and upload input audio to influence new generations.
🔬 AI research
Cool stuff you might get to try one day:
OpenAI is testing a prototype of its SearchGPT tool that pulls data from multiple sources to provide grounded answers to questions. (Join the waitlist.)
Stability AI is working on Stable Video 4D, which turns input videos into controllable videos of the same object from different camera angles. (You can access the research version on Hugging Face if you know what to do with it.)
📖 AI resources
Helpful stuff that teaches you about AI:
“Llama 405b: Full 92-page Analysis” [VIDEO] - superb deep dive into Meta’s Llama 3.1 research paper along with additional benchmarking by AI Explained.
“Open Source AI Is the Path Forward” - an emphatic case for open-source models by Mark Zuckerberg.
🔀 AI random
Other notable AI stories of the week:
Two models from Google DeepMind—AlphaProof and AlphaGeometry 2—have managed to solve the International Mathematical Olympiad up to silver medalist standards.
🤦♂️ AI fail of the week
Anything to share?
Sadly, Substack doesn’t allow free subscribers to comment on posts with paid sections, but I am always open to your feedback. You can message me here:
💰 Sunday Bonus #21: How to combine --sref and /describe in Midjourney
Long before Midjourney had the --sref (style reference) for consistent styles and the --cref (character reference) for consistent characters, it had the /describe command.
This command attempted to reverse-engineer a prompt from an uploaded image to create a similar one. It didn’t always work 100% but would typically give you good ideas for descriptors and style terms to use.
Now that --sref largely takes care of “copying” the style of an image, you might think /describe is obsolete.
But you’d be wrong: There’s a way to combine --sref and /describe to get even closer to your intended style.
Here’s how: